diff --git "a/nemotron_coreml_160ms/encoder/encoder_int8.mlmodelc/model.mil" "b/nemotron_coreml_160ms/encoder/encoder_int8.mlmodelc/model.mil"
new file mode 100644--- /dev/null
+++ "b/nemotron_coreml_160ms/encoder/encoder_int8.mlmodelc/model.mil"
@@ -0,0 +1,4190 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 24, 70, 1024]> cache_channel, tensor<int32, [1]> cache_len, tensor<fp32, [1, 24, 1024, 8]> cache_time, tensor<fp32, [1, 128, 25]> mel, tensor<int32, [1]> mel_length) {
+            tensor<fp32, [256]> module_pre_encode_conv_0_bias = const()[name = tensor<string, []>("module_pre_encode_conv_0_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [256, 1, 3, 3]> module_pre_encode_conv_0_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_pre_encode_conv_0_weight_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1152))), scale = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3840))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3520)))];
+            tensor<fp32, [256]> module_pre_encode_conv_2_bias = const()[name = tensor<string, []>("module_pre_encode_conv_2_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4928)))];
+            tensor<fp32, [256, 1, 3, 3]> module_pre_encode_conv_2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_pre_encode_conv_2_weight_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(6016))), scale = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8704))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8384)))];
+            tensor<fp32, [256]> module_pre_encode_conv_3_bias = const()[name = tensor<string, []>("module_pre_encode_conv_3_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9792)))];
+            tensor<fp32, [256, 256, 1, 1]> module_pre_encode_conv_3_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_pre_encode_conv_3_weight_quantized"), quantized_data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(10880))), scale = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76800))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(76480)))];
+            tensor<fp32, [256]> module_pre_encode_conv_5_bias = const()[name = tensor<string, []>("module_pre_encode_conv_5_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77888)))];
+            tensor<fp32, [256, 1, 3, 3]> module_pre_encode_conv_5_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_pre_encode_conv_5_weight_quantized"), quantized_data = tensor<int8, [256, 1, 3, 3]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(78976))), scale = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81664))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81344)))];
+            tensor<fp32, [256]> module_pre_encode_conv_6_bias = const()[name = tensor<string, []>("module_pre_encode_conv_6_bias"), val = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82752)))];
+            tensor<fp32, [256, 256, 1, 1]> module_pre_encode_conv_6_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_pre_encode_conv_6_weight_quantized"), quantized_data = tensor<int8, [256, 256, 1, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(83840))), scale = tensor<fp32, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149760))), zero_point = tensor<int8, [256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(149440)))];
+            tensor<fp32, [1024]> module_pre_encode_out_bias = const()[name = tensor<string, []>("module_pre_encode_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150848)))];
+            tensor<fp32, [1024, 4352]> module_pre_encode_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_pre_encode_out_weight_quantized"), quantized_data = tensor<int8, [1024, 4352]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(155008))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4612608))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4611520)))];
+            tensor<fp32, [1024]> module_layers_0_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_0_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4616768)))];
+            tensor<fp32, [1024]> module_layers_0_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_0_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4620928)))];
+            tensor<fp32, [4096, 1024]> module_layers_0_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4625088))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8823616))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8819456)))];
+            tensor<fp32, [1024, 4096]> module_layers_0_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(8840064))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13035520))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13034432)))];
+            tensor<fp32, [1024]> module_layers_0_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_0_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13039680)))];
+            tensor<fp32, [1024]> module_layers_0_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_0_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13043840)))];
+            tensor<fp32, [8, 128]> module_layers_0_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_0_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13048000)))];
+            tensor<fp32, [8, 128]> module_layers_0_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_0_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13052160)))];
+            tensor<fp32, [1024, 1024]> module_layers_0_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(13056320))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14106048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14104960)))];
+            tensor<fp32, [1024, 1024]> module_layers_0_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(14110208))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15159936))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15158848)))];
+            tensor<fp32, [1024, 1024]> module_layers_0_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15164096))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16213824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16212736)))];
+            tensor<fp32, [1024, 1024]> module_layers_0_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(16217984))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17267712))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17266624)))];
+            tensor<fp32, [1024]> module_layers_0_norm_conv_bias = const()[name = tensor<string, []>("module_layers_0_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17271872)))];
+            tensor<fp32, [1024]> module_layers_0_norm_conv_weight = const()[name = tensor<string, []>("module_layers_0_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17276032)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_0_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(17280192))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19379520))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19377408)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_0_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19387776))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19398144))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19397056)))];
+            tensor<fp32, [1024]> module_layers_0_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_0_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19402304)))];
+            tensor<fp32, [1024]> module_layers_0_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_0_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19406464)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_0_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(19410624))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20460352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20459264)))];
+            tensor<fp32, [1024]> module_layers_0_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_0_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20464512)))];
+            tensor<fp32, [1024]> module_layers_0_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_0_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20468672)))];
+            tensor<fp32, [4096, 1024]> module_layers_0_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(20472832))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24671360))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24667200)))];
+            tensor<fp32, [1024, 4096]> module_layers_0_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_0_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(24687808))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28883264))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28882176)))];
+            tensor<fp32, [1024]> module_layers_0_norm_out_bias = const()[name = tensor<string, []>("module_layers_0_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28887424)))];
+            tensor<fp32, [1024]> module_layers_0_norm_out_weight = const()[name = tensor<string, []>("module_layers_0_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28891584)))];
+            tensor<fp32, [1024]> module_layers_1_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_1_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28895744)))];
+            tensor<fp32, [1024]> module_layers_1_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_1_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28899904)))];
+            tensor<fp32, [4096, 1024]> module_layers_1_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(28904064))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33102592))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33098432)))];
+            tensor<fp32, [1024, 4096]> module_layers_1_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(33119040))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37314496))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37313408)))];
+            tensor<fp32, [1024]> module_layers_1_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_1_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37318656)))];
+            tensor<fp32, [1024]> module_layers_1_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_1_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37322816)))];
+            tensor<fp32, [8, 128]> module_layers_1_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_1_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37326976)))];
+            tensor<fp32, [8, 128]> module_layers_1_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_1_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37331136)))];
+            tensor<fp32, [1024, 1024]> module_layers_1_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(37335296))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38385024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38383936)))];
+            tensor<fp32, [1024, 1024]> module_layers_1_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(38389184))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39438912))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39437824)))];
+            tensor<fp32, [1024, 1024]> module_layers_1_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(39443072))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40492800))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40491712)))];
+            tensor<fp32, [1024, 1024]> module_layers_1_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40496960))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41546688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41545600)))];
+            tensor<fp32, [1024]> module_layers_1_norm_conv_bias = const()[name = tensor<string, []>("module_layers_1_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41550848)))];
+            tensor<fp32, [1024]> module_layers_1_norm_conv_weight = const()[name = tensor<string, []>("module_layers_1_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41555008)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_1_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(41559168))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43658496))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43656384)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_1_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43666752))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43677120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43676032)))];
+            tensor<fp32, [1024]> module_layers_1_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_1_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43681280)))];
+            tensor<fp32, [1024]> module_layers_1_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_1_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43685440)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_1_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(43689600))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44739328))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44738240)))];
+            tensor<fp32, [1024]> module_layers_1_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_1_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44743488)))];
+            tensor<fp32, [1024]> module_layers_1_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_1_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44747648)))];
+            tensor<fp32, [4096, 1024]> module_layers_1_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(44751808))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48950336))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48946176)))];
+            tensor<fp32, [1024, 4096]> module_layers_1_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_1_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(48966784))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53162240))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53161152)))];
+            tensor<fp32, [1024]> module_layers_1_norm_out_bias = const()[name = tensor<string, []>("module_layers_1_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53166400)))];
+            tensor<fp32, [1024]> module_layers_1_norm_out_weight = const()[name = tensor<string, []>("module_layers_1_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53170560)))];
+            tensor<fp32, [1024]> module_layers_2_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_2_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53174720)))];
+            tensor<fp32, [1024]> module_layers_2_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_2_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53178880)))];
+            tensor<fp32, [4096, 1024]> module_layers_2_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(53183040))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57381568))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57377408)))];
+            tensor<fp32, [1024, 4096]> module_layers_2_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(57398016))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61593472))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61592384)))];
+            tensor<fp32, [1024]> module_layers_2_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_2_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61597632)))];
+            tensor<fp32, [1024]> module_layers_2_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_2_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61601792)))];
+            tensor<fp32, [8, 128]> module_layers_2_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_2_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61605952)))];
+            tensor<fp32, [8, 128]> module_layers_2_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_2_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61610112)))];
+            tensor<fp32, [1024, 1024]> module_layers_2_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(61614272))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62664000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62662912)))];
+            tensor<fp32, [1024, 1024]> module_layers_2_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(62668160))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63717888))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63716800)))];
+            tensor<fp32, [1024, 1024]> module_layers_2_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(63722048))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64771776))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64770688)))];
+            tensor<fp32, [1024, 1024]> module_layers_2_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64775936))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65825664))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65824576)))];
+            tensor<fp32, [1024]> module_layers_2_norm_conv_bias = const()[name = tensor<string, []>("module_layers_2_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65829824)))];
+            tensor<fp32, [1024]> module_layers_2_norm_conv_weight = const()[name = tensor<string, []>("module_layers_2_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65833984)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_2_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(65838144))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67937472))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67935360)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_2_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67945728))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67956096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67955008)))];
+            tensor<fp32, [1024]> module_layers_2_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_2_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67960256)))];
+            tensor<fp32, [1024]> module_layers_2_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_2_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67964416)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_2_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(67968576))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69018304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69017216)))];
+            tensor<fp32, [1024]> module_layers_2_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_2_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69022464)))];
+            tensor<fp32, [1024]> module_layers_2_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_2_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69026624)))];
+            tensor<fp32, [4096, 1024]> module_layers_2_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(69030784))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73229312))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73225152)))];
+            tensor<fp32, [1024, 4096]> module_layers_2_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_2_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(73245760))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77441216))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77440128)))];
+            tensor<fp32, [1024]> module_layers_2_norm_out_bias = const()[name = tensor<string, []>("module_layers_2_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77445376)))];
+            tensor<fp32, [1024]> module_layers_2_norm_out_weight = const()[name = tensor<string, []>("module_layers_2_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77449536)))];
+            tensor<fp32, [1024]> module_layers_3_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_3_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77453696)))];
+            tensor<fp32, [1024]> module_layers_3_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_3_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77457856)))];
+            tensor<fp32, [4096, 1024]> module_layers_3_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(77462016))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81660544))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81656384)))];
+            tensor<fp32, [1024, 4096]> module_layers_3_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(81676992))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85872448))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85871360)))];
+            tensor<fp32, [1024]> module_layers_3_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_3_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85876608)))];
+            tensor<fp32, [1024]> module_layers_3_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_3_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85880768)))];
+            tensor<fp32, [8, 128]> module_layers_3_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_3_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85884928)))];
+            tensor<fp32, [8, 128]> module_layers_3_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_3_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85889088)))];
+            tensor<fp32, [1024, 1024]> module_layers_3_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(85893248))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86942976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86941888)))];
+            tensor<fp32, [1024, 1024]> module_layers_3_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(86947136))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87996864))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(87995776)))];
+            tensor<fp32, [1024, 1024]> module_layers_3_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88001024))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89050752))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89049664)))];
+            tensor<fp32, [1024, 1024]> module_layers_3_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89054912))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90104640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90103552)))];
+            tensor<fp32, [1024]> module_layers_3_norm_conv_bias = const()[name = tensor<string, []>("module_layers_3_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90108800)))];
+            tensor<fp32, [1024]> module_layers_3_norm_conv_weight = const()[name = tensor<string, []>("module_layers_3_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90112960)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_3_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(90117120))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92216448))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92214336)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_3_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92224704))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92235072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92233984)))];
+            tensor<fp32, [1024]> module_layers_3_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_3_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92239232)))];
+            tensor<fp32, [1024]> module_layers_3_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_3_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92243392)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_3_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(92247552))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93297280))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93296192)))];
+            tensor<fp32, [1024]> module_layers_3_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_3_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93301440)))];
+            tensor<fp32, [1024]> module_layers_3_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_3_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93305600)))];
+            tensor<fp32, [4096, 1024]> module_layers_3_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(93309760))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97508288))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97504128)))];
+            tensor<fp32, [1024, 4096]> module_layers_3_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_3_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97524736))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101720192))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101719104)))];
+            tensor<fp32, [1024]> module_layers_3_norm_out_bias = const()[name = tensor<string, []>("module_layers_3_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101724352)))];
+            tensor<fp32, [1024]> module_layers_3_norm_out_weight = const()[name = tensor<string, []>("module_layers_3_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101728512)))];
+            tensor<fp32, [1024]> module_layers_4_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_4_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101732672)))];
+            tensor<fp32, [1024]> module_layers_4_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_4_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101736832)))];
+            tensor<fp32, [4096, 1024]> module_layers_4_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(101740992))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105939520))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105935360)))];
+            tensor<fp32, [1024, 4096]> module_layers_4_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(105955968))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110151424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110150336)))];
+            tensor<fp32, [1024]> module_layers_4_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_4_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110155584)))];
+            tensor<fp32, [1024]> module_layers_4_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_4_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110159744)))];
+            tensor<fp32, [8, 128]> module_layers_4_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_4_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110163904)))];
+            tensor<fp32, [8, 128]> module_layers_4_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_4_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110168064)))];
+            tensor<fp32, [1024, 1024]> module_layers_4_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(110172224))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111221952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111220864)))];
+            tensor<fp32, [1024, 1024]> module_layers_4_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(111226112))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112275840))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112274752)))];
+            tensor<fp32, [1024, 1024]> module_layers_4_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(112280000))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113329728))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113328640)))];
+            tensor<fp32, [1024, 1024]> module_layers_4_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(113333888))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114383616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114382528)))];
+            tensor<fp32, [1024]> module_layers_4_norm_conv_bias = const()[name = tensor<string, []>("module_layers_4_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114387776)))];
+            tensor<fp32, [1024]> module_layers_4_norm_conv_weight = const()[name = tensor<string, []>("module_layers_4_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114391936)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_4_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(114396096))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116495424))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116493312)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_4_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116503680))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116514048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116512960)))];
+            tensor<fp32, [1024]> module_layers_4_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_4_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116518208)))];
+            tensor<fp32, [1024]> module_layers_4_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_4_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116522368)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_4_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(116526528))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117576256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117575168)))];
+            tensor<fp32, [1024]> module_layers_4_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_4_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117580416)))];
+            tensor<fp32, [1024]> module_layers_4_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_4_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117584576)))];
+            tensor<fp32, [4096, 1024]> module_layers_4_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(117588736))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121787264))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121783104)))];
+            tensor<fp32, [1024, 4096]> module_layers_4_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_4_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(121803712))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125999168))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125998080)))];
+            tensor<fp32, [1024]> module_layers_4_norm_out_bias = const()[name = tensor<string, []>("module_layers_4_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126003328)))];
+            tensor<fp32, [1024]> module_layers_4_norm_out_weight = const()[name = tensor<string, []>("module_layers_4_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126007488)))];
+            tensor<fp32, [1024]> module_layers_5_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_5_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126011648)))];
+            tensor<fp32, [1024]> module_layers_5_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_5_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126015808)))];
+            tensor<fp32, [4096, 1024]> module_layers_5_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126019968))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130218496))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130214336)))];
+            tensor<fp32, [1024, 4096]> module_layers_5_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(130234944))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134430400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134429312)))];
+            tensor<fp32, [1024]> module_layers_5_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_5_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134434560)))];
+            tensor<fp32, [1024]> module_layers_5_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_5_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134438720)))];
+            tensor<fp32, [8, 128]> module_layers_5_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_5_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134442880)))];
+            tensor<fp32, [8, 128]> module_layers_5_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_5_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134447040)))];
+            tensor<fp32, [1024, 1024]> module_layers_5_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(134451200))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135500928))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135499840)))];
+            tensor<fp32, [1024, 1024]> module_layers_5_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(135505088))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136554816))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136553728)))];
+            tensor<fp32, [1024, 1024]> module_layers_5_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(136558976))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137608704))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137607616)))];
+            tensor<fp32, [1024, 1024]> module_layers_5_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(137612864))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138662592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138661504)))];
+            tensor<fp32, [1024]> module_layers_5_norm_conv_bias = const()[name = tensor<string, []>("module_layers_5_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138666752)))];
+            tensor<fp32, [1024]> module_layers_5_norm_conv_weight = const()[name = tensor<string, []>("module_layers_5_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138670912)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_5_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(138675072))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140774400))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140772288)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_5_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140782656))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140793024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140791936)))];
+            tensor<fp32, [1024]> module_layers_5_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_5_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140797184)))];
+            tensor<fp32, [1024]> module_layers_5_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_5_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140801344)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_5_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(140805504))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141855232))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141854144)))];
+            tensor<fp32, [1024]> module_layers_5_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_5_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141859392)))];
+            tensor<fp32, [1024]> module_layers_5_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_5_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141863552)))];
+            tensor<fp32, [4096, 1024]> module_layers_5_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(141867712))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146066240))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146062080)))];
+            tensor<fp32, [1024, 4096]> module_layers_5_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_5_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(146082688))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150278144))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150277056)))];
+            tensor<fp32, [1024]> module_layers_5_norm_out_bias = const()[name = tensor<string, []>("module_layers_5_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150282304)))];
+            tensor<fp32, [1024]> module_layers_5_norm_out_weight = const()[name = tensor<string, []>("module_layers_5_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150286464)))];
+            tensor<fp32, [1024]> module_layers_6_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_6_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150290624)))];
+            tensor<fp32, [1024]> module_layers_6_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_6_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150294784)))];
+            tensor<fp32, [4096, 1024]> module_layers_6_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(150298944))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154497472))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154493312)))];
+            tensor<fp32, [1024, 4096]> module_layers_6_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(154513920))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158709376))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158708288)))];
+            tensor<fp32, [1024]> module_layers_6_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_6_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158713536)))];
+            tensor<fp32, [1024]> module_layers_6_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_6_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158717696)))];
+            tensor<fp32, [8, 128]> module_layers_6_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_6_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158721856)))];
+            tensor<fp32, [8, 128]> module_layers_6_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_6_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158726016)))];
+            tensor<fp32, [1024, 1024]> module_layers_6_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(158730176))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159779904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159778816)))];
+            tensor<fp32, [1024, 1024]> module_layers_6_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(159784064))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160833792))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160832704)))];
+            tensor<fp32, [1024, 1024]> module_layers_6_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(160837952))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161887680))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161886592)))];
+            tensor<fp32, [1024, 1024]> module_layers_6_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(161891840))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162941568))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162940480)))];
+            tensor<fp32, [1024]> module_layers_6_norm_conv_bias = const()[name = tensor<string, []>("module_layers_6_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162945728)))];
+            tensor<fp32, [1024]> module_layers_6_norm_conv_weight = const()[name = tensor<string, []>("module_layers_6_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162949888)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_6_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(162954048))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165053376))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165051264)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_6_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165061632))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165072000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165070912)))];
+            tensor<fp32, [1024]> module_layers_6_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_6_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165076160)))];
+            tensor<fp32, [1024]> module_layers_6_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_6_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165080320)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_6_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(165084480))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166134208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166133120)))];
+            tensor<fp32, [1024]> module_layers_6_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_6_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166138368)))];
+            tensor<fp32, [1024]> module_layers_6_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_6_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166142528)))];
+            tensor<fp32, [4096, 1024]> module_layers_6_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(166146688))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170345216))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170341056)))];
+            tensor<fp32, [1024, 4096]> module_layers_6_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_6_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170361664))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174557120))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174556032)))];
+            tensor<fp32, [1024]> module_layers_6_norm_out_bias = const()[name = tensor<string, []>("module_layers_6_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174561280)))];
+            tensor<fp32, [1024]> module_layers_6_norm_out_weight = const()[name = tensor<string, []>("module_layers_6_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174565440)))];
+            tensor<fp32, [1024]> module_layers_7_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_7_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174569600)))];
+            tensor<fp32, [1024]> module_layers_7_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_7_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174573760)))];
+            tensor<fp32, [4096, 1024]> module_layers_7_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(174577920))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178776448))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178772288)))];
+            tensor<fp32, [1024, 4096]> module_layers_7_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(178792896))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182988352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182987264)))];
+            tensor<fp32, [1024]> module_layers_7_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_7_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182992512)))];
+            tensor<fp32, [1024]> module_layers_7_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_7_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(182996672)))];
+            tensor<fp32, [8, 128]> module_layers_7_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_7_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183000832)))];
+            tensor<fp32, [8, 128]> module_layers_7_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_7_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183004992)))];
+            tensor<fp32, [1024, 1024]> module_layers_7_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(183009152))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184058880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184057792)))];
+            tensor<fp32, [1024, 1024]> module_layers_7_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(184063040))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185112768))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185111680)))];
+            tensor<fp32, [1024, 1024]> module_layers_7_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(185116928))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186166656))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186165568)))];
+            tensor<fp32, [1024, 1024]> module_layers_7_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(186170816))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187220544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187219456)))];
+            tensor<fp32, [1024]> module_layers_7_norm_conv_bias = const()[name = tensor<string, []>("module_layers_7_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187224704)))];
+            tensor<fp32, [1024]> module_layers_7_norm_conv_weight = const()[name = tensor<string, []>("module_layers_7_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187228864)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_7_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187233024))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189332352))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189330240)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_7_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189340608))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189350976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189349888)))];
+            tensor<fp32, [1024]> module_layers_7_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_7_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189355136)))];
+            tensor<fp32, [1024]> module_layers_7_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_7_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189359296)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_7_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(189363456))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190413184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190412096)))];
+            tensor<fp32, [1024]> module_layers_7_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_7_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190417344)))];
+            tensor<fp32, [1024]> module_layers_7_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_7_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190421504)))];
+            tensor<fp32, [4096, 1024]> module_layers_7_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(190425664))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194624192))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194620032)))];
+            tensor<fp32, [1024, 4096]> module_layers_7_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_7_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(194640640))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198836096))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198835008)))];
+            tensor<fp32, [1024]> module_layers_7_norm_out_bias = const()[name = tensor<string, []>("module_layers_7_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198840256)))];
+            tensor<fp32, [1024]> module_layers_7_norm_out_weight = const()[name = tensor<string, []>("module_layers_7_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198844416)))];
+            tensor<fp32, [1024]> module_layers_8_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_8_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198848576)))];
+            tensor<fp32, [1024]> module_layers_8_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_8_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198852736)))];
+            tensor<fp32, [4096, 1024]> module_layers_8_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(198856896))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203055424))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203051264)))];
+            tensor<fp32, [1024, 4096]> module_layers_8_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(203071872))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207267328))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207266240)))];
+            tensor<fp32, [1024]> module_layers_8_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_8_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207271488)))];
+            tensor<fp32, [1024]> module_layers_8_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_8_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207275648)))];
+            tensor<fp32, [8, 128]> module_layers_8_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_8_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207279808)))];
+            tensor<fp32, [8, 128]> module_layers_8_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_8_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207283968)))];
+            tensor<fp32, [1024, 1024]> module_layers_8_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(207288128))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208337856))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208336768)))];
+            tensor<fp32, [1024, 1024]> module_layers_8_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(208342016))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209391744))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209390656)))];
+            tensor<fp32, [1024, 1024]> module_layers_8_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(209395904))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210445632))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210444544)))];
+            tensor<fp32, [1024, 1024]> module_layers_8_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(210449792))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211499520))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211498432)))];
+            tensor<fp32, [1024]> module_layers_8_norm_conv_bias = const()[name = tensor<string, []>("module_layers_8_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211503680)))];
+            tensor<fp32, [1024]> module_layers_8_norm_conv_weight = const()[name = tensor<string, []>("module_layers_8_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211507840)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_8_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(211512000))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213611328))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213609216)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_8_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213619584))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213629952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213628864)))];
+            tensor<fp32, [1024]> module_layers_8_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_8_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213634112)))];
+            tensor<fp32, [1024]> module_layers_8_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_8_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213638272)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_8_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(213642432))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214692160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214691072)))];
+            tensor<fp32, [1024]> module_layers_8_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_8_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214696320)))];
+            tensor<fp32, [1024]> module_layers_8_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_8_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214700480)))];
+            tensor<fp32, [4096, 1024]> module_layers_8_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(214704640))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218903168))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218899008)))];
+            tensor<fp32, [1024, 4096]> module_layers_8_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_8_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(218919616))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223115072))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223113984)))];
+            tensor<fp32, [1024]> module_layers_8_norm_out_bias = const()[name = tensor<string, []>("module_layers_8_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223119232)))];
+            tensor<fp32, [1024]> module_layers_8_norm_out_weight = const()[name = tensor<string, []>("module_layers_8_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223123392)))];
+            tensor<fp32, [1024]> module_layers_9_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_9_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223127552)))];
+            tensor<fp32, [1024]> module_layers_9_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_9_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223131712)))];
+            tensor<fp32, [4096, 1024]> module_layers_9_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(223135872))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227334400))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227330240)))];
+            tensor<fp32, [1024, 4096]> module_layers_9_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(227350848))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231546304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231545216)))];
+            tensor<fp32, [1024]> module_layers_9_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_9_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231550464)))];
+            tensor<fp32, [1024]> module_layers_9_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_9_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231554624)))];
+            tensor<fp32, [8, 128]> module_layers_9_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_9_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231558784)))];
+            tensor<fp32, [8, 128]> module_layers_9_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_9_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231562944)))];
+            tensor<fp32, [1024, 1024]> module_layers_9_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(231567104))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232616832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232615744)))];
+            tensor<fp32, [1024, 1024]> module_layers_9_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(232620992))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233670720))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233669632)))];
+            tensor<fp32, [1024, 1024]> module_layers_9_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(233674880))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234724608))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234723520)))];
+            tensor<fp32, [1024, 1024]> module_layers_9_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(234728768))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235778496))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235777408)))];
+            tensor<fp32, [1024]> module_layers_9_norm_conv_bias = const()[name = tensor<string, []>("module_layers_9_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235782656)))];
+            tensor<fp32, [1024]> module_layers_9_norm_conv_weight = const()[name = tensor<string, []>("module_layers_9_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235786816)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_9_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(235790976))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237890304))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237888192)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_9_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237898560))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237908928))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237907840)))];
+            tensor<fp32, [1024]> module_layers_9_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_9_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237913088)))];
+            tensor<fp32, [1024]> module_layers_9_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_9_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237917248)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_9_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(237921408))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238971136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238970048)))];
+            tensor<fp32, [1024]> module_layers_9_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_9_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238975296)))];
+            tensor<fp32, [1024]> module_layers_9_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_9_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238979456)))];
+            tensor<fp32, [4096, 1024]> module_layers_9_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(238983616))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243182144))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243177984)))];
+            tensor<fp32, [1024, 4096]> module_layers_9_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_9_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(243198592))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247394048))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247392960)))];
+            tensor<fp32, [1024]> module_layers_9_norm_out_bias = const()[name = tensor<string, []>("module_layers_9_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247398208)))];
+            tensor<fp32, [1024]> module_layers_9_norm_out_weight = const()[name = tensor<string, []>("module_layers_9_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247402368)))];
+            tensor<fp32, [1024]> module_layers_10_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_10_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247406528)))];
+            tensor<fp32, [1024]> module_layers_10_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_10_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247410688)))];
+            tensor<fp32, [4096, 1024]> module_layers_10_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(247414848))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251613376))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251609216)))];
+            tensor<fp32, [1024, 4096]> module_layers_10_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(251629824))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255825280))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255824192)))];
+            tensor<fp32, [1024]> module_layers_10_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_10_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255829440)))];
+            tensor<fp32, [1024]> module_layers_10_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_10_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255833600)))];
+            tensor<fp32, [8, 128]> module_layers_10_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_10_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255837760)))];
+            tensor<fp32, [8, 128]> module_layers_10_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_10_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255841920)))];
+            tensor<fp32, [1024, 1024]> module_layers_10_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(255846080))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256895808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256894720)))];
+            tensor<fp32, [1024, 1024]> module_layers_10_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(256899968))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257949696))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257948608)))];
+            tensor<fp32, [1024, 1024]> module_layers_10_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(257953856))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259003584))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259002496)))];
+            tensor<fp32, [1024, 1024]> module_layers_10_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(259007744))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260057472))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260056384)))];
+            tensor<fp32, [1024]> module_layers_10_norm_conv_bias = const()[name = tensor<string, []>("module_layers_10_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260061632)))];
+            tensor<fp32, [1024]> module_layers_10_norm_conv_weight = const()[name = tensor<string, []>("module_layers_10_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260065792)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_10_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(260069952))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262169280))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262167168)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_10_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262177536))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262187904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262186816)))];
+            tensor<fp32, [1024]> module_layers_10_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_10_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262192064)))];
+            tensor<fp32, [1024]> module_layers_10_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_10_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262196224)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_10_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(262200384))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263250112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263249024)))];
+            tensor<fp32, [1024]> module_layers_10_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_10_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263254272)))];
+            tensor<fp32, [1024]> module_layers_10_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_10_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263258432)))];
+            tensor<fp32, [4096, 1024]> module_layers_10_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(263262592))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267461120))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267456960)))];
+            tensor<fp32, [1024, 4096]> module_layers_10_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_10_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(267477568))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271673024))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271671936)))];
+            tensor<fp32, [1024]> module_layers_10_norm_out_bias = const()[name = tensor<string, []>("module_layers_10_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271677184)))];
+            tensor<fp32, [1024]> module_layers_10_norm_out_weight = const()[name = tensor<string, []>("module_layers_10_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271681344)))];
+            tensor<fp32, [1024]> module_layers_11_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_11_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271685504)))];
+            tensor<fp32, [1024]> module_layers_11_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_11_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271689664)))];
+            tensor<fp32, [4096, 1024]> module_layers_11_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(271693824))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275892352))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275888192)))];
+            tensor<fp32, [1024, 4096]> module_layers_11_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(275908800))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280104256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280103168)))];
+            tensor<fp32, [1024]> module_layers_11_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_11_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280108416)))];
+            tensor<fp32, [1024]> module_layers_11_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_11_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280112576)))];
+            tensor<fp32, [8, 128]> module_layers_11_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_11_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280116736)))];
+            tensor<fp32, [8, 128]> module_layers_11_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_11_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280120896)))];
+            tensor<fp32, [1024, 1024]> module_layers_11_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(280125056))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281174784))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281173696)))];
+            tensor<fp32, [1024, 1024]> module_layers_11_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(281178944))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282228672))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282227584)))];
+            tensor<fp32, [1024, 1024]> module_layers_11_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(282232832))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283282560))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283281472)))];
+            tensor<fp32, [1024, 1024]> module_layers_11_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(283286720))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284336448))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284335360)))];
+            tensor<fp32, [1024]> module_layers_11_norm_conv_bias = const()[name = tensor<string, []>("module_layers_11_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284340608)))];
+            tensor<fp32, [1024]> module_layers_11_norm_conv_weight = const()[name = tensor<string, []>("module_layers_11_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284344768)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_11_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(284348928))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286448256))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286446144)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_11_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286456512))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286466880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286465792)))];
+            tensor<fp32, [1024]> module_layers_11_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_11_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286471040)))];
+            tensor<fp32, [1024]> module_layers_11_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_11_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286475200)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_11_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(286479360))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287529088))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287528000)))];
+            tensor<fp32, [1024]> module_layers_11_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_11_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287533248)))];
+            tensor<fp32, [1024]> module_layers_11_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_11_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287537408)))];
+            tensor<fp32, [4096, 1024]> module_layers_11_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(287541568))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291740096))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291735936)))];
+            tensor<fp32, [1024, 4096]> module_layers_11_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_11_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(291756544))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295952000))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295950912)))];
+            tensor<fp32, [1024]> module_layers_11_norm_out_bias = const()[name = tensor<string, []>("module_layers_11_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295956160)))];
+            tensor<fp32, [1024]> module_layers_11_norm_out_weight = const()[name = tensor<string, []>("module_layers_11_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295960320)))];
+            tensor<fp32, [1024]> module_layers_12_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_12_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295964480)))];
+            tensor<fp32, [1024]> module_layers_12_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_12_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295968640)))];
+            tensor<fp32, [4096, 1024]> module_layers_12_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(295972800))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300171328))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300167168)))];
+            tensor<fp32, [1024, 4096]> module_layers_12_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(300187776))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304383232))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304382144)))];
+            tensor<fp32, [1024]> module_layers_12_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_12_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304387392)))];
+            tensor<fp32, [1024]> module_layers_12_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_12_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304391552)))];
+            tensor<fp32, [8, 128]> module_layers_12_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_12_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304395712)))];
+            tensor<fp32, [8, 128]> module_layers_12_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_12_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304399872)))];
+            tensor<fp32, [1024, 1024]> module_layers_12_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(304404032))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305453760))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305452672)))];
+            tensor<fp32, [1024, 1024]> module_layers_12_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(305457920))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306507648))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306506560)))];
+            tensor<fp32, [1024, 1024]> module_layers_12_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(306511808))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307561536))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307560448)))];
+            tensor<fp32, [1024, 1024]> module_layers_12_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(307565696))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308615424))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308614336)))];
+            tensor<fp32, [1024]> module_layers_12_norm_conv_bias = const()[name = tensor<string, []>("module_layers_12_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308619584)))];
+            tensor<fp32, [1024]> module_layers_12_norm_conv_weight = const()[name = tensor<string, []>("module_layers_12_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308623744)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_12_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(308627904))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310727232))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310725120)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_12_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310735488))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310745856))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310744768)))];
+            tensor<fp32, [1024]> module_layers_12_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_12_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310750016)))];
+            tensor<fp32, [1024]> module_layers_12_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_12_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310754176)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_12_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(310758336))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311808064))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311806976)))];
+            tensor<fp32, [1024]> module_layers_12_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_12_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311812224)))];
+            tensor<fp32, [1024]> module_layers_12_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_12_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311816384)))];
+            tensor<fp32, [4096, 1024]> module_layers_12_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(311820544))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316019072))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316014912)))];
+            tensor<fp32, [1024, 4096]> module_layers_12_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_12_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(316035520))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320230976))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320229888)))];
+            tensor<fp32, [1024]> module_layers_12_norm_out_bias = const()[name = tensor<string, []>("module_layers_12_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320235136)))];
+            tensor<fp32, [1024]> module_layers_12_norm_out_weight = const()[name = tensor<string, []>("module_layers_12_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320239296)))];
+            tensor<fp32, [1024]> module_layers_13_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_13_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320243456)))];
+            tensor<fp32, [1024]> module_layers_13_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_13_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320247616)))];
+            tensor<fp32, [4096, 1024]> module_layers_13_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(320251776))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324450304))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324446144)))];
+            tensor<fp32, [1024, 4096]> module_layers_13_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(324466752))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328662208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328661120)))];
+            tensor<fp32, [1024]> module_layers_13_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_13_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328666368)))];
+            tensor<fp32, [1024]> module_layers_13_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_13_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328670528)))];
+            tensor<fp32, [8, 128]> module_layers_13_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_13_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328674688)))];
+            tensor<fp32, [8, 128]> module_layers_13_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_13_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328678848)))];
+            tensor<fp32, [1024, 1024]> module_layers_13_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(328683008))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329732736))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329731648)))];
+            tensor<fp32, [1024, 1024]> module_layers_13_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(329736896))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330786624))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330785536)))];
+            tensor<fp32, [1024, 1024]> module_layers_13_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(330790784))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331840512))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331839424)))];
+            tensor<fp32, [1024, 1024]> module_layers_13_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(331844672))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332894400))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332893312)))];
+            tensor<fp32, [1024]> module_layers_13_norm_conv_bias = const()[name = tensor<string, []>("module_layers_13_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332898560)))];
+            tensor<fp32, [1024]> module_layers_13_norm_conv_weight = const()[name = tensor<string, []>("module_layers_13_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332902720)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_13_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(332906880))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335006208))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335004096)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_13_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335014464))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335024832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335023744)))];
+            tensor<fp32, [1024]> module_layers_13_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_13_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335028992)))];
+            tensor<fp32, [1024]> module_layers_13_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_13_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335033152)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_13_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(335037312))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336087040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336085952)))];
+            tensor<fp32, [1024]> module_layers_13_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_13_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336091200)))];
+            tensor<fp32, [1024]> module_layers_13_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_13_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336095360)))];
+            tensor<fp32, [4096, 1024]> module_layers_13_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(336099520))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340298048))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340293888)))];
+            tensor<fp32, [1024, 4096]> module_layers_13_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_13_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(340314496))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344509952))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344508864)))];
+            tensor<fp32, [1024]> module_layers_13_norm_out_bias = const()[name = tensor<string, []>("module_layers_13_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344514112)))];
+            tensor<fp32, [1024]> module_layers_13_norm_out_weight = const()[name = tensor<string, []>("module_layers_13_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344518272)))];
+            tensor<fp32, [1024]> module_layers_14_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_14_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344522432)))];
+            tensor<fp32, [1024]> module_layers_14_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_14_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344526592)))];
+            tensor<fp32, [4096, 1024]> module_layers_14_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(344530752))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348729280))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348725120)))];
+            tensor<fp32, [1024, 4096]> module_layers_14_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(348745728))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352941184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352940096)))];
+            tensor<fp32, [1024]> module_layers_14_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_14_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352945344)))];
+            tensor<fp32, [1024]> module_layers_14_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_14_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352949504)))];
+            tensor<fp32, [8, 128]> module_layers_14_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_14_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352953664)))];
+            tensor<fp32, [8, 128]> module_layers_14_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_14_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352957824)))];
+            tensor<fp32, [1024, 1024]> module_layers_14_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(352961984))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354011712))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354010624)))];
+            tensor<fp32, [1024, 1024]> module_layers_14_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(354015872))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355065600))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355064512)))];
+            tensor<fp32, [1024, 1024]> module_layers_14_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(355069760))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356119488))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356118400)))];
+            tensor<fp32, [1024, 1024]> module_layers_14_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(356123648))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357173376))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357172288)))];
+            tensor<fp32, [1024]> module_layers_14_norm_conv_bias = const()[name = tensor<string, []>("module_layers_14_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357177536)))];
+            tensor<fp32, [1024]> module_layers_14_norm_conv_weight = const()[name = tensor<string, []>("module_layers_14_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357181696)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_14_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(357185856))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359285184))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359283072)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_14_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359293440))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359303808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359302720)))];
+            tensor<fp32, [1024]> module_layers_14_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_14_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359307968)))];
+            tensor<fp32, [1024]> module_layers_14_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_14_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359312128)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_14_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(359316288))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360366016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360364928)))];
+            tensor<fp32, [1024]> module_layers_14_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_14_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360370176)))];
+            tensor<fp32, [1024]> module_layers_14_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_14_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360374336)))];
+            tensor<fp32, [4096, 1024]> module_layers_14_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(360378496))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364577024))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364572864)))];
+            tensor<fp32, [1024, 4096]> module_layers_14_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_14_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(364593472))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368788928))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368787840)))];
+            tensor<fp32, [1024]> module_layers_14_norm_out_bias = const()[name = tensor<string, []>("module_layers_14_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368793088)))];
+            tensor<fp32, [1024]> module_layers_14_norm_out_weight = const()[name = tensor<string, []>("module_layers_14_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368797248)))];
+            tensor<fp32, [1024]> module_layers_15_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_15_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368801408)))];
+            tensor<fp32, [1024]> module_layers_15_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_15_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368805568)))];
+            tensor<fp32, [4096, 1024]> module_layers_15_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368809728))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373008256))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373004096)))];
+            tensor<fp32, [1024, 4096]> module_layers_15_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(373024704))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377220160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377219072)))];
+            tensor<fp32, [1024]> module_layers_15_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_15_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377224320)))];
+            tensor<fp32, [1024]> module_layers_15_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_15_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377228480)))];
+            tensor<fp32, [8, 128]> module_layers_15_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_15_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377232640)))];
+            tensor<fp32, [8, 128]> module_layers_15_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_15_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377236800)))];
+            tensor<fp32, [1024, 1024]> module_layers_15_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(377240960))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378290688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378289600)))];
+            tensor<fp32, [1024, 1024]> module_layers_15_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(378294848))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379344576))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379343488)))];
+            tensor<fp32, [1024, 1024]> module_layers_15_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(379348736))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380398464))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380397376)))];
+            tensor<fp32, [1024, 1024]> module_layers_15_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(380402624))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381452352))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381451264)))];
+            tensor<fp32, [1024]> module_layers_15_norm_conv_bias = const()[name = tensor<string, []>("module_layers_15_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381456512)))];
+            tensor<fp32, [1024]> module_layers_15_norm_conv_weight = const()[name = tensor<string, []>("module_layers_15_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381460672)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_15_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381464832))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383564160))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383562048)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_15_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383572416))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383582784))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383581696)))];
+            tensor<fp32, [1024]> module_layers_15_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_15_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383586944)))];
+            tensor<fp32, [1024]> module_layers_15_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_15_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383591104)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_15_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(383595264))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384644992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384643904)))];
+            tensor<fp32, [1024]> module_layers_15_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_15_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384649152)))];
+            tensor<fp32, [1024]> module_layers_15_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_15_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384653312)))];
+            tensor<fp32, [4096, 1024]> module_layers_15_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(384657472))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388856000))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388851840)))];
+            tensor<fp32, [1024, 4096]> module_layers_15_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_15_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(388872448))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393067904))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393066816)))];
+            tensor<fp32, [1024]> module_layers_15_norm_out_bias = const()[name = tensor<string, []>("module_layers_15_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393072064)))];
+            tensor<fp32, [1024]> module_layers_15_norm_out_weight = const()[name = tensor<string, []>("module_layers_15_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393076224)))];
+            tensor<fp32, [1024]> module_layers_16_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_16_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393080384)))];
+            tensor<fp32, [1024]> module_layers_16_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_16_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393084544)))];
+            tensor<fp32, [4096, 1024]> module_layers_16_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(393088704))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397287232))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397283072)))];
+            tensor<fp32, [1024, 4096]> module_layers_16_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(397303680))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401499136))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401498048)))];
+            tensor<fp32, [1024]> module_layers_16_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_16_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401503296)))];
+            tensor<fp32, [1024]> module_layers_16_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_16_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401507456)))];
+            tensor<fp32, [8, 128]> module_layers_16_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_16_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401511616)))];
+            tensor<fp32, [8, 128]> module_layers_16_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_16_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401515776)))];
+            tensor<fp32, [1024, 1024]> module_layers_16_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(401519936))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402569664))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402568576)))];
+            tensor<fp32, [1024, 1024]> module_layers_16_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(402573824))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403623552))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403622464)))];
+            tensor<fp32, [1024, 1024]> module_layers_16_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(403627712))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404677440))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404676352)))];
+            tensor<fp32, [1024, 1024]> module_layers_16_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(404681600))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405731328))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405730240)))];
+            tensor<fp32, [1024]> module_layers_16_norm_conv_bias = const()[name = tensor<string, []>("module_layers_16_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405735488)))];
+            tensor<fp32, [1024]> module_layers_16_norm_conv_weight = const()[name = tensor<string, []>("module_layers_16_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405739648)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_16_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(405743808))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407843136))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407841024)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_16_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407851392))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407861760))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407860672)))];
+            tensor<fp32, [1024]> module_layers_16_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_16_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407865920)))];
+            tensor<fp32, [1024]> module_layers_16_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_16_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407870080)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_16_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(407874240))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408923968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408922880)))];
+            tensor<fp32, [1024]> module_layers_16_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_16_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408928128)))];
+            tensor<fp32, [1024]> module_layers_16_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_16_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408932288)))];
+            tensor<fp32, [4096, 1024]> module_layers_16_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(408936448))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(413134976))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(413130816)))];
+            tensor<fp32, [1024, 4096]> module_layers_16_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_16_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(413151424))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417346880))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417345792)))];
+            tensor<fp32, [1024]> module_layers_16_norm_out_bias = const()[name = tensor<string, []>("module_layers_16_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417351040)))];
+            tensor<fp32, [1024]> module_layers_16_norm_out_weight = const()[name = tensor<string, []>("module_layers_16_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417355200)))];
+            tensor<fp32, [1024]> module_layers_17_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_17_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417359360)))];
+            tensor<fp32, [1024]> module_layers_17_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_17_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417363520)))];
+            tensor<fp32, [4096, 1024]> module_layers_17_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(417367680))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421566208))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421562048)))];
+            tensor<fp32, [1024, 4096]> module_layers_17_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(421582656))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425778112))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425777024)))];
+            tensor<fp32, [1024]> module_layers_17_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_17_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425782272)))];
+            tensor<fp32, [1024]> module_layers_17_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_17_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425786432)))];
+            tensor<fp32, [8, 128]> module_layers_17_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_17_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425790592)))];
+            tensor<fp32, [8, 128]> module_layers_17_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_17_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425794752)))];
+            tensor<fp32, [1024, 1024]> module_layers_17_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(425798912))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426848640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426847552)))];
+            tensor<fp32, [1024, 1024]> module_layers_17_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(426852800))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427902528))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427901440)))];
+            tensor<fp32, [1024, 1024]> module_layers_17_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(427906688))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428956416))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428955328)))];
+            tensor<fp32, [1024, 1024]> module_layers_17_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(428960576))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430010304))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430009216)))];
+            tensor<fp32, [1024]> module_layers_17_norm_conv_bias = const()[name = tensor<string, []>("module_layers_17_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430014464)))];
+            tensor<fp32, [1024]> module_layers_17_norm_conv_weight = const()[name = tensor<string, []>("module_layers_17_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430018624)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_17_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(430022784))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432122112))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432120000)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_17_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432130368))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432140736))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432139648)))];
+            tensor<fp32, [1024]> module_layers_17_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_17_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432144896)))];
+            tensor<fp32, [1024]> module_layers_17_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_17_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432149056)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_17_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(432153216))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433202944))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433201856)))];
+            tensor<fp32, [1024]> module_layers_17_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_17_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433207104)))];
+            tensor<fp32, [1024]> module_layers_17_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_17_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433211264)))];
+            tensor<fp32, [4096, 1024]> module_layers_17_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(433215424))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437413952))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437409792)))];
+            tensor<fp32, [1024, 4096]> module_layers_17_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_17_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(437430400))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441625856))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441624768)))];
+            tensor<fp32, [1024]> module_layers_17_norm_out_bias = const()[name = tensor<string, []>("module_layers_17_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441630016)))];
+            tensor<fp32, [1024]> module_layers_17_norm_out_weight = const()[name = tensor<string, []>("module_layers_17_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441634176)))];
+            tensor<fp32, [1024]> module_layers_18_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_18_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441638336)))];
+            tensor<fp32, [1024]> module_layers_18_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_18_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441642496)))];
+            tensor<fp32, [4096, 1024]> module_layers_18_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(441646656))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445845184))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445841024)))];
+            tensor<fp32, [1024, 4096]> module_layers_18_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(445861632))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450057088))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450056000)))];
+            tensor<fp32, [1024]> module_layers_18_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_18_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450061248)))];
+            tensor<fp32, [1024]> module_layers_18_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_18_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450065408)))];
+            tensor<fp32, [8, 128]> module_layers_18_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_18_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450069568)))];
+            tensor<fp32, [8, 128]> module_layers_18_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_18_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450073728)))];
+            tensor<fp32, [1024, 1024]> module_layers_18_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(450077888))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451127616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451126528)))];
+            tensor<fp32, [1024, 1024]> module_layers_18_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(451131776))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452181504))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452180416)))];
+            tensor<fp32, [1024, 1024]> module_layers_18_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452185664))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453235392))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453234304)))];
+            tensor<fp32, [1024, 1024]> module_layers_18_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(453239552))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454289280))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454288192)))];
+            tensor<fp32, [1024]> module_layers_18_norm_conv_bias = const()[name = tensor<string, []>("module_layers_18_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454293440)))];
+            tensor<fp32, [1024]> module_layers_18_norm_conv_weight = const()[name = tensor<string, []>("module_layers_18_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454297600)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_18_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454301760))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456401088))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456398976)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_18_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456409344))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456419712))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456418624)))];
+            tensor<fp32, [1024]> module_layers_18_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_18_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456423872)))];
+            tensor<fp32, [1024]> module_layers_18_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_18_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456428032)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_18_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(456432192))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457481920))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457480832)))];
+            tensor<fp32, [1024]> module_layers_18_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_18_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457486080)))];
+            tensor<fp32, [1024]> module_layers_18_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_18_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457490240)))];
+            tensor<fp32, [4096, 1024]> module_layers_18_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457494400))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461692928))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461688768)))];
+            tensor<fp32, [1024, 4096]> module_layers_18_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_18_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(461709376))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465904832))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465903744)))];
+            tensor<fp32, [1024]> module_layers_18_norm_out_bias = const()[name = tensor<string, []>("module_layers_18_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465908992)))];
+            tensor<fp32, [1024]> module_layers_18_norm_out_weight = const()[name = tensor<string, []>("module_layers_18_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465913152)))];
+            tensor<fp32, [1024]> module_layers_19_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_19_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465917312)))];
+            tensor<fp32, [1024]> module_layers_19_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_19_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465921472)))];
+            tensor<fp32, [4096, 1024]> module_layers_19_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(465925632))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(470124160))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(470120000)))];
+            tensor<fp32, [1024, 4096]> module_layers_19_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(470140608))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474336064))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474334976)))];
+            tensor<fp32, [1024]> module_layers_19_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_19_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474340224)))];
+            tensor<fp32, [1024]> module_layers_19_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_19_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474344384)))];
+            tensor<fp32, [8, 128]> module_layers_19_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_19_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474348544)))];
+            tensor<fp32, [8, 128]> module_layers_19_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_19_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474352704)))];
+            tensor<fp32, [1024, 1024]> module_layers_19_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(474356864))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475406592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475405504)))];
+            tensor<fp32, [1024, 1024]> module_layers_19_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(475410752))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476460480))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476459392)))];
+            tensor<fp32, [1024, 1024]> module_layers_19_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(476464640))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477514368))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477513280)))];
+            tensor<fp32, [1024, 1024]> module_layers_19_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(477518528))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478568256))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478567168)))];
+            tensor<fp32, [1024]> module_layers_19_norm_conv_bias = const()[name = tensor<string, []>("module_layers_19_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478572416)))];
+            tensor<fp32, [1024]> module_layers_19_norm_conv_weight = const()[name = tensor<string, []>("module_layers_19_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478576576)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_19_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(478580736))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480680064))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480677952)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_19_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480688320))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480698688))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480697600)))];
+            tensor<fp32, [1024]> module_layers_19_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_19_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480702848)))];
+            tensor<fp32, [1024]> module_layers_19_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_19_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480707008)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_19_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(480711168))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481760896))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481759808)))];
+            tensor<fp32, [1024]> module_layers_19_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_19_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481765056)))];
+            tensor<fp32, [1024]> module_layers_19_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_19_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481769216)))];
+            tensor<fp32, [4096, 1024]> module_layers_19_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(481773376))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485971904))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485967744)))];
+            tensor<fp32, [1024, 4096]> module_layers_19_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_19_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(485988352))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490183808))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490182720)))];
+            tensor<fp32, [1024]> module_layers_19_norm_out_bias = const()[name = tensor<string, []>("module_layers_19_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490187968)))];
+            tensor<fp32, [1024]> module_layers_19_norm_out_weight = const()[name = tensor<string, []>("module_layers_19_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490192128)))];
+            tensor<fp32, [1024]> module_layers_20_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_20_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490196288)))];
+            tensor<fp32, [1024]> module_layers_20_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_20_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490200448)))];
+            tensor<fp32, [4096, 1024]> module_layers_20_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(490204608))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494403136))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494398976)))];
+            tensor<fp32, [1024, 4096]> module_layers_20_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(494419584))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498615040))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498613952)))];
+            tensor<fp32, [1024]> module_layers_20_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_20_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498619200)))];
+            tensor<fp32, [1024]> module_layers_20_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_20_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498623360)))];
+            tensor<fp32, [8, 128]> module_layers_20_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_20_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498627520)))];
+            tensor<fp32, [8, 128]> module_layers_20_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_20_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498631680)))];
+            tensor<fp32, [1024, 1024]> module_layers_20_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(498635840))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(499685568))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(499684480)))];
+            tensor<fp32, [1024, 1024]> module_layers_20_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(499689728))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500739456))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500738368)))];
+            tensor<fp32, [1024, 1024]> module_layers_20_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(500743616))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501793344))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501792256)))];
+            tensor<fp32, [1024, 1024]> module_layers_20_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(501797504))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502847232))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502846144)))];
+            tensor<fp32, [1024]> module_layers_20_norm_conv_bias = const()[name = tensor<string, []>("module_layers_20_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502851392)))];
+            tensor<fp32, [1024]> module_layers_20_norm_conv_weight = const()[name = tensor<string, []>("module_layers_20_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502855552)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_20_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(502859712))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504959040))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504956928)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_20_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504967296))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504977664))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504976576)))];
+            tensor<fp32, [1024]> module_layers_20_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_20_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504981824)))];
+            tensor<fp32, [1024]> module_layers_20_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_20_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504985984)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_20_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(504990144))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506039872))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506038784)))];
+            tensor<fp32, [1024]> module_layers_20_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_20_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506044032)))];
+            tensor<fp32, [1024]> module_layers_20_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_20_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506048192)))];
+            tensor<fp32, [4096, 1024]> module_layers_20_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(506052352))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510250880))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510246720)))];
+            tensor<fp32, [1024, 4096]> module_layers_20_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_20_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(510267328))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514462784))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514461696)))];
+            tensor<fp32, [1024]> module_layers_20_norm_out_bias = const()[name = tensor<string, []>("module_layers_20_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514466944)))];
+            tensor<fp32, [1024]> module_layers_20_norm_out_weight = const()[name = tensor<string, []>("module_layers_20_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514471104)))];
+            tensor<fp32, [1024]> module_layers_21_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_21_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514475264)))];
+            tensor<fp32, [1024]> module_layers_21_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_21_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514479424)))];
+            tensor<fp32, [4096, 1024]> module_layers_21_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(514483584))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518682112))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518677952)))];
+            tensor<fp32, [1024, 4096]> module_layers_21_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(518698560))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522894016))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522892928)))];
+            tensor<fp32, [1024]> module_layers_21_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_21_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522898176)))];
+            tensor<fp32, [1024]> module_layers_21_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_21_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522902336)))];
+            tensor<fp32, [8, 128]> module_layers_21_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_21_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522906496)))];
+            tensor<fp32, [8, 128]> module_layers_21_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_21_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522910656)))];
+            tensor<fp32, [1024, 1024]> module_layers_21_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(522914816))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523964544))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523963456)))];
+            tensor<fp32, [1024, 1024]> module_layers_21_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(523968704))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(525018432))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(525017344)))];
+            tensor<fp32, [1024, 1024]> module_layers_21_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(525022592))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526072320))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526071232)))];
+            tensor<fp32, [1024, 1024]> module_layers_21_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526076480))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527126208))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527125120)))];
+            tensor<fp32, [1024]> module_layers_21_norm_conv_bias = const()[name = tensor<string, []>("module_layers_21_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527130368)))];
+            tensor<fp32, [1024]> module_layers_21_norm_conv_weight = const()[name = tensor<string, []>("module_layers_21_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527134528)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_21_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(527138688))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529238016))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529235904)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_21_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529246272))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529256640))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529255552)))];
+            tensor<fp32, [1024]> module_layers_21_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_21_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529260800)))];
+            tensor<fp32, [1024]> module_layers_21_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_21_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529264960)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_21_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(529269120))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530318848))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530317760)))];
+            tensor<fp32, [1024]> module_layers_21_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_21_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530323008)))];
+            tensor<fp32, [1024]> module_layers_21_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_21_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530327168)))];
+            tensor<fp32, [4096, 1024]> module_layers_21_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(530331328))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534529856))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534525696)))];
+            tensor<fp32, [1024, 4096]> module_layers_21_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_21_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(534546304))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538741760))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538740672)))];
+            tensor<fp32, [1024]> module_layers_21_norm_out_bias = const()[name = tensor<string, []>("module_layers_21_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538745920)))];
+            tensor<fp32, [1024]> module_layers_21_norm_out_weight = const()[name = tensor<string, []>("module_layers_21_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538750080)))];
+            tensor<fp32, [1024]> module_layers_22_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_22_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538754240)))];
+            tensor<fp32, [1024]> module_layers_22_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_22_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538758400)))];
+            tensor<fp32, [4096, 1024]> module_layers_22_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(538762560))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542961088))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542956928)))];
+            tensor<fp32, [1024, 4096]> module_layers_22_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(542977536))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547172992))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547171904)))];
+            tensor<fp32, [1024]> module_layers_22_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_22_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547177152)))];
+            tensor<fp32, [1024]> module_layers_22_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_22_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547181312)))];
+            tensor<fp32, [8, 128]> module_layers_22_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_22_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547185472)))];
+            tensor<fp32, [8, 128]> module_layers_22_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_22_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547189632)))];
+            tensor<fp32, [1024, 1024]> module_layers_22_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(547193792))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548243520))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548242432)))];
+            tensor<fp32, [1024, 1024]> module_layers_22_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(548247680))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549297408))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549296320)))];
+            tensor<fp32, [1024, 1024]> module_layers_22_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(549301568))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550351296))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550350208)))];
+            tensor<fp32, [1024, 1024]> module_layers_22_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(550355456))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551405184))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551404096)))];
+            tensor<fp32, [1024]> module_layers_22_norm_conv_bias = const()[name = tensor<string, []>("module_layers_22_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551409344)))];
+            tensor<fp32, [1024]> module_layers_22_norm_conv_weight = const()[name = tensor<string, []>("module_layers_22_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551413504)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_22_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(551417664))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553516992))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553514880)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_22_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553525248))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553535616))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553534528)))];
+            tensor<fp32, [1024]> module_layers_22_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_22_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553539776)))];
+            tensor<fp32, [1024]> module_layers_22_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_22_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553543936)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_22_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(553548096))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554597824))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554596736)))];
+            tensor<fp32, [1024]> module_layers_22_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_22_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554601984)))];
+            tensor<fp32, [1024]> module_layers_22_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_22_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554606144)))];
+            tensor<fp32, [4096, 1024]> module_layers_22_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(554610304))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558808832))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558804672)))];
+            tensor<fp32, [1024, 4096]> module_layers_22_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_22_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(558825280))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563020736))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563019648)))];
+            tensor<fp32, [1024]> module_layers_22_norm_out_bias = const()[name = tensor<string, []>("module_layers_22_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563024896)))];
+            tensor<fp32, [1024]> module_layers_22_norm_out_weight = const()[name = tensor<string, []>("module_layers_22_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563029056)))];
+            tensor<fp32, [1024]> module_layers_23_norm_feed_forward1_bias = const()[name = tensor<string, []>("module_layers_23_norm_feed_forward1_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563033216)))];
+            tensor<fp32, [1024]> module_layers_23_norm_feed_forward1_weight = const()[name = tensor<string, []>("module_layers_23_norm_feed_forward1_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563037376)))];
+            tensor<fp32, [4096, 1024]> module_layers_23_feed_forward1_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_feed_forward1_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(563041536))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(567240064))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(567235904)))];
+            tensor<fp32, [1024, 4096]> module_layers_23_feed_forward1_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_feed_forward1_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(567256512))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571451968))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571450880)))];
+            tensor<fp32, [1024]> module_layers_23_norm_self_att_bias = const()[name = tensor<string, []>("module_layers_23_norm_self_att_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571456128)))];
+            tensor<fp32, [1024]> module_layers_23_norm_self_att_weight = const()[name = tensor<string, []>("module_layers_23_norm_self_att_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571460288)))];
+            tensor<fp32, [8, 128]> module_layers_23_self_attn_pos_bias_v = const()[name = tensor<string, []>("module_layers_23_self_attn_pos_bias_v"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571464448)))];
+            tensor<fp32, [8, 128]> module_layers_23_self_attn_pos_bias_u = const()[name = tensor<string, []>("module_layers_23_self_attn_pos_bias_u"), val = tensor<fp32, [8, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571468608)))];
+            tensor<fp32, [1024, 1024]> module_layers_23_self_attn_linear_q_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_self_attn_linear_q_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(571472768))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572522496))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572521408)))];
+            tensor<fp32, [1024, 1024]> module_layers_23_self_attn_linear_k_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_self_attn_linear_k_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(572526656))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(573576384))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(573575296)))];
+            tensor<fp32, [1024, 1024]> module_layers_23_self_attn_linear_v_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_self_attn_linear_v_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(573580544))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(574630272))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(574629184)))];
+            tensor<fp32, [1024, 1024]> module_layers_23_self_attn_linear_out_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_self_attn_linear_out_weight_quantized"), quantized_data = tensor<int8, [1024, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(574634432))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575684160))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575683072)))];
+            tensor<fp32, [1024]> module_layers_23_norm_conv_bias = const()[name = tensor<string, []>("module_layers_23_norm_conv_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575688320)))];
+            tensor<fp32, [1024]> module_layers_23_norm_conv_weight = const()[name = tensor<string, []>("module_layers_23_norm_conv_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575692480)))];
+            tensor<fp32, [2048, 1024, 1]> module_layers_23_conv_pointwise_conv1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_conv_pointwise_conv1_weight_quantized"), quantized_data = tensor<int8, [2048, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(575696640))), scale = tensor<fp32, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577795968))), zero_point = tensor<int8, [2048]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577793856)))];
+            tensor<fp32, [1024, 1, 9]> module_layers_23_conv_depthwise_conv_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_conv_depthwise_conv_weight_quantized"), quantized_data = tensor<int8, [1024, 1, 9]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577804224))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577814592))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577813504)))];
+            tensor<fp32, [1024]> module_layers_23_conv_batch_norm_bias = const()[name = tensor<string, []>("module_layers_23_conv_batch_norm_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577818752)))];
+            tensor<fp32, [1024]> module_layers_23_conv_batch_norm_weight = const()[name = tensor<string, []>("module_layers_23_conv_batch_norm_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577822912)))];
+            tensor<fp32, [1024, 1024, 1]> module_layers_23_conv_pointwise_conv2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_conv_pointwise_conv2_weight_quantized"), quantized_data = tensor<int8, [1024, 1024, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(577827072))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578876800))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578875712)))];
+            tensor<fp32, [1024]> module_layers_23_norm_feed_forward2_bias = const()[name = tensor<string, []>("module_layers_23_norm_feed_forward2_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578880960)))];
+            tensor<fp32, [1024]> module_layers_23_norm_feed_forward2_weight = const()[name = tensor<string, []>("module_layers_23_norm_feed_forward2_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578885120)))];
+            tensor<fp32, [4096, 1024]> module_layers_23_feed_forward2_linear1_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_feed_forward2_linear1_weight_quantized"), quantized_data = tensor<int8, [4096, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(578889280))), scale = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(583087808))), zero_point = tensor<int8, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(583083648)))];
+            tensor<fp32, [1024, 4096]> module_layers_23_feed_forward2_linear2_weight_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(0), name = tensor<string, []>("module_layers_23_feed_forward2_linear2_weight_quantized"), quantized_data = tensor<int8, [1024, 4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(583104256))), scale = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587299712))), zero_point = tensor<int8, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587298624)))];
+            tensor<fp32, [1024]> module_layers_23_norm_out_bias = const()[name = tensor<string, []>("module_layers_23_norm_out_bias"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587303872)))];
+            tensor<fp32, [1024]> module_layers_23_norm_out_weight = const()[name = tensor<string, []>("module_layers_23_norm_out_weight"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587308032)))];
+            tensor<int32, [4]> value_3_perm_0 = const()[name = tensor<string, []>("value_3_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> value_5_perm_0 = const()[name = tensor<string, []>("value_5_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<fp32, []> var_38 = const()[name = tensor<string, []>("op_38"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
+            tensor<fp32, []> var_40 = const()[name = tensor<string, []>("op_40"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<fp32, []> var_41 = const()[name = tensor<string, []>("op_41"), val = tensor<fp32, []>(-0x1.388p+13)];
+            tensor<int32, []> var_56 = const()[name = tensor<string, []>("op_56"), val = tensor<int32, []>(-1)];
+            tensor<int32, []> var_65 = const()[name = tensor<string, []>("op_65"), val = tensor<int32, []>(1)];
+            tensor<int32, [3]> x_1_perm_0 = const()[name = tensor<string, []>("x_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> tensor_1_axes_0 = const()[name = tensor<string, []>("tensor_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 25, 128]> x_1 = transpose(perm = x_1_perm_0, x = mel)[name = tensor<string, []>("transpose_367")];
+            tensor<fp32, [1, 1, 25, 128]> tensor_1 = expand_dims(axes = tensor_1_axes_0, x = x_1)[name = tensor<string, []>("tensor_1")];
+            tensor<string, []> current_lengths_1_dtype_0 = const()[name = tensor<string, []>("current_lengths_1_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1, 25]> expand_dims_0 = const()[name = tensor<string, []>("expand_dims_0"), val = tensor<int32, [1, 25]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24]])];
+            tensor<int32, [1]> var_134_axes_0 = const()[name = tensor<string, []>("op_134_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_134 = expand_dims(axes = var_134_axes_0, x = mel_length)[name = tensor<string, []>("op_134")];
+            tensor<bool, [1, 25]> time_mask_1 = less(x = expand_dims_0, y = var_134)[name = tensor<string, []>("time_mask_1")];
+            tensor<int32, [1]> var_136_axes_0 = const()[name = tensor<string, []>("op_136_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 25, 1]> var_136 = expand_dims(axes = var_136_axes_0, x = time_mask_1)[name = tensor<string, []>("op_136")];
+            tensor<int32, [3]> var_138_reps_0 = const()[name = tensor<string, []>("op_138_reps_0"), val = tensor<int32, [3]>([1, 1, 128])];
+            tensor<bool, [1, 25, 128]> var_138 = tile(reps = var_138_reps_0, x = var_136)[name = tensor<string, []>("op_138")];
+            tensor<string, []> mask_1_dtype_0 = const()[name = tensor<string, []>("mask_1_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_144_axes_0 = const()[name = tensor<string, []>("op_144_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 25, 128]> mask_1 = cast(dtype = mask_1_dtype_0, x = var_138)[name = tensor<string, []>("cast_11")];
+            tensor<fp32, [1, 1, 25, 128]> var_144 = expand_dims(axes = var_144_axes_0, x = mask_1)[name = tensor<string, []>("op_144")];
+            tensor<fp32, [1, 1, 25, 128]> input_1 = mul(x = tensor_1, y = var_144)[name = tensor<string, []>("input_1")];
+            tensor<fp32, []> const_9 = const()[name = tensor<string, []>("const_9"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_3_mode_0 = const()[name = tensor<string, []>("input_3_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 1, 28, 131]> input_3 = pad(constant_val = const_9, mode = input_3_mode_0, pad = input_3_pad_0, x = input_1)[name = tensor<string, []>("input_3")];
+            tensor<string, []> tensor_3_pad_type_0 = const()[name = tensor<string, []>("tensor_3_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_3_strides_0 = const()[name = tensor<string, []>("tensor_3_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, [4]> tensor_3_pad_0 = const()[name = tensor<string, []>("tensor_3_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_3_dilations_0 = const()[name = tensor<string, []>("tensor_3_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_3_groups_0 = const()[name = tensor<string, []>("tensor_3_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 256, 13, 65]> tensor_3 = conv(bias = module_pre_encode_conv_0_bias, dilations = tensor_3_dilations_0, groups = tensor_3_groups_0, pad = tensor_3_pad_0, pad_type = tensor_3_pad_type_0, strides = tensor_3_strides_0, weight = module_pre_encode_conv_0_weight_quantized, x = input_3)[name = tensor<string, []>("tensor_3")];
+            tensor<fp32, []> var_157_promoted = const()[name = tensor<string, []>("op_157_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> current_lengths_1 = cast(dtype = current_lengths_1_dtype_0, x = mel_length)[name = tensor<string, []>("cast_10")];
+            tensor<fp32, [1]> var_158 = add(x = current_lengths_1, y = var_157_promoted)[name = tensor<string, []>("op_158")];
+            tensor<fp32, []> var_159_promoted = const()[name = tensor<string, []>("op_159_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> var_160 = add(x = var_158, y = var_159_promoted)[name = tensor<string, []>("op_160")];
+            tensor<fp32, []> var_161_promoted = const()[name = tensor<string, []>("op_161_promoted"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> var_162 = sub(x = var_160, y = var_161_promoted)[name = tensor<string, []>("op_162")];
+            tensor<fp32, []> var_53_promoted = const()[name = tensor<string, []>("op_53_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> floor_div_0 = floor_div(x = var_162, y = var_53_promoted)[name = tensor<string, []>("floor_div_0")];
+            tensor<fp32, []> var_164_promoted = const()[name = tensor<string, []>("op_164_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> current_lengths_3 = add(x = floor_div_0, y = var_164_promoted)[name = tensor<string, []>("current_lengths_3")];
+            tensor<string, []> lengths_19_dtype_0 = const()[name = tensor<string, []>("lengths_19_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 13]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<int32, [1, 13]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]])];
+            tensor<int32, [1]> var_173_axes_0 = const()[name = tensor<string, []>("op_173_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> lengths_19 = cast(dtype = lengths_19_dtype_0, x = current_lengths_3)[name = tensor<string, []>("cast_9")];
+            tensor<int32, [1, 1]> var_173 = expand_dims(axes = var_173_axes_0, x = lengths_19)[name = tensor<string, []>("op_173")];
+            tensor<bool, [1, 13]> time_mask_3 = less(x = expand_dims_1, y = var_173)[name = tensor<string, []>("time_mask_3")];
+            tensor<int32, [1]> var_175_axes_0 = const()[name = tensor<string, []>("op_175_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 13, 1]> var_175 = expand_dims(axes = var_175_axes_0, x = time_mask_3)[name = tensor<string, []>("op_175")];
+            tensor<int32, [3]> var_177_reps_0 = const()[name = tensor<string, []>("op_177_reps_0"), val = tensor<int32, [3]>([1, 1, 65])];
+            tensor<bool, [1, 13, 65]> var_177 = tile(reps = var_177_reps_0, x = var_175)[name = tensor<string, []>("op_177")];
+            tensor<string, []> mask_3_dtype_0 = const()[name = tensor<string, []>("mask_3_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_183_axes_0 = const()[name = tensor<string, []>("op_183_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 13, 65]> mask_3 = cast(dtype = mask_3_dtype_0, x = var_177)[name = tensor<string, []>("cast_8")];
+            tensor<fp32, [1, 1, 13, 65]> var_183 = expand_dims(axes = var_183_axes_0, x = mask_3)[name = tensor<string, []>("op_183")];
+            tensor<int32, [4]> expanded_mask_3_reps_0 = const()[name = tensor<string, []>("expanded_mask_3_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp32, [1, 256, 13, 65]> expanded_mask_3 = tile(reps = expanded_mask_3_reps_0, x = var_183)[name = tensor<string, []>("expanded_mask_3")];
+            tensor<fp32, [1, 256, 13, 65]> input_5 = mul(x = tensor_3, y = expanded_mask_3)[name = tensor<string, []>("input_5")];
+            tensor<fp32, [1, 256, 13, 65]> tensor_5 = relu(x = input_5)[name = tensor<string, []>("tensor_5")];
+            tensor<fp32, [1, 256, 13, 65]> input_7 = mul(x = tensor_5, y = expanded_mask_3)[name = tensor<string, []>("input_7")];
+            tensor<fp32, []> const_23 = const()[name = tensor<string, []>("const_23"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_9_mode_0 = const()[name = tensor<string, []>("input_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 256, 16, 68]> input_9 = pad(constant_val = const_23, mode = input_9_mode_0, pad = input_9_pad_0, x = input_7)[name = tensor<string, []>("input_9")];
+            tensor<string, []> tensor_7_pad_type_0 = const()[name = tensor<string, []>("tensor_7_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_7_strides_0 = const()[name = tensor<string, []>("tensor_7_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_7_groups_0 = const()[name = tensor<string, []>("tensor_7_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_7_pad_0 = const()[name = tensor<string, []>("tensor_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_7_dilations_0 = const()[name = tensor<string, []>("tensor_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp32, [1, 256, 7, 33]> tensor_7 = conv(bias = module_pre_encode_conv_2_bias, dilations = tensor_7_dilations_0, groups = tensor_7_groups_0, pad = tensor_7_pad_0, pad_type = tensor_7_pad_type_0, strides = tensor_7_strides_0, weight = module_pre_encode_conv_2_weight_quantized, x = input_9)[name = tensor<string, []>("tensor_7")];
+            tensor<fp32, []> var_205_promoted = const()[name = tensor<string, []>("op_205_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> var_206 = add(x = current_lengths_3, y = var_205_promoted)[name = tensor<string, []>("op_206")];
+            tensor<fp32, []> var_207_promoted = const()[name = tensor<string, []>("op_207_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> var_208 = add(x = var_206, y = var_207_promoted)[name = tensor<string, []>("op_208")];
+            tensor<fp32, []> var_209_promoted = const()[name = tensor<string, []>("op_209_promoted"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> var_210 = sub(x = var_208, y = var_209_promoted)[name = tensor<string, []>("op_210")];
+            tensor<fp32, []> var_53_promoted_1 = const()[name = tensor<string, []>("op_53_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> floor_div_1 = floor_div(x = var_210, y = var_53_promoted_1)[name = tensor<string, []>("floor_div_1")];
+            tensor<fp32, []> var_212_promoted = const()[name = tensor<string, []>("op_212_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> current_lengths_5 = add(x = floor_div_1, y = var_212_promoted)[name = tensor<string, []>("current_lengths_5")];
+            tensor<string, []> lengths_21_dtype_0 = const()[name = tensor<string, []>("lengths_21_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 7]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<int32, [1, 7]>([[0, 1, 2, 3, 4, 5, 6]])];
+            tensor<int32, [1]> var_221_axes_0 = const()[name = tensor<string, []>("op_221_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> lengths_21 = cast(dtype = lengths_21_dtype_0, x = current_lengths_5)[name = tensor<string, []>("cast_7")];
+            tensor<int32, [1, 1]> var_221 = expand_dims(axes = var_221_axes_0, x = lengths_21)[name = tensor<string, []>("op_221")];
+            tensor<bool, [1, 7]> time_mask_5 = less(x = expand_dims_2, y = var_221)[name = tensor<string, []>("time_mask_5")];
+            tensor<int32, [1]> var_223_axes_0 = const()[name = tensor<string, []>("op_223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 7, 1]> var_223 = expand_dims(axes = var_223_axes_0, x = time_mask_5)[name = tensor<string, []>("op_223")];
+            tensor<int32, [3]> var_225_reps_0 = const()[name = tensor<string, []>("op_225_reps_0"), val = tensor<int32, [3]>([1, 1, 33])];
+            tensor<bool, [1, 7, 33]> var_225 = tile(reps = var_225_reps_0, x = var_223)[name = tensor<string, []>("op_225")];
+            tensor<string, []> mask_5_dtype_0 = const()[name = tensor<string, []>("mask_5_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_231_axes_0 = const()[name = tensor<string, []>("op_231_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 7, 33]> mask_5 = cast(dtype = mask_5_dtype_0, x = var_225)[name = tensor<string, []>("cast_6")];
+            tensor<fp32, [1, 1, 7, 33]> var_231 = expand_dims(axes = var_231_axes_0, x = mask_5)[name = tensor<string, []>("op_231")];
+            tensor<int32, [4]> expanded_mask_7_reps_0 = const()[name = tensor<string, []>("expanded_mask_7_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp32, [1, 256, 7, 33]> expanded_mask_7 = tile(reps = expanded_mask_7_reps_0, x = var_231)[name = tensor<string, []>("expanded_mask_7")];
+            tensor<fp32, [1, 256, 7, 33]> input_11 = mul(x = tensor_7, y = expanded_mask_7)[name = tensor<string, []>("input_11")];
+            tensor<string, []> tensor_9_pad_type_0 = const()[name = tensor<string, []>("tensor_9_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_9_strides_0 = const()[name = tensor<string, []>("tensor_9_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_9_pad_0 = const()[name = tensor<string, []>("tensor_9_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_9_dilations_0 = const()[name = tensor<string, []>("tensor_9_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_9_groups_0 = const()[name = tensor<string, []>("tensor_9_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 256, 7, 33]> tensor_9 = conv(bias = module_pre_encode_conv_3_bias, dilations = tensor_9_dilations_0, groups = tensor_9_groups_0, pad = tensor_9_pad_0, pad_type = tensor_9_pad_type_0, strides = tensor_9_strides_0, weight = module_pre_encode_conv_3_weight_quantized, x = input_11)[name = tensor<string, []>("tensor_9")];
+            tensor<fp32, [1, 256, 7, 33]> input_13 = mul(x = tensor_9, y = expanded_mask_7)[name = tensor<string, []>("input_13")];
+            tensor<fp32, [1, 256, 7, 33]> tensor_11 = relu(x = input_13)[name = tensor<string, []>("tensor_11")];
+            tensor<fp32, [1, 256, 7, 33]> input_15 = mul(x = tensor_11, y = expanded_mask_7)[name = tensor<string, []>("input_15")];
+            tensor<fp32, []> const_41 = const()[name = tensor<string, []>("const_41"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 1, 2, 1])];
+            tensor<string, []> input_17_mode_0 = const()[name = tensor<string, []>("input_17_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 256, 10, 36]> input_17 = pad(constant_val = const_41, mode = input_17_mode_0, pad = input_17_pad_0, x = input_15)[name = tensor<string, []>("input_17")];
+            tensor<string, []> tensor_13_pad_type_0 = const()[name = tensor<string, []>("tensor_13_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_13_strides_0 = const()[name = tensor<string, []>("tensor_13_strides_0"), val = tensor<int32, [2]>([2, 2])];
+            tensor<int32, []> tensor_13_groups_0 = const()[name = tensor<string, []>("tensor_13_groups_0"), val = tensor<int32, []>(256)];
+            tensor<int32, [4]> tensor_13_pad_0 = const()[name = tensor<string, []>("tensor_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_13_dilations_0 = const()[name = tensor<string, []>("tensor_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<fp32, [1, 256, 4, 17]> tensor_13 = conv(bias = module_pre_encode_conv_5_bias, dilations = tensor_13_dilations_0, groups = tensor_13_groups_0, pad = tensor_13_pad_0, pad_type = tensor_13_pad_type_0, strides = tensor_13_strides_0, weight = module_pre_encode_conv_5_weight_quantized, x = input_17)[name = tensor<string, []>("tensor_13")];
+            tensor<fp32, []> var_268_promoted = const()[name = tensor<string, []>("op_268_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> var_269 = add(x = current_lengths_5, y = var_268_promoted)[name = tensor<string, []>("op_269")];
+            tensor<fp32, []> var_270_promoted = const()[name = tensor<string, []>("op_270_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> var_271 = add(x = var_269, y = var_270_promoted)[name = tensor<string, []>("op_271")];
+            tensor<fp32, []> var_272_promoted = const()[name = tensor<string, []>("op_272_promoted"), val = tensor<fp32, []>(0x1.8p+1)];
+            tensor<fp32, [1]> var_273 = sub(x = var_271, y = var_272_promoted)[name = tensor<string, []>("op_273")];
+            tensor<fp32, []> var_53_promoted_2 = const()[name = tensor<string, []>("op_53_promoted_2"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1]> floor_div_2 = floor_div(x = var_273, y = var_53_promoted_2)[name = tensor<string, []>("floor_div_2")];
+            tensor<fp32, []> var_275_promoted = const()[name = tensor<string, []>("op_275_promoted"), val = tensor<fp32, []>(0x1p+0)];
+            tensor<fp32, [1]> current_lengths = add(x = floor_div_2, y = var_275_promoted)[name = tensor<string, []>("current_lengths")];
+            tensor<string, []> lengths_dtype_0 = const()[name = tensor<string, []>("lengths_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<int32, [1, 4]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1, 4]>([[0, 1, 2, 3]])];
+            tensor<int32, [1]> var_284_axes_0 = const()[name = tensor<string, []>("op_284_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1]> lengths = cast(dtype = lengths_dtype_0, x = current_lengths)[name = tensor<string, []>("cast_5")];
+            tensor<int32, [1, 1]> var_284 = expand_dims(axes = var_284_axes_0, x = lengths)[name = tensor<string, []>("op_284")];
+            tensor<bool, [1, 4]> time_mask = less(x = expand_dims_3, y = var_284)[name = tensor<string, []>("time_mask")];
+            tensor<int32, [1]> var_286_axes_0 = const()[name = tensor<string, []>("op_286_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, [1, 4, 1]> var_286 = expand_dims(axes = var_286_axes_0, x = time_mask)[name = tensor<string, []>("op_286")];
+            tensor<int32, [3]> var_288_reps_0 = const()[name = tensor<string, []>("op_288_reps_0"), val = tensor<int32, [3]>([1, 1, 17])];
+            tensor<bool, [1, 4, 17]> var_288 = tile(reps = var_288_reps_0, x = var_286)[name = tensor<string, []>("op_288")];
+            tensor<string, []> mask_7_dtype_0 = const()[name = tensor<string, []>("mask_7_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<int32, [1]> var_294_axes_0 = const()[name = tensor<string, []>("op_294_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 4, 17]> mask_7 = cast(dtype = mask_7_dtype_0, x = var_288)[name = tensor<string, []>("cast_4")];
+            tensor<fp32, [1, 1, 4, 17]> var_294 = expand_dims(axes = var_294_axes_0, x = mask_7)[name = tensor<string, []>("op_294")];
+            tensor<int32, [4]> expanded_mask_13_reps_0 = const()[name = tensor<string, []>("expanded_mask_13_reps_0"), val = tensor<int32, [4]>([1, 256, 1, 1])];
+            tensor<fp32, [1, 256, 4, 17]> expanded_mask_13 = tile(reps = expanded_mask_13_reps_0, x = var_294)[name = tensor<string, []>("expanded_mask_13")];
+            tensor<fp32, [1, 256, 4, 17]> input_19 = mul(x = tensor_13, y = expanded_mask_13)[name = tensor<string, []>("input_19")];
+            tensor<string, []> tensor_15_pad_type_0 = const()[name = tensor<string, []>("tensor_15_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> tensor_15_strides_0 = const()[name = tensor<string, []>("tensor_15_strides_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, [4]> tensor_15_pad_0 = const()[name = tensor<string, []>("tensor_15_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [2]> tensor_15_dilations_0 = const()[name = tensor<string, []>("tensor_15_dilations_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<int32, []> tensor_15_groups_0 = const()[name = tensor<string, []>("tensor_15_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 256, 4, 17]> tensor_15 = conv(bias = module_pre_encode_conv_6_bias, dilations = tensor_15_dilations_0, groups = tensor_15_groups_0, pad = tensor_15_pad_0, pad_type = tensor_15_pad_type_0, strides = tensor_15_strides_0, weight = module_pre_encode_conv_6_weight_quantized, x = input_19)[name = tensor<string, []>("tensor_15")];
+            tensor<fp32, [1, 256, 4, 17]> input_21 = mul(x = tensor_15, y = expanded_mask_13)[name = tensor<string, []>("input_21")];
+            tensor<fp32, [1, 256, 4, 17]> tensor_workaround = relu(x = input_21)[name = tensor<string, []>("tensor_workaround")];
+            tensor<fp32, [1, 256, 4, 17]> x_3 = mul(x = tensor_workaround, y = expanded_mask_13)[name = tensor<string, []>("x_3")];
+            tensor<int32, [4]> var_328_perm_0 = const()[name = tensor<string, []>("op_328_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_329 = const()[name = tensor<string, []>("op_329"), val = tensor<int32, [3]>([1, 4, -1])];
+            tensor<fp32, [1, 4, 256, 17]> var_328 = transpose(perm = var_328_perm_0, x = x_3)[name = tensor<string, []>("transpose_366")];
+            tensor<fp32, [1, 4, 4352]> input_23 = reshape(shape = var_329, x = var_328)[name = tensor<string, []>("input_23")];
+            tensor<fp32, [1, 4, 1024]> audio_signal_1 = linear(bias = module_pre_encode_out_bias, weight = module_pre_encode_out_weight_quantized, x = input_23)[name = tensor<string, []>("linear_0")];
+            tensor<int32, [3]> var_339_begin_0 = const()[name = tensor<string, []>("op_339_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_339_end_0 = const()[name = tensor<string, []>("op_339_end_0"), val = tensor<int32, [3]>([1, 4, 1024])];
+            tensor<bool, [3]> var_339_end_mask_0 = const()[name = tensor<string, []>("op_339_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 2, 1024]> var_339 = slice_by_index(begin = var_339_begin_0, end = var_339_end_0, end_mask = var_339_end_mask_0, x = audio_signal_1)[name = tensor<string, []>("op_339")];
+            tensor<int32, []> var_341 = const()[name = tensor<string, []>("op_341"), val = tensor<int32, []>(2)];
+            tensor<int32, [1]> var_342 = sub(x = lengths, y = var_341)[name = tensor<string, []>("op_342")];
+            tensor<fp32, []> const_61 = const()[name = tensor<string, []>("const_61"), val = tensor<fp32, []>(0x1.fffffep+127)];
+            tensor<string, []> var_342_promoted_dtype_0 = const()[name = tensor<string, []>("op_342_promoted_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> var_59_promoted = const()[name = tensor<string, []>("op_59_promoted"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<fp32, [1]> var_342_promoted = cast(dtype = var_342_promoted_dtype_0, x = var_342)[name = tensor<string, []>("cast_3")];
+            tensor<fp32, [1]> clip_0 = clip(alpha = var_59_promoted, beta = const_61, x = var_342_promoted)[name = tensor<string, []>("clip_0")];
+            tensor<int32, [1]> max_audio_length_1 = const()[name = tensor<string, []>("max_audio_length_1"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, []> var_358_promoted = const()[name = tensor<string, []>("op_358_promoted"), val = tensor<fp32, []>(0x1.18p+6)];
+            tensor<fp32, [1]> padding_length = add(x = clip_0, y = var_358_promoted)[name = tensor<string, []>("padding_length")];
+            tensor<int32, []> const_63 = const()[name = tensor<string, []>("const_63"), val = tensor<int32, []>(-1)];
+            tensor<int32, [1]> var_360 = mul(x = cache_len, y = const_63)[name = tensor<string, []>("op_360")];
+            tensor<int32, []> var_361 = const()[name = tensor<string, []>("op_361"), val = tensor<int32, []>(70)];
+            tensor<int32, [1]> offset = add(x = var_360, y = var_361)[name = tensor<string, []>("offset")];
+            tensor<int32, [1]> var_401_axes_0 = const()[name = tensor<string, []>("op_401_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 1]> var_401 = expand_dims(axes = var_401_axes_0, x = padding_length)[name = tensor<string, []>("op_401")];
+            tensor<fp32, [1, 72]> var_400_promoted = const()[name = tensor<string, []>("op_400_promoted"), val = tensor<fp32, [1, 72]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587312192)))];
+            tensor<bool, [1, 72]> pad_mask_1 = less(x = var_400_promoted, y = var_401)[name = tensor<string, []>("pad_mask_1")];
+            tensor<int32, [1, 72]> expand_dims_5 = const()[name = tensor<string, []>("expand_dims_5"), val = tensor<int32, [1, 72]>([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71]])];
+            tensor<int32, [1]> var_407_axes_0 = const()[name = tensor<string, []>("op_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<int32, [1, 1]> var_407 = expand_dims(axes = var_407_axes_0, x = offset)[name = tensor<string, []>("op_407")];
+            tensor<bool, [1, 72]> pad_mask_off = greater_equal(x = expand_dims_5, y = var_407)[name = tensor<string, []>("pad_mask_off")];
+            tensor<bool, [1, 72]> pad_mask_3 = logical_and(x = pad_mask_off, y = pad_mask_1)[name = tensor<string, []>("pad_mask_3")];
+            tensor<int32, [1]> var_410_axes_0 = const()[name = tensor<string, []>("op_410_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 72]> var_410 = expand_dims(axes = var_410_axes_0, x = pad_mask_3)[name = tensor<string, []>("op_410")];
+            tensor<int32, [3]> var_411 = const()[name = tensor<string, []>("op_411"), val = tensor<int32, [3]>([1, 72, 1])];
+            tensor<bool, [1, 72, 72]> pad_mask_for_att_mask_1 = tile(reps = var_411, x = var_410)[name = tensor<string, []>("pad_mask_for_att_mask_1")];
+            tensor<int32, [3]> var_413_perm_0 = const()[name = tensor<string, []>("op_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<bool, [1, 72, 72]> var_413 = transpose(perm = var_413_perm_0, x = pad_mask_for_att_mask_1)[name = tensor<string, []>("transpose_365")];
+            tensor<bool, [1, 72, 72]> pad_mask_for_att_mask = logical_and(x = pad_mask_for_att_mask_1, y = var_413)[name = tensor<string, []>("pad_mask_for_att_mask")];
+            tensor<bool, [1, 72, 72]> const_71 = const()[name = tensor<string, []>("const_71"), val = tensor<bool, [1, 72, 72]>([[[true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, false, false], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true], [true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true]]])];
+            tensor<bool, [1, 72, 72]> att_mask_9 = logical_and(x = pad_mask_for_att_mask, y = const_71)[name = tensor<string, []>("att_mask_9")];
+            tensor<bool, [1, 72, 72]> att_mask = logical_not(x = att_mask_9)[name = tensor<string, []>("att_mask")];
+            tensor<bool, [1, 72]> pad_mask_5 = logical_not(x = pad_mask_3)[name = tensor<string, []>("pad_mask_5")];
+            tensor<int32, [2]> pad_mask_begin_0 = const()[name = tensor<string, []>("pad_mask_begin_0"), val = tensor<int32, [2]>([0, 70])];
+            tensor<int32, [2]> pad_mask_end_0 = const()[name = tensor<string, []>("pad_mask_end_0"), val = tensor<int32, [2]>([1, 72])];
+            tensor<bool, [2]> pad_mask_end_mask_0 = const()[name = tensor<string, []>("pad_mask_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<bool, [1, 2]> pad_mask = slice_by_index(begin = pad_mask_begin_0, end = pad_mask_end_0, end_mask = pad_mask_end_mask_0, x = pad_mask_5)[name = tensor<string, []>("pad_mask")];
+            tensor<int32, [3]> mask_9_begin_0 = const()[name = tensor<string, []>("mask_9_begin_0"), val = tensor<int32, [3]>([0, 70, 0])];
+            tensor<int32, [3]> mask_9_end_0 = const()[name = tensor<string, []>("mask_9_end_0"), val = tensor<int32, [3]>([1, 72, 72])];
+            tensor<bool, [3]> mask_9_end_mask_0 = const()[name = tensor<string, []>("mask_9_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<bool, [1, 2, 72]> mask_9 = slice_by_index(begin = mask_9_begin_0, end = mask_9_end_0, end_mask = mask_9_end_mask_0, x = att_mask)[name = tensor<string, []>("mask_9")];
+            tensor<int32, [4]> cache_1_begin_0 = const()[name = tensor<string, []>("cache_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_1_end_0 = const()[name = tensor<string, []>("cache_1_end_0"), val = tensor<int32, [4]>([1, 1, 70, 1024])];
+            tensor<bool, [4]> cache_1_end_mask_0 = const()[name = tensor<string, []>("cache_1_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_1_squeeze_mask_0 = const()[name = tensor<string, []>("cache_1_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [24, 1, 70, 1024]> value_3 = transpose(perm = value_3_perm_0, x = cache_channel)[name = tensor<string, []>("transpose_364")];
+            tensor<fp32, [1, 70, 1024]> cache_1 = slice_by_index(begin = cache_1_begin_0, end = cache_1_end_0, end_mask = cache_1_end_mask_0, squeeze_mask = cache_1_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_1")];
+            tensor<int32, [4]> cache_3_begin_0 = const()[name = tensor<string, []>("cache_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> cache_3_end_0 = const()[name = tensor<string, []>("cache_3_end_0"), val = tensor<int32, [4]>([1, 1, 1024, 8])];
+            tensor<bool, [4]> cache_3_end_mask_0 = const()[name = tensor<string, []>("cache_3_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_3_squeeze_mask_0 = const()[name = tensor<string, []>("cache_3_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [24, 1, 1024, 8]> value_5 = transpose(perm = value_5_perm_0, x = cache_time)[name = tensor<string, []>("transpose_363")];
+            tensor<fp32, [1, 1024, 8]> cache_3 = slice_by_index(begin = cache_3_begin_0, end = cache_3_end_0, end_mask = cache_3_end_mask_0, squeeze_mask = cache_3_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_3")];
+            tensor<int32, [1]> input_27_axes_0 = const()[name = tensor<string, []>("input_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_27 = layer_norm(axes = input_27_axes_0, beta = module_layers_0_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_0_norm_feed_forward1_weight, x = var_339)[name = tensor<string, []>("input_27")];
+            tensor<fp32, [4096]> linear_1_bias_0 = const()[name = tensor<string, []>("linear_1_bias_0"), val = tensor<fp32, [4096]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587312576)))];
+            tensor<fp32, [1, 2, 4096]> input_29 = linear(bias = linear_1_bias_0, weight = module_layers_0_feed_forward1_linear1_weight_quantized, x = input_27)[name = tensor<string, []>("linear_1")];
+            tensor<fp32, [1, 2, 4096]> input_31 = silu(x = input_29)[name = tensor<string, []>("input_31")];
+            tensor<fp32, [1024]> linear_2_bias_0 = const()[name = tensor<string, []>("linear_2_bias_0"), val = tensor<fp32, [1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587329024)))];
+            tensor<fp32, [1, 2, 1024]> input_35 = linear(bias = linear_2_bias_0, weight = module_layers_0_feed_forward1_linear2_weight_quantized, x = input_31)[name = tensor<string, []>("linear_2")];
+            tensor<fp32, []> var_450 = const()[name = tensor<string, []>("op_450"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_451 = mul(x = input_35, y = var_450)[name = tensor<string, []>("op_451")];
+            tensor<fp32, [1, 2, 1024]> input_37 = add(x = var_339, y = var_451)[name = tensor<string, []>("input_37")];
+            tensor<int32, [1]> key_1_axes_0 = const()[name = tensor<string, []>("key_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_1 = layer_norm(axes = key_1_axes_0, beta = module_layers_0_norm_self_att_bias, epsilon = var_38, gamma = module_layers_0_norm_self_att_weight, x = input_37)[name = tensor<string, []>("key_1")];
+            tensor<bool, []> input_39_interleave_0 = const()[name = tensor<string, []>("input_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_39 = concat(axis = var_65, interleave = input_39_interleave_0, values = (cache_1, key_1))[name = tensor<string, []>("input_39")];
+            tensor<int32, [3]> var_473_begin_0 = const()[name = tensor<string, []>("op_473_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_473_end_0 = const()[name = tensor<string, []>("op_473_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_473_end_mask_0 = const()[name = tensor<string, []>("op_473_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_473 = slice_by_index(begin = var_473_begin_0, end = var_473_end_0, end_mask = var_473_end_mask_0, x = cache_1)[name = tensor<string, []>("op_473")];
+            tensor<bool, []> var_479_interleave_0 = const()[name = tensor<string, []>("op_479_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_479 = concat(axis = var_65, interleave = var_479_interleave_0, values = (var_473, key_1))[name = tensor<string, []>("op_479")];
+            tensor<fp32, [1, 2, 1024]> var_482 = linear(bias = linear_2_bias_0, weight = module_layers_0_self_attn_linear_q_weight_quantized, x = key_1)[name = tensor<string, []>("linear_3")];
+            tensor<int32, [4]> var_483 = const()[name = tensor<string, []>("op_483"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_1 = reshape(shape = var_483, x = var_482)[name = tensor<string, []>("q_1")];
+            tensor<fp32, [1, 72, 1024]> var_486 = linear(bias = linear_2_bias_0, weight = module_layers_0_self_attn_linear_k_weight_quantized, x = input_39)[name = tensor<string, []>("linear_4")];
+            tensor<int32, [4]> var_487 = const()[name = tensor<string, []>("op_487"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_1 = reshape(shape = var_487, x = var_486)[name = tensor<string, []>("k_1")];
+            tensor<fp32, [1, 72, 1024]> var_490 = linear(bias = linear_2_bias_0, weight = module_layers_0_self_attn_linear_v_weight_quantized, x = input_39)[name = tensor<string, []>("linear_5")];
+            tensor<int32, [4]> var_491 = const()[name = tensor<string, []>("op_491"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_1 = reshape(shape = var_491, x = var_490)[name = tensor<string, []>("v_1")];
+            tensor<int32, [4]> value_9_perm_0 = const()[name = tensor<string, []>("value_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_503 = add(x = q_1, y = module_layers_0_self_attn_pos_bias_u)[name = tensor<string, []>("op_503")];
+            tensor<fp32, [1, 2, 8, 128]> var_505 = add(x = q_1, y = module_layers_0_self_attn_pos_bias_v)[name = tensor<string, []>("op_505")];
+            tensor<int32, [4]> q_with_bias_v_1_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_507_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_507_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587333184))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587479936))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587479680)))];
+            tensor<bool, []> x_7_transpose_x_0 = const()[name = tensor<string, []>("x_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_7_transpose_y_0 = const()[name = tensor<string, []>("x_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_1 = transpose(perm = q_with_bias_v_1_perm_0, x = var_505)[name = tensor<string, []>("transpose_362")];
+            tensor<fp32, [1, 8, 2, 143]> x_7 = matmul(transpose_x = x_7_transpose_x_0, transpose_y = x_7_transpose_y_0, x = q_with_bias_v_1, y = op_507_quantized)[name = tensor<string, []>("x_7")];
+            tensor<fp32, []> const_79 = const()[name = tensor<string, []>("const_79"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_9_pad_0 = const()[name = tensor<string, []>("x_9_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_9_mode_0 = const()[name = tensor<string, []>("x_9_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_9 = pad(constant_val = const_79, mode = x_9_mode_0, pad = x_9_pad_0, x = x_7)[name = tensor<string, []>("x_9")];
+            tensor<int32, [4]> var_515 = const()[name = tensor<string, []>("op_515"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_11 = reshape(shape = var_515, x = x_9)[name = tensor<string, []>("x_11")];
+            tensor<int32, [4]> var_519_begin_0 = const()[name = tensor<string, []>("op_519_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_519_end_0 = const()[name = tensor<string, []>("op_519_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_519_end_mask_0 = const()[name = tensor<string, []>("op_519_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_519 = slice_by_index(begin = var_519_begin_0, end = var_519_end_0, end_mask = var_519_end_mask_0, x = x_11)[name = tensor<string, []>("op_519")];
+            tensor<int32, [4]> var_520 = const()[name = tensor<string, []>("op_520"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_1 = reshape(shape = var_520, x = var_519)[name = tensor<string, []>("matrix_bd_1")];
+            tensor<bool, []> matrix_ac_1_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_1_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_1_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_96_perm_0 = const()[name = tensor<string, []>("transpose_96_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_97_perm_0 = const()[name = tensor<string, []>("transpose_97_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_97 = transpose(perm = transpose_97_perm_0, x = k_1)[name = tensor<string, []>("transpose_360")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_96 = transpose(perm = transpose_96_perm_0, x = var_503)[name = tensor<string, []>("transpose_361")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_1 = matmul(transpose_x = matrix_ac_1_transpose_x_0, transpose_y = matrix_ac_1_transpose_y_0, x = transpose_96, y = transpose_97)[name = tensor<string, []>("matrix_ac_1")];
+            tensor<int32, [4]> matrix_bd_3_begin_0 = const()[name = tensor<string, []>("matrix_bd_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_3_end_0 = const()[name = tensor<string, []>("matrix_bd_3_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_3_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_3 = slice_by_index(begin = matrix_bd_3_begin_0, end = matrix_bd_3_end_0, end_mask = matrix_bd_3_end_mask_0, x = matrix_bd_1)[name = tensor<string, []>("matrix_bd_3")];
+            tensor<fp32, [1, 8, 2, 72]> var_529 = add(x = matrix_ac_1, y = matrix_bd_3)[name = tensor<string, []>("op_529")];
+            tensor<fp32, []> _inversed_scores_1_y_0 = const()[name = tensor<string, []>("_inversed_scores_1_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_1 = mul(x = var_529, y = _inversed_scores_1_y_0)[name = tensor<string, []>("_inversed_scores_1")];
+            tensor<int32, [1]> mask_11_axes_0 = const()[name = tensor<string, []>("mask_11_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 2, 72]> mask_11 = expand_dims(axes = mask_11_axes_0, x = mask_9)[name = tensor<string, []>("mask_11")];
+            tensor<fp32, [1, 8, 2, 72]> scores_3 = select(a = var_41, b = _inversed_scores_1, cond = mask_11)[name = tensor<string, []>("scores_3")];
+            tensor<fp32, [1, 8, 2, 72]> var_535 = softmax(axis = var_56, x = scores_3)[name = tensor<string, []>("op_535")];
+            tensor<fp32, [1, 8, 2, 72]> input_41 = select(a = var_40, b = var_535, cond = mask_11)[name = tensor<string, []>("input_41")];
+            tensor<bool, []> x_13_transpose_x_0 = const()[name = tensor<string, []>("x_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_13_transpose_y_0 = const()[name = tensor<string, []>("x_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_9 = transpose(perm = value_9_perm_0, x = v_1)[name = tensor<string, []>("transpose_359")];
+            tensor<fp32, [1, 8, 2, 128]> x_13 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = input_41, y = value_9)[name = tensor<string, []>("x_13")];
+            tensor<int32, [4]> var_539_perm_0 = const()[name = tensor<string, []>("op_539_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_540 = const()[name = tensor<string, []>("op_540"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_539 = transpose(perm = var_539_perm_0, x = x_13)[name = tensor<string, []>("transpose_358")];
+            tensor<fp32, [1, 2, 1024]> input_43 = reshape(shape = var_540, x = var_539)[name = tensor<string, []>("input_43")];
+            tensor<fp32, [1, 2, 1024]> input_45 = linear(bias = linear_2_bias_0, weight = module_layers_0_self_attn_linear_out_weight_quantized, x = input_43)[name = tensor<string, []>("linear_7")];
+            tensor<fp32, [1, 2, 1024]> input_47 = add(x = input_37, y = input_45)[name = tensor<string, []>("input_47")];
+            tensor<int32, [1]> x_17_axes_0 = const()[name = tensor<string, []>("x_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_17 = layer_norm(axes = x_17_axes_0, beta = module_layers_0_norm_conv_bias, epsilon = var_38, gamma = module_layers_0_norm_conv_weight, x = input_47)[name = tensor<string, []>("x_17")];
+            tensor<int32, [3]> input_49_perm_0 = const()[name = tensor<string, []>("input_49_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_51_pad_type_0 = const()[name = tensor<string, []>("input_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_51_strides_0 = const()[name = tensor<string, []>("input_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_51_pad_0 = const()[name = tensor<string, []>("input_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_51_dilations_0 = const()[name = tensor<string, []>("input_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_51_groups_0 = const()[name = tensor<string, []>("input_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_49 = transpose(perm = input_49_perm_0, x = x_17)[name = tensor<string, []>("transpose_357")];
+            tensor<fp32, [1, 2048, 2]> input_51 = conv(dilations = input_51_dilations_0, groups = input_51_groups_0, pad = input_51_pad_0, pad_type = input_51_pad_type_0, strides = input_51_strides_0, weight = module_layers_0_conv_pointwise_conv1_weight_quantized, x = input_49)[name = tensor<string, []>("input_51")];
+            tensor<int32, []> x_19_split_num_splits_0 = const()[name = tensor<string, []>("x_19_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_19_split_axis_0 = const()[name = tensor<string, []>("x_19_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_19_split_0, tensor<fp32, [1, 1024, 2]> x_19_split_1 = split(axis = x_19_split_axis_0, num_splits = x_19_split_num_splits_0, x = input_51)[name = tensor<string, []>("x_19_split")];
+            tensor<fp32, [1, 1024, 2]> x_19_split_1_sigmoid = sigmoid(x = x_19_split_1)[name = tensor<string, []>("x_19_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_19 = mul(x = x_19_split_0, y = x_19_split_1_sigmoid)[name = tensor<string, []>("x_19")];
+            tensor<int32, [1]> var_565_axes_0 = const()[name = tensor<string, []>("op_565_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, 2]> var_565 = expand_dims(axes = var_565_axes_0, x = pad_mask)[name = tensor<string, []>("op_565")];
+            tensor<fp32, [1, 1024, 2]> input_53 = select(a = var_40, b = x_19, cond = var_565)[name = tensor<string, []>("input_53")];
+            tensor<bool, []> new_x_3_interleave_0 = const()[name = tensor<string, []>("new_x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_3 = concat(axis = var_56, interleave = new_x_3_interleave_0, values = (cache_3, input_53))[name = tensor<string, []>("new_x_3")];
+            tensor<int32, [3]> var_578_begin_0 = const()[name = tensor<string, []>("op_578_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_578_end_0 = const()[name = tensor<string, []>("op_578_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_578_end_mask_0 = const()[name = tensor<string, []>("op_578_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_578 = slice_by_index(begin = var_578_begin_0, end = var_578_end_0, end_mask = var_578_end_mask_0, x = new_x_3)[name = tensor<string, []>("op_578")];
+            tensor<string, []> x_21_pad_type_0 = const()[name = tensor<string, []>("x_21_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_21_groups_0 = const()[name = tensor<string, []>("x_21_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_21_strides_0 = const()[name = tensor<string, []>("x_21_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_21_pad_0 = const()[name = tensor<string, []>("x_21_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_21_dilations_0 = const()[name = tensor<string, []>("x_21_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_21 = conv(dilations = x_21_dilations_0, groups = x_21_groups_0, pad = x_21_pad_0, pad_type = x_21_pad_type_0, strides = x_21_strides_0, weight = module_layers_0_conv_depthwise_conv_weight_quantized, x = new_x_3)[name = tensor<string, []>("x_21")];
+            tensor<int32, [3]> input_55_perm_0 = const()[name = tensor<string, []>("input_55_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_23_axes_0 = const()[name = tensor<string, []>("x_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_55 = transpose(perm = input_55_perm_0, x = x_21)[name = tensor<string, []>("transpose_356")];
+            tensor<fp32, [1, 2, 1024]> x_23 = layer_norm(axes = x_23_axes_0, beta = module_layers_0_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_0_conv_batch_norm_weight, x = input_55)[name = tensor<string, []>("x_23")];
+            tensor<int32, [3]> input_57_perm_0 = const()[name = tensor<string, []>("input_57_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_57 = transpose(perm = input_57_perm_0, x = x_23)[name = tensor<string, []>("transpose_355")];
+            tensor<fp32, [1, 1024, 2]> input_59 = silu(x = input_57)[name = tensor<string, []>("input_59")];
+            tensor<string, []> x_25_pad_type_0 = const()[name = tensor<string, []>("x_25_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_25_strides_0 = const()[name = tensor<string, []>("x_25_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_25_pad_0 = const()[name = tensor<string, []>("x_25_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_25_dilations_0 = const()[name = tensor<string, []>("x_25_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_25_groups_0 = const()[name = tensor<string, []>("x_25_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_25 = conv(dilations = x_25_dilations_0, groups = x_25_groups_0, pad = x_25_pad_0, pad_type = x_25_pad_type_0, strides = x_25_strides_0, weight = module_layers_0_conv_pointwise_conv2_weight_quantized, x = input_59)[name = tensor<string, []>("x_25")];
+            tensor<int32, [3]> input_61_perm_0 = const()[name = tensor<string, []>("input_61_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_61 = transpose(perm = input_61_perm_0, x = x_25)[name = tensor<string, []>("transpose_354")];
+            tensor<fp32, [1, 2, 1024]> input_63 = add(x = input_47, y = input_61)[name = tensor<string, []>("input_63")];
+            tensor<int32, [1]> input_65_axes_0 = const()[name = tensor<string, []>("input_65_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_65 = layer_norm(axes = input_65_axes_0, beta = module_layers_0_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_0_norm_feed_forward2_weight, x = input_63)[name = tensor<string, []>("input_65")];
+            tensor<fp32, [1, 2, 4096]> input_67 = linear(bias = linear_1_bias_0, weight = module_layers_0_feed_forward2_linear1_weight_quantized, x = input_65)[name = tensor<string, []>("linear_8")];
+            tensor<fp32, [1, 2, 4096]> input_69 = silu(x = input_67)[name = tensor<string, []>("input_69")];
+            tensor<fp32, [1, 2, 1024]> input_73 = linear(bias = linear_2_bias_0, weight = module_layers_0_feed_forward2_linear2_weight_quantized, x = input_69)[name = tensor<string, []>("linear_9")];
+            tensor<fp32, []> var_619 = const()[name = tensor<string, []>("op_619"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_620 = mul(x = input_73, y = var_619)[name = tensor<string, []>("op_620")];
+            tensor<fp32, [1, 2, 1024]> input_75 = add(x = input_63, y = var_620)[name = tensor<string, []>("input_75")];
+            tensor<int32, [1]> input_77_axes_0 = const()[name = tensor<string, []>("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_77 = layer_norm(axes = input_77_axes_0, beta = module_layers_0_norm_out_bias, epsilon = var_38, gamma = module_layers_0_norm_out_weight, x = input_75)[name = tensor<string, []>("input_77")];
+            tensor<int32, [4]> cache_5_begin_0 = const()[name = tensor<string, []>("cache_5_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_5_end_0 = const()[name = tensor<string, []>("cache_5_end_0"), val = tensor<int32, [4]>([2, 1, 70, 1024])];
+            tensor<bool, [4]> cache_5_end_mask_0 = const()[name = tensor<string, []>("cache_5_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_5_squeeze_mask_0 = const()[name = tensor<string, []>("cache_5_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_5 = slice_by_index(begin = cache_5_begin_0, end = cache_5_end_0, end_mask = cache_5_end_mask_0, squeeze_mask = cache_5_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_5")];
+            tensor<int32, [4]> cache_7_begin_0 = const()[name = tensor<string, []>("cache_7_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
+            tensor<int32, [4]> cache_7_end_0 = const()[name = tensor<string, []>("cache_7_end_0"), val = tensor<int32, [4]>([2, 1, 1024, 8])];
+            tensor<bool, [4]> cache_7_end_mask_0 = const()[name = tensor<string, []>("cache_7_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_7_squeeze_mask_0 = const()[name = tensor<string, []>("cache_7_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_7 = slice_by_index(begin = cache_7_begin_0, end = cache_7_end_0, end_mask = cache_7_end_mask_0, squeeze_mask = cache_7_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_7")];
+            tensor<int32, [1]> input_79_axes_0 = const()[name = tensor<string, []>("input_79_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_79 = layer_norm(axes = input_79_axes_0, beta = module_layers_1_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_1_norm_feed_forward1_weight, x = input_77)[name = tensor<string, []>("input_79")];
+            tensor<fp32, [1, 2, 4096]> input_81 = linear(bias = linear_1_bias_0, weight = module_layers_1_feed_forward1_linear1_weight_quantized, x = input_79)[name = tensor<string, []>("linear_10")];
+            tensor<fp32, [1, 2, 4096]> input_83 = silu(x = input_81)[name = tensor<string, []>("input_83")];
+            tensor<fp32, [1, 2, 1024]> input_87 = linear(bias = linear_2_bias_0, weight = module_layers_1_feed_forward1_linear2_weight_quantized, x = input_83)[name = tensor<string, []>("linear_11")];
+            tensor<fp32, []> var_654 = const()[name = tensor<string, []>("op_654"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_655 = mul(x = input_87, y = var_654)[name = tensor<string, []>("op_655")];
+            tensor<fp32, [1, 2, 1024]> input_89 = add(x = input_77, y = var_655)[name = tensor<string, []>("input_89")];
+            tensor<int32, [1]> key_3_axes_0 = const()[name = tensor<string, []>("key_3_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_3 = layer_norm(axes = key_3_axes_0, beta = module_layers_1_norm_self_att_bias, epsilon = var_38, gamma = module_layers_1_norm_self_att_weight, x = input_89)[name = tensor<string, []>("key_3")];
+            tensor<bool, []> input_91_interleave_0 = const()[name = tensor<string, []>("input_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_91 = concat(axis = var_65, interleave = input_91_interleave_0, values = (cache_5, key_3))[name = tensor<string, []>("input_91")];
+            tensor<int32, [3]> var_677_begin_0 = const()[name = tensor<string, []>("op_677_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_677_end_0 = const()[name = tensor<string, []>("op_677_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_677_end_mask_0 = const()[name = tensor<string, []>("op_677_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_677 = slice_by_index(begin = var_677_begin_0, end = var_677_end_0, end_mask = var_677_end_mask_0, x = cache_5)[name = tensor<string, []>("op_677")];
+            tensor<bool, []> var_683_interleave_0 = const()[name = tensor<string, []>("op_683_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_683 = concat(axis = var_65, interleave = var_683_interleave_0, values = (var_677, key_3))[name = tensor<string, []>("op_683")];
+            tensor<fp32, [1, 2, 1024]> var_686 = linear(bias = linear_2_bias_0, weight = module_layers_1_self_attn_linear_q_weight_quantized, x = key_3)[name = tensor<string, []>("linear_12")];
+            tensor<int32, [4]> var_687 = const()[name = tensor<string, []>("op_687"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_7 = reshape(shape = var_687, x = var_686)[name = tensor<string, []>("q_7")];
+            tensor<fp32, [1, 72, 1024]> var_690 = linear(bias = linear_2_bias_0, weight = module_layers_1_self_attn_linear_k_weight_quantized, x = input_91)[name = tensor<string, []>("linear_13")];
+            tensor<int32, [4]> var_691 = const()[name = tensor<string, []>("op_691"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_5 = reshape(shape = var_691, x = var_690)[name = tensor<string, []>("k_5")];
+            tensor<fp32, [1, 72, 1024]> var_694 = linear(bias = linear_2_bias_0, weight = module_layers_1_self_attn_linear_v_weight_quantized, x = input_91)[name = tensor<string, []>("linear_14")];
+            tensor<int32, [4]> var_695 = const()[name = tensor<string, []>("op_695"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_3 = reshape(shape = var_695, x = var_694)[name = tensor<string, []>("v_3")];
+            tensor<int32, [4]> value_11_perm_0 = const()[name = tensor<string, []>("value_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_707 = add(x = q_7, y = module_layers_1_self_attn_pos_bias_u)[name = tensor<string, []>("op_707")];
+            tensor<fp32, [1, 2, 8, 128]> var_709 = add(x = q_7, y = module_layers_1_self_attn_pos_bias_v)[name = tensor<string, []>("op_709")];
+            tensor<int32, [4]> q_with_bias_v_3_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_711_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_711_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587480576))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587627328))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587627072)))];
+            tensor<bool, []> x_33_transpose_x_0 = const()[name = tensor<string, []>("x_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_33_transpose_y_0 = const()[name = tensor<string, []>("x_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_3 = transpose(perm = q_with_bias_v_3_perm_0, x = var_709)[name = tensor<string, []>("transpose_353")];
+            tensor<fp32, [1, 8, 2, 143]> x_33 = matmul(transpose_x = x_33_transpose_x_0, transpose_y = x_33_transpose_y_0, x = q_with_bias_v_3, y = op_711_quantized)[name = tensor<string, []>("x_33")];
+            tensor<fp32, []> const_92 = const()[name = tensor<string, []>("const_92"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_35_pad_0 = const()[name = tensor<string, []>("x_35_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_35_mode_0 = const()[name = tensor<string, []>("x_35_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_35 = pad(constant_val = const_92, mode = x_35_mode_0, pad = x_35_pad_0, x = x_33)[name = tensor<string, []>("x_35")];
+            tensor<int32, [4]> var_719 = const()[name = tensor<string, []>("op_719"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_37 = reshape(shape = var_719, x = x_35)[name = tensor<string, []>("x_37")];
+            tensor<int32, [4]> var_723_begin_0 = const()[name = tensor<string, []>("op_723_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_723_end_0 = const()[name = tensor<string, []>("op_723_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_723_end_mask_0 = const()[name = tensor<string, []>("op_723_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_723 = slice_by_index(begin = var_723_begin_0, end = var_723_end_0, end_mask = var_723_end_mask_0, x = x_37)[name = tensor<string, []>("op_723")];
+            tensor<int32, [4]> var_724 = const()[name = tensor<string, []>("op_724"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_5 = reshape(shape = var_724, x = var_723)[name = tensor<string, []>("matrix_bd_5")];
+            tensor<bool, []> matrix_ac_3_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_3_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_3_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_98_perm_0 = const()[name = tensor<string, []>("transpose_98_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_99_perm_0 = const()[name = tensor<string, []>("transpose_99_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_99 = transpose(perm = transpose_99_perm_0, x = k_5)[name = tensor<string, []>("transpose_351")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_98 = transpose(perm = transpose_98_perm_0, x = var_707)[name = tensor<string, []>("transpose_352")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_3 = matmul(transpose_x = matrix_ac_3_transpose_x_0, transpose_y = matrix_ac_3_transpose_y_0, x = transpose_98, y = transpose_99)[name = tensor<string, []>("matrix_ac_3")];
+            tensor<int32, [4]> matrix_bd_7_begin_0 = const()[name = tensor<string, []>("matrix_bd_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_7_end_0 = const()[name = tensor<string, []>("matrix_bd_7_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_7_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_7 = slice_by_index(begin = matrix_bd_7_begin_0, end = matrix_bd_7_end_0, end_mask = matrix_bd_7_end_mask_0, x = matrix_bd_5)[name = tensor<string, []>("matrix_bd_7")];
+            tensor<fp32, [1, 8, 2, 72]> var_733 = add(x = matrix_ac_3, y = matrix_bd_7)[name = tensor<string, []>("op_733")];
+            tensor<fp32, []> _inversed_scores_5_y_0 = const()[name = tensor<string, []>("_inversed_scores_5_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_5 = mul(x = var_733, y = _inversed_scores_5_y_0)[name = tensor<string, []>("_inversed_scores_5")];
+            tensor<fp32, [1, 8, 2, 72]> scores_7 = select(a = var_41, b = _inversed_scores_5, cond = mask_11)[name = tensor<string, []>("scores_7")];
+            tensor<fp32, [1, 8, 2, 72]> var_739 = softmax(axis = var_56, x = scores_7)[name = tensor<string, []>("op_739")];
+            tensor<fp32, [1, 8, 2, 72]> input_93 = select(a = var_40, b = var_739, cond = mask_11)[name = tensor<string, []>("input_93")];
+            tensor<bool, []> x_39_transpose_x_0 = const()[name = tensor<string, []>("x_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_39_transpose_y_0 = const()[name = tensor<string, []>("x_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_11 = transpose(perm = value_11_perm_0, x = v_3)[name = tensor<string, []>("transpose_350")];
+            tensor<fp32, [1, 8, 2, 128]> x_39 = matmul(transpose_x = x_39_transpose_x_0, transpose_y = x_39_transpose_y_0, x = input_93, y = value_11)[name = tensor<string, []>("x_39")];
+            tensor<int32, [4]> var_743_perm_0 = const()[name = tensor<string, []>("op_743_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_744 = const()[name = tensor<string, []>("op_744"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_743 = transpose(perm = var_743_perm_0, x = x_39)[name = tensor<string, []>("transpose_349")];
+            tensor<fp32, [1, 2, 1024]> input_95 = reshape(shape = var_744, x = var_743)[name = tensor<string, []>("input_95")];
+            tensor<fp32, [1, 2, 1024]> input_97 = linear(bias = linear_2_bias_0, weight = module_layers_1_self_attn_linear_out_weight_quantized, x = input_95)[name = tensor<string, []>("linear_16")];
+            tensor<fp32, [1, 2, 1024]> input_99 = add(x = input_89, y = input_97)[name = tensor<string, []>("input_99")];
+            tensor<int32, [1]> x_43_axes_0 = const()[name = tensor<string, []>("x_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_43 = layer_norm(axes = x_43_axes_0, beta = module_layers_1_norm_conv_bias, epsilon = var_38, gamma = module_layers_1_norm_conv_weight, x = input_99)[name = tensor<string, []>("x_43")];
+            tensor<int32, [3]> input_101_perm_0 = const()[name = tensor<string, []>("input_101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_103_pad_type_0 = const()[name = tensor<string, []>("input_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_103_strides_0 = const()[name = tensor<string, []>("input_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_103_pad_0 = const()[name = tensor<string, []>("input_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_103_dilations_0 = const()[name = tensor<string, []>("input_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_103_groups_0 = const()[name = tensor<string, []>("input_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_101 = transpose(perm = input_101_perm_0, x = x_43)[name = tensor<string, []>("transpose_348")];
+            tensor<fp32, [1, 2048, 2]> input_103 = conv(dilations = input_103_dilations_0, groups = input_103_groups_0, pad = input_103_pad_0, pad_type = input_103_pad_type_0, strides = input_103_strides_0, weight = module_layers_1_conv_pointwise_conv1_weight_quantized, x = input_101)[name = tensor<string, []>("input_103")];
+            tensor<int32, []> x_45_split_num_splits_0 = const()[name = tensor<string, []>("x_45_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_45_split_axis_0 = const()[name = tensor<string, []>("x_45_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_45_split_0, tensor<fp32, [1, 1024, 2]> x_45_split_1 = split(axis = x_45_split_axis_0, num_splits = x_45_split_num_splits_0, x = input_103)[name = tensor<string, []>("x_45_split")];
+            tensor<fp32, [1, 1024, 2]> x_45_split_1_sigmoid = sigmoid(x = x_45_split_1)[name = tensor<string, []>("x_45_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_45 = mul(x = x_45_split_0, y = x_45_split_1_sigmoid)[name = tensor<string, []>("x_45")];
+            tensor<fp32, [1, 1024, 2]> input_105 = select(a = var_40, b = x_45, cond = var_565)[name = tensor<string, []>("input_105")];
+            tensor<bool, []> new_x_7_interleave_0 = const()[name = tensor<string, []>("new_x_7_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_7 = concat(axis = var_56, interleave = new_x_7_interleave_0, values = (cache_7, input_105))[name = tensor<string, []>("new_x_7")];
+            tensor<int32, [3]> var_782_begin_0 = const()[name = tensor<string, []>("op_782_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_782_end_0 = const()[name = tensor<string, []>("op_782_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_782_end_mask_0 = const()[name = tensor<string, []>("op_782_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_782 = slice_by_index(begin = var_782_begin_0, end = var_782_end_0, end_mask = var_782_end_mask_0, x = new_x_7)[name = tensor<string, []>("op_782")];
+            tensor<string, []> x_47_pad_type_0 = const()[name = tensor<string, []>("x_47_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_47_groups_0 = const()[name = tensor<string, []>("x_47_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_47_strides_0 = const()[name = tensor<string, []>("x_47_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_47_pad_0 = const()[name = tensor<string, []>("x_47_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_47_dilations_0 = const()[name = tensor<string, []>("x_47_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_47 = conv(dilations = x_47_dilations_0, groups = x_47_groups_0, pad = x_47_pad_0, pad_type = x_47_pad_type_0, strides = x_47_strides_0, weight = module_layers_1_conv_depthwise_conv_weight_quantized, x = new_x_7)[name = tensor<string, []>("x_47")];
+            tensor<int32, [3]> input_107_perm_0 = const()[name = tensor<string, []>("input_107_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_49_axes_0 = const()[name = tensor<string, []>("x_49_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_107 = transpose(perm = input_107_perm_0, x = x_47)[name = tensor<string, []>("transpose_347")];
+            tensor<fp32, [1, 2, 1024]> x_49 = layer_norm(axes = x_49_axes_0, beta = module_layers_1_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_1_conv_batch_norm_weight, x = input_107)[name = tensor<string, []>("x_49")];
+            tensor<int32, [3]> input_109_perm_0 = const()[name = tensor<string, []>("input_109_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_109 = transpose(perm = input_109_perm_0, x = x_49)[name = tensor<string, []>("transpose_346")];
+            tensor<fp32, [1, 1024, 2]> input_111 = silu(x = input_109)[name = tensor<string, []>("input_111")];
+            tensor<string, []> x_51_pad_type_0 = const()[name = tensor<string, []>("x_51_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_51_strides_0 = const()[name = tensor<string, []>("x_51_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_51_pad_0 = const()[name = tensor<string, []>("x_51_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_51_dilations_0 = const()[name = tensor<string, []>("x_51_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_51_groups_0 = const()[name = tensor<string, []>("x_51_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_51 = conv(dilations = x_51_dilations_0, groups = x_51_groups_0, pad = x_51_pad_0, pad_type = x_51_pad_type_0, strides = x_51_strides_0, weight = module_layers_1_conv_pointwise_conv2_weight_quantized, x = input_111)[name = tensor<string, []>("x_51")];
+            tensor<int32, [3]> input_113_perm_0 = const()[name = tensor<string, []>("input_113_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_113 = transpose(perm = input_113_perm_0, x = x_51)[name = tensor<string, []>("transpose_345")];
+            tensor<fp32, [1, 2, 1024]> input_115 = add(x = input_99, y = input_113)[name = tensor<string, []>("input_115")];
+            tensor<int32, [1]> input_117_axes_0 = const()[name = tensor<string, []>("input_117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_117 = layer_norm(axes = input_117_axes_0, beta = module_layers_1_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_1_norm_feed_forward2_weight, x = input_115)[name = tensor<string, []>("input_117")];
+            tensor<fp32, [1, 2, 4096]> input_119 = linear(bias = linear_1_bias_0, weight = module_layers_1_feed_forward2_linear1_weight_quantized, x = input_117)[name = tensor<string, []>("linear_17")];
+            tensor<fp32, [1, 2, 4096]> input_121 = silu(x = input_119)[name = tensor<string, []>("input_121")];
+            tensor<fp32, [1, 2, 1024]> input_125 = linear(bias = linear_2_bias_0, weight = module_layers_1_feed_forward2_linear2_weight_quantized, x = input_121)[name = tensor<string, []>("linear_18")];
+            tensor<fp32, []> var_823 = const()[name = tensor<string, []>("op_823"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_824 = mul(x = input_125, y = var_823)[name = tensor<string, []>("op_824")];
+            tensor<fp32, [1, 2, 1024]> input_127 = add(x = input_115, y = var_824)[name = tensor<string, []>("input_127")];
+            tensor<int32, [1]> input_129_axes_0 = const()[name = tensor<string, []>("input_129_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_129 = layer_norm(axes = input_129_axes_0, beta = module_layers_1_norm_out_bias, epsilon = var_38, gamma = module_layers_1_norm_out_weight, x = input_127)[name = tensor<string, []>("input_129")];
+            tensor<int32, [4]> cache_9_begin_0 = const()[name = tensor<string, []>("cache_9_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_9_end_0 = const()[name = tensor<string, []>("cache_9_end_0"), val = tensor<int32, [4]>([3, 1, 70, 1024])];
+            tensor<bool, [4]> cache_9_end_mask_0 = const()[name = tensor<string, []>("cache_9_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_9_squeeze_mask_0 = const()[name = tensor<string, []>("cache_9_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_9 = slice_by_index(begin = cache_9_begin_0, end = cache_9_end_0, end_mask = cache_9_end_mask_0, squeeze_mask = cache_9_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_9")];
+            tensor<int32, [4]> cache_11_begin_0 = const()[name = tensor<string, []>("cache_11_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
+            tensor<int32, [4]> cache_11_end_0 = const()[name = tensor<string, []>("cache_11_end_0"), val = tensor<int32, [4]>([3, 1, 1024, 8])];
+            tensor<bool, [4]> cache_11_end_mask_0 = const()[name = tensor<string, []>("cache_11_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_11_squeeze_mask_0 = const()[name = tensor<string, []>("cache_11_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_11 = slice_by_index(begin = cache_11_begin_0, end = cache_11_end_0, end_mask = cache_11_end_mask_0, squeeze_mask = cache_11_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_11")];
+            tensor<int32, [1]> input_131_axes_0 = const()[name = tensor<string, []>("input_131_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_131 = layer_norm(axes = input_131_axes_0, beta = module_layers_2_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_2_norm_feed_forward1_weight, x = input_129)[name = tensor<string, []>("input_131")];
+            tensor<fp32, [1, 2, 4096]> input_133 = linear(bias = linear_1_bias_0, weight = module_layers_2_feed_forward1_linear1_weight_quantized, x = input_131)[name = tensor<string, []>("linear_19")];
+            tensor<fp32, [1, 2, 4096]> input_135 = silu(x = input_133)[name = tensor<string, []>("input_135")];
+            tensor<fp32, [1, 2, 1024]> input_139 = linear(bias = linear_2_bias_0, weight = module_layers_2_feed_forward1_linear2_weight_quantized, x = input_135)[name = tensor<string, []>("linear_20")];
+            tensor<fp32, []> var_858 = const()[name = tensor<string, []>("op_858"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_859 = mul(x = input_139, y = var_858)[name = tensor<string, []>("op_859")];
+            tensor<fp32, [1, 2, 1024]> input_141 = add(x = input_129, y = var_859)[name = tensor<string, []>("input_141")];
+            tensor<int32, [1]> key_5_axes_0 = const()[name = tensor<string, []>("key_5_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_5 = layer_norm(axes = key_5_axes_0, beta = module_layers_2_norm_self_att_bias, epsilon = var_38, gamma = module_layers_2_norm_self_att_weight, x = input_141)[name = tensor<string, []>("key_5")];
+            tensor<bool, []> input_143_interleave_0 = const()[name = tensor<string, []>("input_143_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_143 = concat(axis = var_65, interleave = input_143_interleave_0, values = (cache_9, key_5))[name = tensor<string, []>("input_143")];
+            tensor<int32, [3]> var_881_begin_0 = const()[name = tensor<string, []>("op_881_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_881_end_0 = const()[name = tensor<string, []>("op_881_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_881_end_mask_0 = const()[name = tensor<string, []>("op_881_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_881 = slice_by_index(begin = var_881_begin_0, end = var_881_end_0, end_mask = var_881_end_mask_0, x = cache_9)[name = tensor<string, []>("op_881")];
+            tensor<bool, []> var_887_interleave_0 = const()[name = tensor<string, []>("op_887_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_887 = concat(axis = var_65, interleave = var_887_interleave_0, values = (var_881, key_5))[name = tensor<string, []>("op_887")];
+            tensor<fp32, [1, 2, 1024]> var_890 = linear(bias = linear_2_bias_0, weight = module_layers_2_self_attn_linear_q_weight_quantized, x = key_5)[name = tensor<string, []>("linear_21")];
+            tensor<int32, [4]> var_891 = const()[name = tensor<string, []>("op_891"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_13 = reshape(shape = var_891, x = var_890)[name = tensor<string, []>("q_13")];
+            tensor<fp32, [1, 72, 1024]> var_894 = linear(bias = linear_2_bias_0, weight = module_layers_2_self_attn_linear_k_weight_quantized, x = input_143)[name = tensor<string, []>("linear_22")];
+            tensor<int32, [4]> var_895 = const()[name = tensor<string, []>("op_895"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_9 = reshape(shape = var_895, x = var_894)[name = tensor<string, []>("k_9")];
+            tensor<fp32, [1, 72, 1024]> var_898 = linear(bias = linear_2_bias_0, weight = module_layers_2_self_attn_linear_v_weight_quantized, x = input_143)[name = tensor<string, []>("linear_23")];
+            tensor<int32, [4]> var_899 = const()[name = tensor<string, []>("op_899"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_5 = reshape(shape = var_899, x = var_898)[name = tensor<string, []>("v_5")];
+            tensor<int32, [4]> value_13_perm_0 = const()[name = tensor<string, []>("value_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_911 = add(x = q_13, y = module_layers_2_self_attn_pos_bias_u)[name = tensor<string, []>("op_911")];
+            tensor<fp32, [1, 2, 8, 128]> var_913 = add(x = q_13, y = module_layers_2_self_attn_pos_bias_v)[name = tensor<string, []>("op_913")];
+            tensor<int32, [4]> q_with_bias_v_5_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_915_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_915_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587627968))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587774720))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587774464)))];
+            tensor<bool, []> x_59_transpose_x_0 = const()[name = tensor<string, []>("x_59_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_59_transpose_y_0 = const()[name = tensor<string, []>("x_59_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_5 = transpose(perm = q_with_bias_v_5_perm_0, x = var_913)[name = tensor<string, []>("transpose_344")];
+            tensor<fp32, [1, 8, 2, 143]> x_59 = matmul(transpose_x = x_59_transpose_x_0, transpose_y = x_59_transpose_y_0, x = q_with_bias_v_5, y = op_915_quantized)[name = tensor<string, []>("x_59")];
+            tensor<fp32, []> const_105 = const()[name = tensor<string, []>("const_105"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_61_pad_0 = const()[name = tensor<string, []>("x_61_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_61_mode_0 = const()[name = tensor<string, []>("x_61_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_61 = pad(constant_val = const_105, mode = x_61_mode_0, pad = x_61_pad_0, x = x_59)[name = tensor<string, []>("x_61")];
+            tensor<int32, [4]> var_923 = const()[name = tensor<string, []>("op_923"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_63 = reshape(shape = var_923, x = x_61)[name = tensor<string, []>("x_63")];
+            tensor<int32, [4]> var_927_begin_0 = const()[name = tensor<string, []>("op_927_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_927_end_0 = const()[name = tensor<string, []>("op_927_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_927_end_mask_0 = const()[name = tensor<string, []>("op_927_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_927 = slice_by_index(begin = var_927_begin_0, end = var_927_end_0, end_mask = var_927_end_mask_0, x = x_63)[name = tensor<string, []>("op_927")];
+            tensor<int32, [4]> var_928 = const()[name = tensor<string, []>("op_928"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_9 = reshape(shape = var_928, x = var_927)[name = tensor<string, []>("matrix_bd_9")];
+            tensor<bool, []> matrix_ac_5_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_5_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_5_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_100_perm_0 = const()[name = tensor<string, []>("transpose_100_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_101_perm_0 = const()[name = tensor<string, []>("transpose_101_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_101 = transpose(perm = transpose_101_perm_0, x = k_9)[name = tensor<string, []>("transpose_342")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_100 = transpose(perm = transpose_100_perm_0, x = var_911)[name = tensor<string, []>("transpose_343")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_5 = matmul(transpose_x = matrix_ac_5_transpose_x_0, transpose_y = matrix_ac_5_transpose_y_0, x = transpose_100, y = transpose_101)[name = tensor<string, []>("matrix_ac_5")];
+            tensor<int32, [4]> matrix_bd_11_begin_0 = const()[name = tensor<string, []>("matrix_bd_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_11_end_0 = const()[name = tensor<string, []>("matrix_bd_11_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_11_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_11 = slice_by_index(begin = matrix_bd_11_begin_0, end = matrix_bd_11_end_0, end_mask = matrix_bd_11_end_mask_0, x = matrix_bd_9)[name = tensor<string, []>("matrix_bd_11")];
+            tensor<fp32, [1, 8, 2, 72]> var_937 = add(x = matrix_ac_5, y = matrix_bd_11)[name = tensor<string, []>("op_937")];
+            tensor<fp32, []> _inversed_scores_9_y_0 = const()[name = tensor<string, []>("_inversed_scores_9_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_9 = mul(x = var_937, y = _inversed_scores_9_y_0)[name = tensor<string, []>("_inversed_scores_9")];
+            tensor<fp32, [1, 8, 2, 72]> scores_11 = select(a = var_41, b = _inversed_scores_9, cond = mask_11)[name = tensor<string, []>("scores_11")];
+            tensor<fp32, [1, 8, 2, 72]> var_943 = softmax(axis = var_56, x = scores_11)[name = tensor<string, []>("op_943")];
+            tensor<fp32, [1, 8, 2, 72]> input_145 = select(a = var_40, b = var_943, cond = mask_11)[name = tensor<string, []>("input_145")];
+            tensor<bool, []> x_65_transpose_x_0 = const()[name = tensor<string, []>("x_65_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_65_transpose_y_0 = const()[name = tensor<string, []>("x_65_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_13 = transpose(perm = value_13_perm_0, x = v_5)[name = tensor<string, []>("transpose_341")];
+            tensor<fp32, [1, 8, 2, 128]> x_65 = matmul(transpose_x = x_65_transpose_x_0, transpose_y = x_65_transpose_y_0, x = input_145, y = value_13)[name = tensor<string, []>("x_65")];
+            tensor<int32, [4]> var_947_perm_0 = const()[name = tensor<string, []>("op_947_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_948 = const()[name = tensor<string, []>("op_948"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_947 = transpose(perm = var_947_perm_0, x = x_65)[name = tensor<string, []>("transpose_340")];
+            tensor<fp32, [1, 2, 1024]> input_147 = reshape(shape = var_948, x = var_947)[name = tensor<string, []>("input_147")];
+            tensor<fp32, [1, 2, 1024]> input_149 = linear(bias = linear_2_bias_0, weight = module_layers_2_self_attn_linear_out_weight_quantized, x = input_147)[name = tensor<string, []>("linear_25")];
+            tensor<fp32, [1, 2, 1024]> input_151 = add(x = input_141, y = input_149)[name = tensor<string, []>("input_151")];
+            tensor<int32, [1]> x_69_axes_0 = const()[name = tensor<string, []>("x_69_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_69 = layer_norm(axes = x_69_axes_0, beta = module_layers_2_norm_conv_bias, epsilon = var_38, gamma = module_layers_2_norm_conv_weight, x = input_151)[name = tensor<string, []>("x_69")];
+            tensor<int32, [3]> input_153_perm_0 = const()[name = tensor<string, []>("input_153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_155_pad_type_0 = const()[name = tensor<string, []>("input_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_155_strides_0 = const()[name = tensor<string, []>("input_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_155_pad_0 = const()[name = tensor<string, []>("input_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_155_dilations_0 = const()[name = tensor<string, []>("input_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_155_groups_0 = const()[name = tensor<string, []>("input_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_153 = transpose(perm = input_153_perm_0, x = x_69)[name = tensor<string, []>("transpose_339")];
+            tensor<fp32, [1, 2048, 2]> input_155 = conv(dilations = input_155_dilations_0, groups = input_155_groups_0, pad = input_155_pad_0, pad_type = input_155_pad_type_0, strides = input_155_strides_0, weight = module_layers_2_conv_pointwise_conv1_weight_quantized, x = input_153)[name = tensor<string, []>("input_155")];
+            tensor<int32, []> x_71_split_num_splits_0 = const()[name = tensor<string, []>("x_71_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_71_split_axis_0 = const()[name = tensor<string, []>("x_71_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_71_split_0, tensor<fp32, [1, 1024, 2]> x_71_split_1 = split(axis = x_71_split_axis_0, num_splits = x_71_split_num_splits_0, x = input_155)[name = tensor<string, []>("x_71_split")];
+            tensor<fp32, [1, 1024, 2]> x_71_split_1_sigmoid = sigmoid(x = x_71_split_1)[name = tensor<string, []>("x_71_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_71 = mul(x = x_71_split_0, y = x_71_split_1_sigmoid)[name = tensor<string, []>("x_71")];
+            tensor<fp32, [1, 1024, 2]> input_157 = select(a = var_40, b = x_71, cond = var_565)[name = tensor<string, []>("input_157")];
+            tensor<bool, []> new_x_11_interleave_0 = const()[name = tensor<string, []>("new_x_11_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_11 = concat(axis = var_56, interleave = new_x_11_interleave_0, values = (cache_11, input_157))[name = tensor<string, []>("new_x_11")];
+            tensor<int32, [3]> var_986_begin_0 = const()[name = tensor<string, []>("op_986_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_986_end_0 = const()[name = tensor<string, []>("op_986_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_986_end_mask_0 = const()[name = tensor<string, []>("op_986_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_986 = slice_by_index(begin = var_986_begin_0, end = var_986_end_0, end_mask = var_986_end_mask_0, x = new_x_11)[name = tensor<string, []>("op_986")];
+            tensor<string, []> x_73_pad_type_0 = const()[name = tensor<string, []>("x_73_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_73_groups_0 = const()[name = tensor<string, []>("x_73_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_73_strides_0 = const()[name = tensor<string, []>("x_73_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_73_pad_0 = const()[name = tensor<string, []>("x_73_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_73_dilations_0 = const()[name = tensor<string, []>("x_73_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_73 = conv(dilations = x_73_dilations_0, groups = x_73_groups_0, pad = x_73_pad_0, pad_type = x_73_pad_type_0, strides = x_73_strides_0, weight = module_layers_2_conv_depthwise_conv_weight_quantized, x = new_x_11)[name = tensor<string, []>("x_73")];
+            tensor<int32, [3]> input_159_perm_0 = const()[name = tensor<string, []>("input_159_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_75_axes_0 = const()[name = tensor<string, []>("x_75_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_159 = transpose(perm = input_159_perm_0, x = x_73)[name = tensor<string, []>("transpose_338")];
+            tensor<fp32, [1, 2, 1024]> x_75 = layer_norm(axes = x_75_axes_0, beta = module_layers_2_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_2_conv_batch_norm_weight, x = input_159)[name = tensor<string, []>("x_75")];
+            tensor<int32, [3]> input_161_perm_0 = const()[name = tensor<string, []>("input_161_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_161 = transpose(perm = input_161_perm_0, x = x_75)[name = tensor<string, []>("transpose_337")];
+            tensor<fp32, [1, 1024, 2]> input_163 = silu(x = input_161)[name = tensor<string, []>("input_163")];
+            tensor<string, []> x_77_pad_type_0 = const()[name = tensor<string, []>("x_77_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_77_strides_0 = const()[name = tensor<string, []>("x_77_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_77_pad_0 = const()[name = tensor<string, []>("x_77_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_77_dilations_0 = const()[name = tensor<string, []>("x_77_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_77_groups_0 = const()[name = tensor<string, []>("x_77_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_77 = conv(dilations = x_77_dilations_0, groups = x_77_groups_0, pad = x_77_pad_0, pad_type = x_77_pad_type_0, strides = x_77_strides_0, weight = module_layers_2_conv_pointwise_conv2_weight_quantized, x = input_163)[name = tensor<string, []>("x_77")];
+            tensor<int32, [3]> input_165_perm_0 = const()[name = tensor<string, []>("input_165_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_165 = transpose(perm = input_165_perm_0, x = x_77)[name = tensor<string, []>("transpose_336")];
+            tensor<fp32, [1, 2, 1024]> input_167 = add(x = input_151, y = input_165)[name = tensor<string, []>("input_167")];
+            tensor<int32, [1]> input_169_axes_0 = const()[name = tensor<string, []>("input_169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_169 = layer_norm(axes = input_169_axes_0, beta = module_layers_2_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_2_norm_feed_forward2_weight, x = input_167)[name = tensor<string, []>("input_169")];
+            tensor<fp32, [1, 2, 4096]> input_171 = linear(bias = linear_1_bias_0, weight = module_layers_2_feed_forward2_linear1_weight_quantized, x = input_169)[name = tensor<string, []>("linear_26")];
+            tensor<fp32, [1, 2, 4096]> input_173 = silu(x = input_171)[name = tensor<string, []>("input_173")];
+            tensor<fp32, [1, 2, 1024]> input_177 = linear(bias = linear_2_bias_0, weight = module_layers_2_feed_forward2_linear2_weight_quantized, x = input_173)[name = tensor<string, []>("linear_27")];
+            tensor<fp32, []> var_1027 = const()[name = tensor<string, []>("op_1027"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1028 = mul(x = input_177, y = var_1027)[name = tensor<string, []>("op_1028")];
+            tensor<fp32, [1, 2, 1024]> input_179 = add(x = input_167, y = var_1028)[name = tensor<string, []>("input_179")];
+            tensor<int32, [1]> input_181_axes_0 = const()[name = tensor<string, []>("input_181_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_181 = layer_norm(axes = input_181_axes_0, beta = module_layers_2_norm_out_bias, epsilon = var_38, gamma = module_layers_2_norm_out_weight, x = input_179)[name = tensor<string, []>("input_181")];
+            tensor<int32, [4]> cache_13_begin_0 = const()[name = tensor<string, []>("cache_13_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_13_end_0 = const()[name = tensor<string, []>("cache_13_end_0"), val = tensor<int32, [4]>([4, 1, 70, 1024])];
+            tensor<bool, [4]> cache_13_end_mask_0 = const()[name = tensor<string, []>("cache_13_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_13_squeeze_mask_0 = const()[name = tensor<string, []>("cache_13_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_13 = slice_by_index(begin = cache_13_begin_0, end = cache_13_end_0, end_mask = cache_13_end_mask_0, squeeze_mask = cache_13_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_13")];
+            tensor<int32, [4]> cache_15_begin_0 = const()[name = tensor<string, []>("cache_15_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
+            tensor<int32, [4]> cache_15_end_0 = const()[name = tensor<string, []>("cache_15_end_0"), val = tensor<int32, [4]>([4, 1, 1024, 8])];
+            tensor<bool, [4]> cache_15_end_mask_0 = const()[name = tensor<string, []>("cache_15_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_15_squeeze_mask_0 = const()[name = tensor<string, []>("cache_15_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_15 = slice_by_index(begin = cache_15_begin_0, end = cache_15_end_0, end_mask = cache_15_end_mask_0, squeeze_mask = cache_15_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_15")];
+            tensor<int32, [1]> input_183_axes_0 = const()[name = tensor<string, []>("input_183_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_183 = layer_norm(axes = input_183_axes_0, beta = module_layers_3_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_3_norm_feed_forward1_weight, x = input_181)[name = tensor<string, []>("input_183")];
+            tensor<fp32, [1, 2, 4096]> input_185 = linear(bias = linear_1_bias_0, weight = module_layers_3_feed_forward1_linear1_weight_quantized, x = input_183)[name = tensor<string, []>("linear_28")];
+            tensor<fp32, [1, 2, 4096]> input_187 = silu(x = input_185)[name = tensor<string, []>("input_187")];
+            tensor<fp32, [1, 2, 1024]> input_191 = linear(bias = linear_2_bias_0, weight = module_layers_3_feed_forward1_linear2_weight_quantized, x = input_187)[name = tensor<string, []>("linear_29")];
+            tensor<fp32, []> var_1062 = const()[name = tensor<string, []>("op_1062"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1063 = mul(x = input_191, y = var_1062)[name = tensor<string, []>("op_1063")];
+            tensor<fp32, [1, 2, 1024]> input_193 = add(x = input_181, y = var_1063)[name = tensor<string, []>("input_193")];
+            tensor<int32, [1]> key_7_axes_0 = const()[name = tensor<string, []>("key_7_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_7 = layer_norm(axes = key_7_axes_0, beta = module_layers_3_norm_self_att_bias, epsilon = var_38, gamma = module_layers_3_norm_self_att_weight, x = input_193)[name = tensor<string, []>("key_7")];
+            tensor<bool, []> input_195_interleave_0 = const()[name = tensor<string, []>("input_195_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_195 = concat(axis = var_65, interleave = input_195_interleave_0, values = (cache_13, key_7))[name = tensor<string, []>("input_195")];
+            tensor<int32, [3]> var_1085_begin_0 = const()[name = tensor<string, []>("op_1085_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_1085_end_0 = const()[name = tensor<string, []>("op_1085_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_1085_end_mask_0 = const()[name = tensor<string, []>("op_1085_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_1085 = slice_by_index(begin = var_1085_begin_0, end = var_1085_end_0, end_mask = var_1085_end_mask_0, x = cache_13)[name = tensor<string, []>("op_1085")];
+            tensor<bool, []> var_1091_interleave_0 = const()[name = tensor<string, []>("op_1091_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_1091 = concat(axis = var_65, interleave = var_1091_interleave_0, values = (var_1085, key_7))[name = tensor<string, []>("op_1091")];
+            tensor<fp32, [1, 2, 1024]> var_1094 = linear(bias = linear_2_bias_0, weight = module_layers_3_self_attn_linear_q_weight_quantized, x = key_7)[name = tensor<string, []>("linear_30")];
+            tensor<int32, [4]> var_1095 = const()[name = tensor<string, []>("op_1095"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_19 = reshape(shape = var_1095, x = var_1094)[name = tensor<string, []>("q_19")];
+            tensor<fp32, [1, 72, 1024]> var_1098 = linear(bias = linear_2_bias_0, weight = module_layers_3_self_attn_linear_k_weight_quantized, x = input_195)[name = tensor<string, []>("linear_31")];
+            tensor<int32, [4]> var_1099 = const()[name = tensor<string, []>("op_1099"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_13 = reshape(shape = var_1099, x = var_1098)[name = tensor<string, []>("k_13")];
+            tensor<fp32, [1, 72, 1024]> var_1102 = linear(bias = linear_2_bias_0, weight = module_layers_3_self_attn_linear_v_weight_quantized, x = input_195)[name = tensor<string, []>("linear_32")];
+            tensor<int32, [4]> var_1103 = const()[name = tensor<string, []>("op_1103"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_7 = reshape(shape = var_1103, x = var_1102)[name = tensor<string, []>("v_7")];
+            tensor<int32, [4]> value_15_perm_0 = const()[name = tensor<string, []>("value_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_1115 = add(x = q_19, y = module_layers_3_self_attn_pos_bias_u)[name = tensor<string, []>("op_1115")];
+            tensor<fp32, [1, 2, 8, 128]> var_1117 = add(x = q_19, y = module_layers_3_self_attn_pos_bias_v)[name = tensor<string, []>("op_1117")];
+            tensor<int32, [4]> q_with_bias_v_7_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_1119_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1119_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587775360))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587922112))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587921856)))];
+            tensor<bool, []> x_85_transpose_x_0 = const()[name = tensor<string, []>("x_85_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_85_transpose_y_0 = const()[name = tensor<string, []>("x_85_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_7 = transpose(perm = q_with_bias_v_7_perm_0, x = var_1117)[name = tensor<string, []>("transpose_335")];
+            tensor<fp32, [1, 8, 2, 143]> x_85 = matmul(transpose_x = x_85_transpose_x_0, transpose_y = x_85_transpose_y_0, x = q_with_bias_v_7, y = op_1119_quantized)[name = tensor<string, []>("x_85")];
+            tensor<fp32, []> const_118 = const()[name = tensor<string, []>("const_118"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_87_pad_0 = const()[name = tensor<string, []>("x_87_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_87_mode_0 = const()[name = tensor<string, []>("x_87_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_87 = pad(constant_val = const_118, mode = x_87_mode_0, pad = x_87_pad_0, x = x_85)[name = tensor<string, []>("x_87")];
+            tensor<int32, [4]> var_1127 = const()[name = tensor<string, []>("op_1127"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_89 = reshape(shape = var_1127, x = x_87)[name = tensor<string, []>("x_89")];
+            tensor<int32, [4]> var_1131_begin_0 = const()[name = tensor<string, []>("op_1131_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1131_end_0 = const()[name = tensor<string, []>("op_1131_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_1131_end_mask_0 = const()[name = tensor<string, []>("op_1131_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_1131 = slice_by_index(begin = var_1131_begin_0, end = var_1131_end_0, end_mask = var_1131_end_mask_0, x = x_89)[name = tensor<string, []>("op_1131")];
+            tensor<int32, [4]> var_1132 = const()[name = tensor<string, []>("op_1132"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_13 = reshape(shape = var_1132, x = var_1131)[name = tensor<string, []>("matrix_bd_13")];
+            tensor<bool, []> matrix_ac_7_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_7_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_7_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_102_perm_0 = const()[name = tensor<string, []>("transpose_102_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_103_perm_0 = const()[name = tensor<string, []>("transpose_103_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_103 = transpose(perm = transpose_103_perm_0, x = k_13)[name = tensor<string, []>("transpose_333")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_102 = transpose(perm = transpose_102_perm_0, x = var_1115)[name = tensor<string, []>("transpose_334")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_7 = matmul(transpose_x = matrix_ac_7_transpose_x_0, transpose_y = matrix_ac_7_transpose_y_0, x = transpose_102, y = transpose_103)[name = tensor<string, []>("matrix_ac_7")];
+            tensor<int32, [4]> matrix_bd_15_begin_0 = const()[name = tensor<string, []>("matrix_bd_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_15_end_0 = const()[name = tensor<string, []>("matrix_bd_15_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_15_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_15 = slice_by_index(begin = matrix_bd_15_begin_0, end = matrix_bd_15_end_0, end_mask = matrix_bd_15_end_mask_0, x = matrix_bd_13)[name = tensor<string, []>("matrix_bd_15")];
+            tensor<fp32, [1, 8, 2, 72]> var_1141 = add(x = matrix_ac_7, y = matrix_bd_15)[name = tensor<string, []>("op_1141")];
+            tensor<fp32, []> _inversed_scores_13_y_0 = const()[name = tensor<string, []>("_inversed_scores_13_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_13 = mul(x = var_1141, y = _inversed_scores_13_y_0)[name = tensor<string, []>("_inversed_scores_13")];
+            tensor<fp32, [1, 8, 2, 72]> scores_15 = select(a = var_41, b = _inversed_scores_13, cond = mask_11)[name = tensor<string, []>("scores_15")];
+            tensor<fp32, [1, 8, 2, 72]> var_1147 = softmax(axis = var_56, x = scores_15)[name = tensor<string, []>("op_1147")];
+            tensor<fp32, [1, 8, 2, 72]> input_197 = select(a = var_40, b = var_1147, cond = mask_11)[name = tensor<string, []>("input_197")];
+            tensor<bool, []> x_91_transpose_x_0 = const()[name = tensor<string, []>("x_91_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_91_transpose_y_0 = const()[name = tensor<string, []>("x_91_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_15 = transpose(perm = value_15_perm_0, x = v_7)[name = tensor<string, []>("transpose_332")];
+            tensor<fp32, [1, 8, 2, 128]> x_91 = matmul(transpose_x = x_91_transpose_x_0, transpose_y = x_91_transpose_y_0, x = input_197, y = value_15)[name = tensor<string, []>("x_91")];
+            tensor<int32, [4]> var_1151_perm_0 = const()[name = tensor<string, []>("op_1151_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1152 = const()[name = tensor<string, []>("op_1152"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_1151 = transpose(perm = var_1151_perm_0, x = x_91)[name = tensor<string, []>("transpose_331")];
+            tensor<fp32, [1, 2, 1024]> input_199 = reshape(shape = var_1152, x = var_1151)[name = tensor<string, []>("input_199")];
+            tensor<fp32, [1, 2, 1024]> input_201 = linear(bias = linear_2_bias_0, weight = module_layers_3_self_attn_linear_out_weight_quantized, x = input_199)[name = tensor<string, []>("linear_34")];
+            tensor<fp32, [1, 2, 1024]> input_203 = add(x = input_193, y = input_201)[name = tensor<string, []>("input_203")];
+            tensor<int32, [1]> x_95_axes_0 = const()[name = tensor<string, []>("x_95_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_95 = layer_norm(axes = x_95_axes_0, beta = module_layers_3_norm_conv_bias, epsilon = var_38, gamma = module_layers_3_norm_conv_weight, x = input_203)[name = tensor<string, []>("x_95")];
+            tensor<int32, [3]> input_205_perm_0 = const()[name = tensor<string, []>("input_205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_207_pad_type_0 = const()[name = tensor<string, []>("input_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_207_strides_0 = const()[name = tensor<string, []>("input_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_207_pad_0 = const()[name = tensor<string, []>("input_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_207_dilations_0 = const()[name = tensor<string, []>("input_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_207_groups_0 = const()[name = tensor<string, []>("input_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_205 = transpose(perm = input_205_perm_0, x = x_95)[name = tensor<string, []>("transpose_330")];
+            tensor<fp32, [1, 2048, 2]> input_207 = conv(dilations = input_207_dilations_0, groups = input_207_groups_0, pad = input_207_pad_0, pad_type = input_207_pad_type_0, strides = input_207_strides_0, weight = module_layers_3_conv_pointwise_conv1_weight_quantized, x = input_205)[name = tensor<string, []>("input_207")];
+            tensor<int32, []> x_97_split_num_splits_0 = const()[name = tensor<string, []>("x_97_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_97_split_axis_0 = const()[name = tensor<string, []>("x_97_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_97_split_0, tensor<fp32, [1, 1024, 2]> x_97_split_1 = split(axis = x_97_split_axis_0, num_splits = x_97_split_num_splits_0, x = input_207)[name = tensor<string, []>("x_97_split")];
+            tensor<fp32, [1, 1024, 2]> x_97_split_1_sigmoid = sigmoid(x = x_97_split_1)[name = tensor<string, []>("x_97_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_97 = mul(x = x_97_split_0, y = x_97_split_1_sigmoid)[name = tensor<string, []>("x_97")];
+            tensor<fp32, [1, 1024, 2]> input_209 = select(a = var_40, b = x_97, cond = var_565)[name = tensor<string, []>("input_209")];
+            tensor<bool, []> new_x_15_interleave_0 = const()[name = tensor<string, []>("new_x_15_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_15 = concat(axis = var_56, interleave = new_x_15_interleave_0, values = (cache_15, input_209))[name = tensor<string, []>("new_x_15")];
+            tensor<int32, [3]> var_1190_begin_0 = const()[name = tensor<string, []>("op_1190_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_1190_end_0 = const()[name = tensor<string, []>("op_1190_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_1190_end_mask_0 = const()[name = tensor<string, []>("op_1190_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_1190 = slice_by_index(begin = var_1190_begin_0, end = var_1190_end_0, end_mask = var_1190_end_mask_0, x = new_x_15)[name = tensor<string, []>("op_1190")];
+            tensor<string, []> x_99_pad_type_0 = const()[name = tensor<string, []>("x_99_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_99_groups_0 = const()[name = tensor<string, []>("x_99_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_99_strides_0 = const()[name = tensor<string, []>("x_99_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_99_pad_0 = const()[name = tensor<string, []>("x_99_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_99_dilations_0 = const()[name = tensor<string, []>("x_99_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_99 = conv(dilations = x_99_dilations_0, groups = x_99_groups_0, pad = x_99_pad_0, pad_type = x_99_pad_type_0, strides = x_99_strides_0, weight = module_layers_3_conv_depthwise_conv_weight_quantized, x = new_x_15)[name = tensor<string, []>("x_99")];
+            tensor<int32, [3]> input_211_perm_0 = const()[name = tensor<string, []>("input_211_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_101_axes_0 = const()[name = tensor<string, []>("x_101_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_211 = transpose(perm = input_211_perm_0, x = x_99)[name = tensor<string, []>("transpose_329")];
+            tensor<fp32, [1, 2, 1024]> x_101 = layer_norm(axes = x_101_axes_0, beta = module_layers_3_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_3_conv_batch_norm_weight, x = input_211)[name = tensor<string, []>("x_101")];
+            tensor<int32, [3]> input_213_perm_0 = const()[name = tensor<string, []>("input_213_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_213 = transpose(perm = input_213_perm_0, x = x_101)[name = tensor<string, []>("transpose_328")];
+            tensor<fp32, [1, 1024, 2]> input_215 = silu(x = input_213)[name = tensor<string, []>("input_215")];
+            tensor<string, []> x_103_pad_type_0 = const()[name = tensor<string, []>("x_103_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_103_strides_0 = const()[name = tensor<string, []>("x_103_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_103_pad_0 = const()[name = tensor<string, []>("x_103_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_103_dilations_0 = const()[name = tensor<string, []>("x_103_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_103_groups_0 = const()[name = tensor<string, []>("x_103_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_103 = conv(dilations = x_103_dilations_0, groups = x_103_groups_0, pad = x_103_pad_0, pad_type = x_103_pad_type_0, strides = x_103_strides_0, weight = module_layers_3_conv_pointwise_conv2_weight_quantized, x = input_215)[name = tensor<string, []>("x_103")];
+            tensor<int32, [3]> input_217_perm_0 = const()[name = tensor<string, []>("input_217_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_217 = transpose(perm = input_217_perm_0, x = x_103)[name = tensor<string, []>("transpose_327")];
+            tensor<fp32, [1, 2, 1024]> input_219 = add(x = input_203, y = input_217)[name = tensor<string, []>("input_219")];
+            tensor<int32, [1]> input_221_axes_0 = const()[name = tensor<string, []>("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_221 = layer_norm(axes = input_221_axes_0, beta = module_layers_3_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_3_norm_feed_forward2_weight, x = input_219)[name = tensor<string, []>("input_221")];
+            tensor<fp32, [1, 2, 4096]> input_223 = linear(bias = linear_1_bias_0, weight = module_layers_3_feed_forward2_linear1_weight_quantized, x = input_221)[name = tensor<string, []>("linear_35")];
+            tensor<fp32, [1, 2, 4096]> input_225 = silu(x = input_223)[name = tensor<string, []>("input_225")];
+            tensor<fp32, [1, 2, 1024]> input_229 = linear(bias = linear_2_bias_0, weight = module_layers_3_feed_forward2_linear2_weight_quantized, x = input_225)[name = tensor<string, []>("linear_36")];
+            tensor<fp32, []> var_1231 = const()[name = tensor<string, []>("op_1231"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1232 = mul(x = input_229, y = var_1231)[name = tensor<string, []>("op_1232")];
+            tensor<fp32, [1, 2, 1024]> input_231 = add(x = input_219, y = var_1232)[name = tensor<string, []>("input_231")];
+            tensor<int32, [1]> input_233_axes_0 = const()[name = tensor<string, []>("input_233_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_233 = layer_norm(axes = input_233_axes_0, beta = module_layers_3_norm_out_bias, epsilon = var_38, gamma = module_layers_3_norm_out_weight, x = input_231)[name = tensor<string, []>("input_233")];
+            tensor<int32, [4]> cache_17_begin_0 = const()[name = tensor<string, []>("cache_17_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_17_end_0 = const()[name = tensor<string, []>("cache_17_end_0"), val = tensor<int32, [4]>([5, 1, 70, 1024])];
+            tensor<bool, [4]> cache_17_end_mask_0 = const()[name = tensor<string, []>("cache_17_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_17_squeeze_mask_0 = const()[name = tensor<string, []>("cache_17_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_17 = slice_by_index(begin = cache_17_begin_0, end = cache_17_end_0, end_mask = cache_17_end_mask_0, squeeze_mask = cache_17_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_17")];
+            tensor<int32, [4]> cache_19_begin_0 = const()[name = tensor<string, []>("cache_19_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
+            tensor<int32, [4]> cache_19_end_0 = const()[name = tensor<string, []>("cache_19_end_0"), val = tensor<int32, [4]>([5, 1, 1024, 8])];
+            tensor<bool, [4]> cache_19_end_mask_0 = const()[name = tensor<string, []>("cache_19_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_19_squeeze_mask_0 = const()[name = tensor<string, []>("cache_19_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_19 = slice_by_index(begin = cache_19_begin_0, end = cache_19_end_0, end_mask = cache_19_end_mask_0, squeeze_mask = cache_19_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_19")];
+            tensor<int32, [1]> input_235_axes_0 = const()[name = tensor<string, []>("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_235 = layer_norm(axes = input_235_axes_0, beta = module_layers_4_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_4_norm_feed_forward1_weight, x = input_233)[name = tensor<string, []>("input_235")];
+            tensor<fp32, [1, 2, 4096]> input_237 = linear(bias = linear_1_bias_0, weight = module_layers_4_feed_forward1_linear1_weight_quantized, x = input_235)[name = tensor<string, []>("linear_37")];
+            tensor<fp32, [1, 2, 4096]> input_239 = silu(x = input_237)[name = tensor<string, []>("input_239")];
+            tensor<fp32, [1, 2, 1024]> input_243 = linear(bias = linear_2_bias_0, weight = module_layers_4_feed_forward1_linear2_weight_quantized, x = input_239)[name = tensor<string, []>("linear_38")];
+            tensor<fp32, []> var_1266 = const()[name = tensor<string, []>("op_1266"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1267 = mul(x = input_243, y = var_1266)[name = tensor<string, []>("op_1267")];
+            tensor<fp32, [1, 2, 1024]> input_245 = add(x = input_233, y = var_1267)[name = tensor<string, []>("input_245")];
+            tensor<int32, [1]> key_9_axes_0 = const()[name = tensor<string, []>("key_9_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_9 = layer_norm(axes = key_9_axes_0, beta = module_layers_4_norm_self_att_bias, epsilon = var_38, gamma = module_layers_4_norm_self_att_weight, x = input_245)[name = tensor<string, []>("key_9")];
+            tensor<bool, []> input_247_interleave_0 = const()[name = tensor<string, []>("input_247_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_247 = concat(axis = var_65, interleave = input_247_interleave_0, values = (cache_17, key_9))[name = tensor<string, []>("input_247")];
+            tensor<int32, [3]> var_1289_begin_0 = const()[name = tensor<string, []>("op_1289_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_1289_end_0 = const()[name = tensor<string, []>("op_1289_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_1289_end_mask_0 = const()[name = tensor<string, []>("op_1289_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_1289 = slice_by_index(begin = var_1289_begin_0, end = var_1289_end_0, end_mask = var_1289_end_mask_0, x = cache_17)[name = tensor<string, []>("op_1289")];
+            tensor<bool, []> var_1295_interleave_0 = const()[name = tensor<string, []>("op_1295_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_1295 = concat(axis = var_65, interleave = var_1295_interleave_0, values = (var_1289, key_9))[name = tensor<string, []>("op_1295")];
+            tensor<fp32, [1, 2, 1024]> var_1298 = linear(bias = linear_2_bias_0, weight = module_layers_4_self_attn_linear_q_weight_quantized, x = key_9)[name = tensor<string, []>("linear_39")];
+            tensor<int32, [4]> var_1299 = const()[name = tensor<string, []>("op_1299"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_25 = reshape(shape = var_1299, x = var_1298)[name = tensor<string, []>("q_25")];
+            tensor<fp32, [1, 72, 1024]> var_1302 = linear(bias = linear_2_bias_0, weight = module_layers_4_self_attn_linear_k_weight_quantized, x = input_247)[name = tensor<string, []>("linear_40")];
+            tensor<int32, [4]> var_1303 = const()[name = tensor<string, []>("op_1303"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_17 = reshape(shape = var_1303, x = var_1302)[name = tensor<string, []>("k_17")];
+            tensor<fp32, [1, 72, 1024]> var_1306 = linear(bias = linear_2_bias_0, weight = module_layers_4_self_attn_linear_v_weight_quantized, x = input_247)[name = tensor<string, []>("linear_41")];
+            tensor<int32, [4]> var_1307 = const()[name = tensor<string, []>("op_1307"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_9 = reshape(shape = var_1307, x = var_1306)[name = tensor<string, []>("v_9")];
+            tensor<int32, [4]> value_17_perm_0 = const()[name = tensor<string, []>("value_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_1319 = add(x = q_25, y = module_layers_4_self_attn_pos_bias_u)[name = tensor<string, []>("op_1319")];
+            tensor<fp32, [1, 2, 8, 128]> var_1321 = add(x = q_25, y = module_layers_4_self_attn_pos_bias_v)[name = tensor<string, []>("op_1321")];
+            tensor<int32, [4]> q_with_bias_v_9_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_1323_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1323_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(587922752))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588069504))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588069248)))];
+            tensor<bool, []> x_111_transpose_x_0 = const()[name = tensor<string, []>("x_111_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_111_transpose_y_0 = const()[name = tensor<string, []>("x_111_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_9 = transpose(perm = q_with_bias_v_9_perm_0, x = var_1321)[name = tensor<string, []>("transpose_326")];
+            tensor<fp32, [1, 8, 2, 143]> x_111 = matmul(transpose_x = x_111_transpose_x_0, transpose_y = x_111_transpose_y_0, x = q_with_bias_v_9, y = op_1323_quantized)[name = tensor<string, []>("x_111")];
+            tensor<fp32, []> const_131 = const()[name = tensor<string, []>("const_131"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_113_pad_0 = const()[name = tensor<string, []>("x_113_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_113_mode_0 = const()[name = tensor<string, []>("x_113_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_113 = pad(constant_val = const_131, mode = x_113_mode_0, pad = x_113_pad_0, x = x_111)[name = tensor<string, []>("x_113")];
+            tensor<int32, [4]> var_1331 = const()[name = tensor<string, []>("op_1331"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_115 = reshape(shape = var_1331, x = x_113)[name = tensor<string, []>("x_115")];
+            tensor<int32, [4]> var_1335_begin_0 = const()[name = tensor<string, []>("op_1335_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1335_end_0 = const()[name = tensor<string, []>("op_1335_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_1335_end_mask_0 = const()[name = tensor<string, []>("op_1335_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_1335 = slice_by_index(begin = var_1335_begin_0, end = var_1335_end_0, end_mask = var_1335_end_mask_0, x = x_115)[name = tensor<string, []>("op_1335")];
+            tensor<int32, [4]> var_1336 = const()[name = tensor<string, []>("op_1336"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_17 = reshape(shape = var_1336, x = var_1335)[name = tensor<string, []>("matrix_bd_17")];
+            tensor<bool, []> matrix_ac_9_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_9_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_9_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_104_perm_0 = const()[name = tensor<string, []>("transpose_104_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_105_perm_0 = const()[name = tensor<string, []>("transpose_105_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_105 = transpose(perm = transpose_105_perm_0, x = k_17)[name = tensor<string, []>("transpose_324")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_104 = transpose(perm = transpose_104_perm_0, x = var_1319)[name = tensor<string, []>("transpose_325")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_9 = matmul(transpose_x = matrix_ac_9_transpose_x_0, transpose_y = matrix_ac_9_transpose_y_0, x = transpose_104, y = transpose_105)[name = tensor<string, []>("matrix_ac_9")];
+            tensor<int32, [4]> matrix_bd_19_begin_0 = const()[name = tensor<string, []>("matrix_bd_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_19_end_0 = const()[name = tensor<string, []>("matrix_bd_19_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_19_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_19 = slice_by_index(begin = matrix_bd_19_begin_0, end = matrix_bd_19_end_0, end_mask = matrix_bd_19_end_mask_0, x = matrix_bd_17)[name = tensor<string, []>("matrix_bd_19")];
+            tensor<fp32, [1, 8, 2, 72]> var_1345 = add(x = matrix_ac_9, y = matrix_bd_19)[name = tensor<string, []>("op_1345")];
+            tensor<fp32, []> _inversed_scores_17_y_0 = const()[name = tensor<string, []>("_inversed_scores_17_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_17 = mul(x = var_1345, y = _inversed_scores_17_y_0)[name = tensor<string, []>("_inversed_scores_17")];
+            tensor<fp32, [1, 8, 2, 72]> scores_19 = select(a = var_41, b = _inversed_scores_17, cond = mask_11)[name = tensor<string, []>("scores_19")];
+            tensor<fp32, [1, 8, 2, 72]> var_1351 = softmax(axis = var_56, x = scores_19)[name = tensor<string, []>("op_1351")];
+            tensor<fp32, [1, 8, 2, 72]> input_249 = select(a = var_40, b = var_1351, cond = mask_11)[name = tensor<string, []>("input_249")];
+            tensor<bool, []> x_117_transpose_x_0 = const()[name = tensor<string, []>("x_117_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_117_transpose_y_0 = const()[name = tensor<string, []>("x_117_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_17 = transpose(perm = value_17_perm_0, x = v_9)[name = tensor<string, []>("transpose_323")];
+            tensor<fp32, [1, 8, 2, 128]> x_117 = matmul(transpose_x = x_117_transpose_x_0, transpose_y = x_117_transpose_y_0, x = input_249, y = value_17)[name = tensor<string, []>("x_117")];
+            tensor<int32, [4]> var_1355_perm_0 = const()[name = tensor<string, []>("op_1355_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1356 = const()[name = tensor<string, []>("op_1356"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_1355 = transpose(perm = var_1355_perm_0, x = x_117)[name = tensor<string, []>("transpose_322")];
+            tensor<fp32, [1, 2, 1024]> input_251 = reshape(shape = var_1356, x = var_1355)[name = tensor<string, []>("input_251")];
+            tensor<fp32, [1, 2, 1024]> input_253 = linear(bias = linear_2_bias_0, weight = module_layers_4_self_attn_linear_out_weight_quantized, x = input_251)[name = tensor<string, []>("linear_43")];
+            tensor<fp32, [1, 2, 1024]> input_255 = add(x = input_245, y = input_253)[name = tensor<string, []>("input_255")];
+            tensor<int32, [1]> x_121_axes_0 = const()[name = tensor<string, []>("x_121_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_121 = layer_norm(axes = x_121_axes_0, beta = module_layers_4_norm_conv_bias, epsilon = var_38, gamma = module_layers_4_norm_conv_weight, x = input_255)[name = tensor<string, []>("x_121")];
+            tensor<int32, [3]> input_257_perm_0 = const()[name = tensor<string, []>("input_257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_259_pad_type_0 = const()[name = tensor<string, []>("input_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_259_strides_0 = const()[name = tensor<string, []>("input_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_259_pad_0 = const()[name = tensor<string, []>("input_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_259_dilations_0 = const()[name = tensor<string, []>("input_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_259_groups_0 = const()[name = tensor<string, []>("input_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_257 = transpose(perm = input_257_perm_0, x = x_121)[name = tensor<string, []>("transpose_321")];
+            tensor<fp32, [1, 2048, 2]> input_259 = conv(dilations = input_259_dilations_0, groups = input_259_groups_0, pad = input_259_pad_0, pad_type = input_259_pad_type_0, strides = input_259_strides_0, weight = module_layers_4_conv_pointwise_conv1_weight_quantized, x = input_257)[name = tensor<string, []>("input_259")];
+            tensor<int32, []> x_123_split_num_splits_0 = const()[name = tensor<string, []>("x_123_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_123_split_axis_0 = const()[name = tensor<string, []>("x_123_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_123_split_0, tensor<fp32, [1, 1024, 2]> x_123_split_1 = split(axis = x_123_split_axis_0, num_splits = x_123_split_num_splits_0, x = input_259)[name = tensor<string, []>("x_123_split")];
+            tensor<fp32, [1, 1024, 2]> x_123_split_1_sigmoid = sigmoid(x = x_123_split_1)[name = tensor<string, []>("x_123_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_123 = mul(x = x_123_split_0, y = x_123_split_1_sigmoid)[name = tensor<string, []>("x_123")];
+            tensor<fp32, [1, 1024, 2]> input_261 = select(a = var_40, b = x_123, cond = var_565)[name = tensor<string, []>("input_261")];
+            tensor<bool, []> new_x_19_interleave_0 = const()[name = tensor<string, []>("new_x_19_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_19 = concat(axis = var_56, interleave = new_x_19_interleave_0, values = (cache_19, input_261))[name = tensor<string, []>("new_x_19")];
+            tensor<int32, [3]> var_1394_begin_0 = const()[name = tensor<string, []>("op_1394_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_1394_end_0 = const()[name = tensor<string, []>("op_1394_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_1394_end_mask_0 = const()[name = tensor<string, []>("op_1394_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_1394 = slice_by_index(begin = var_1394_begin_0, end = var_1394_end_0, end_mask = var_1394_end_mask_0, x = new_x_19)[name = tensor<string, []>("op_1394")];
+            tensor<string, []> x_125_pad_type_0 = const()[name = tensor<string, []>("x_125_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_125_groups_0 = const()[name = tensor<string, []>("x_125_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_125_strides_0 = const()[name = tensor<string, []>("x_125_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_125_pad_0 = const()[name = tensor<string, []>("x_125_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_125_dilations_0 = const()[name = tensor<string, []>("x_125_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_125 = conv(dilations = x_125_dilations_0, groups = x_125_groups_0, pad = x_125_pad_0, pad_type = x_125_pad_type_0, strides = x_125_strides_0, weight = module_layers_4_conv_depthwise_conv_weight_quantized, x = new_x_19)[name = tensor<string, []>("x_125")];
+            tensor<int32, [3]> input_263_perm_0 = const()[name = tensor<string, []>("input_263_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_127_axes_0 = const()[name = tensor<string, []>("x_127_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_263 = transpose(perm = input_263_perm_0, x = x_125)[name = tensor<string, []>("transpose_320")];
+            tensor<fp32, [1, 2, 1024]> x_127 = layer_norm(axes = x_127_axes_0, beta = module_layers_4_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_4_conv_batch_norm_weight, x = input_263)[name = tensor<string, []>("x_127")];
+            tensor<int32, [3]> input_265_perm_0 = const()[name = tensor<string, []>("input_265_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_265 = transpose(perm = input_265_perm_0, x = x_127)[name = tensor<string, []>("transpose_319")];
+            tensor<fp32, [1, 1024, 2]> input_267 = silu(x = input_265)[name = tensor<string, []>("input_267")];
+            tensor<string, []> x_129_pad_type_0 = const()[name = tensor<string, []>("x_129_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_129_strides_0 = const()[name = tensor<string, []>("x_129_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_129_pad_0 = const()[name = tensor<string, []>("x_129_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_129_dilations_0 = const()[name = tensor<string, []>("x_129_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_129_groups_0 = const()[name = tensor<string, []>("x_129_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_129 = conv(dilations = x_129_dilations_0, groups = x_129_groups_0, pad = x_129_pad_0, pad_type = x_129_pad_type_0, strides = x_129_strides_0, weight = module_layers_4_conv_pointwise_conv2_weight_quantized, x = input_267)[name = tensor<string, []>("x_129")];
+            tensor<int32, [3]> input_269_perm_0 = const()[name = tensor<string, []>("input_269_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_269 = transpose(perm = input_269_perm_0, x = x_129)[name = tensor<string, []>("transpose_318")];
+            tensor<fp32, [1, 2, 1024]> input_271 = add(x = input_255, y = input_269)[name = tensor<string, []>("input_271")];
+            tensor<int32, [1]> input_273_axes_0 = const()[name = tensor<string, []>("input_273_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_273 = layer_norm(axes = input_273_axes_0, beta = module_layers_4_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_4_norm_feed_forward2_weight, x = input_271)[name = tensor<string, []>("input_273")];
+            tensor<fp32, [1, 2, 4096]> input_275 = linear(bias = linear_1_bias_0, weight = module_layers_4_feed_forward2_linear1_weight_quantized, x = input_273)[name = tensor<string, []>("linear_44")];
+            tensor<fp32, [1, 2, 4096]> input_277 = silu(x = input_275)[name = tensor<string, []>("input_277")];
+            tensor<fp32, [1, 2, 1024]> input_281 = linear(bias = linear_2_bias_0, weight = module_layers_4_feed_forward2_linear2_weight_quantized, x = input_277)[name = tensor<string, []>("linear_45")];
+            tensor<fp32, []> var_1435 = const()[name = tensor<string, []>("op_1435"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1436 = mul(x = input_281, y = var_1435)[name = tensor<string, []>("op_1436")];
+            tensor<fp32, [1, 2, 1024]> input_283 = add(x = input_271, y = var_1436)[name = tensor<string, []>("input_283")];
+            tensor<int32, [1]> input_285_axes_0 = const()[name = tensor<string, []>("input_285_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_285 = layer_norm(axes = input_285_axes_0, beta = module_layers_4_norm_out_bias, epsilon = var_38, gamma = module_layers_4_norm_out_weight, x = input_283)[name = tensor<string, []>("input_285")];
+            tensor<int32, [4]> cache_21_begin_0 = const()[name = tensor<string, []>("cache_21_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_21_end_0 = const()[name = tensor<string, []>("cache_21_end_0"), val = tensor<int32, [4]>([6, 1, 70, 1024])];
+            tensor<bool, [4]> cache_21_end_mask_0 = const()[name = tensor<string, []>("cache_21_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_21_squeeze_mask_0 = const()[name = tensor<string, []>("cache_21_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_21 = slice_by_index(begin = cache_21_begin_0, end = cache_21_end_0, end_mask = cache_21_end_mask_0, squeeze_mask = cache_21_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_21")];
+            tensor<int32, [4]> cache_23_begin_0 = const()[name = tensor<string, []>("cache_23_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
+            tensor<int32, [4]> cache_23_end_0 = const()[name = tensor<string, []>("cache_23_end_0"), val = tensor<int32, [4]>([6, 1, 1024, 8])];
+            tensor<bool, [4]> cache_23_end_mask_0 = const()[name = tensor<string, []>("cache_23_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_23_squeeze_mask_0 = const()[name = tensor<string, []>("cache_23_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_23 = slice_by_index(begin = cache_23_begin_0, end = cache_23_end_0, end_mask = cache_23_end_mask_0, squeeze_mask = cache_23_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_23")];
+            tensor<int32, [1]> input_287_axes_0 = const()[name = tensor<string, []>("input_287_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_287 = layer_norm(axes = input_287_axes_0, beta = module_layers_5_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_5_norm_feed_forward1_weight, x = input_285)[name = tensor<string, []>("input_287")];
+            tensor<fp32, [1, 2, 4096]> input_289 = linear(bias = linear_1_bias_0, weight = module_layers_5_feed_forward1_linear1_weight_quantized, x = input_287)[name = tensor<string, []>("linear_46")];
+            tensor<fp32, [1, 2, 4096]> input_291 = silu(x = input_289)[name = tensor<string, []>("input_291")];
+            tensor<fp32, [1, 2, 1024]> input_295 = linear(bias = linear_2_bias_0, weight = module_layers_5_feed_forward1_linear2_weight_quantized, x = input_291)[name = tensor<string, []>("linear_47")];
+            tensor<fp32, []> var_1470 = const()[name = tensor<string, []>("op_1470"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1471 = mul(x = input_295, y = var_1470)[name = tensor<string, []>("op_1471")];
+            tensor<fp32, [1, 2, 1024]> input_297 = add(x = input_285, y = var_1471)[name = tensor<string, []>("input_297")];
+            tensor<int32, [1]> key_11_axes_0 = const()[name = tensor<string, []>("key_11_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_11 = layer_norm(axes = key_11_axes_0, beta = module_layers_5_norm_self_att_bias, epsilon = var_38, gamma = module_layers_5_norm_self_att_weight, x = input_297)[name = tensor<string, []>("key_11")];
+            tensor<bool, []> input_299_interleave_0 = const()[name = tensor<string, []>("input_299_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_299 = concat(axis = var_65, interleave = input_299_interleave_0, values = (cache_21, key_11))[name = tensor<string, []>("input_299")];
+            tensor<int32, [3]> var_1493_begin_0 = const()[name = tensor<string, []>("op_1493_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_1493_end_0 = const()[name = tensor<string, []>("op_1493_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_1493_end_mask_0 = const()[name = tensor<string, []>("op_1493_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_1493 = slice_by_index(begin = var_1493_begin_0, end = var_1493_end_0, end_mask = var_1493_end_mask_0, x = cache_21)[name = tensor<string, []>("op_1493")];
+            tensor<bool, []> var_1499_interleave_0 = const()[name = tensor<string, []>("op_1499_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_1499 = concat(axis = var_65, interleave = var_1499_interleave_0, values = (var_1493, key_11))[name = tensor<string, []>("op_1499")];
+            tensor<fp32, [1, 2, 1024]> var_1502 = linear(bias = linear_2_bias_0, weight = module_layers_5_self_attn_linear_q_weight_quantized, x = key_11)[name = tensor<string, []>("linear_48")];
+            tensor<int32, [4]> var_1503 = const()[name = tensor<string, []>("op_1503"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_31 = reshape(shape = var_1503, x = var_1502)[name = tensor<string, []>("q_31")];
+            tensor<fp32, [1, 72, 1024]> var_1506 = linear(bias = linear_2_bias_0, weight = module_layers_5_self_attn_linear_k_weight_quantized, x = input_299)[name = tensor<string, []>("linear_49")];
+            tensor<int32, [4]> var_1507 = const()[name = tensor<string, []>("op_1507"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_21 = reshape(shape = var_1507, x = var_1506)[name = tensor<string, []>("k_21")];
+            tensor<fp32, [1, 72, 1024]> var_1510 = linear(bias = linear_2_bias_0, weight = module_layers_5_self_attn_linear_v_weight_quantized, x = input_299)[name = tensor<string, []>("linear_50")];
+            tensor<int32, [4]> var_1511 = const()[name = tensor<string, []>("op_1511"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_11 = reshape(shape = var_1511, x = var_1510)[name = tensor<string, []>("v_11")];
+            tensor<int32, [4]> value_19_perm_0 = const()[name = tensor<string, []>("value_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_1523 = add(x = q_31, y = module_layers_5_self_attn_pos_bias_u)[name = tensor<string, []>("op_1523")];
+            tensor<fp32, [1, 2, 8, 128]> var_1525 = add(x = q_31, y = module_layers_5_self_attn_pos_bias_v)[name = tensor<string, []>("op_1525")];
+            tensor<int32, [4]> q_with_bias_v_11_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_1527_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1527_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588070144))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588216896))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588216640)))];
+            tensor<bool, []> x_137_transpose_x_0 = const()[name = tensor<string, []>("x_137_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_137_transpose_y_0 = const()[name = tensor<string, []>("x_137_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_11 = transpose(perm = q_with_bias_v_11_perm_0, x = var_1525)[name = tensor<string, []>("transpose_317")];
+            tensor<fp32, [1, 8, 2, 143]> x_137 = matmul(transpose_x = x_137_transpose_x_0, transpose_y = x_137_transpose_y_0, x = q_with_bias_v_11, y = op_1527_quantized)[name = tensor<string, []>("x_137")];
+            tensor<fp32, []> const_144 = const()[name = tensor<string, []>("const_144"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_139_pad_0 = const()[name = tensor<string, []>("x_139_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_139_mode_0 = const()[name = tensor<string, []>("x_139_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_139 = pad(constant_val = const_144, mode = x_139_mode_0, pad = x_139_pad_0, x = x_137)[name = tensor<string, []>("x_139")];
+            tensor<int32, [4]> var_1535 = const()[name = tensor<string, []>("op_1535"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_141 = reshape(shape = var_1535, x = x_139)[name = tensor<string, []>("x_141")];
+            tensor<int32, [4]> var_1539_begin_0 = const()[name = tensor<string, []>("op_1539_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1539_end_0 = const()[name = tensor<string, []>("op_1539_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_1539_end_mask_0 = const()[name = tensor<string, []>("op_1539_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_1539 = slice_by_index(begin = var_1539_begin_0, end = var_1539_end_0, end_mask = var_1539_end_mask_0, x = x_141)[name = tensor<string, []>("op_1539")];
+            tensor<int32, [4]> var_1540 = const()[name = tensor<string, []>("op_1540"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_21 = reshape(shape = var_1540, x = var_1539)[name = tensor<string, []>("matrix_bd_21")];
+            tensor<bool, []> matrix_ac_11_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_11_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_11_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_106_perm_0 = const()[name = tensor<string, []>("transpose_106_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_107_perm_0 = const()[name = tensor<string, []>("transpose_107_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_107 = transpose(perm = transpose_107_perm_0, x = k_21)[name = tensor<string, []>("transpose_315")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_106 = transpose(perm = transpose_106_perm_0, x = var_1523)[name = tensor<string, []>("transpose_316")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_11 = matmul(transpose_x = matrix_ac_11_transpose_x_0, transpose_y = matrix_ac_11_transpose_y_0, x = transpose_106, y = transpose_107)[name = tensor<string, []>("matrix_ac_11")];
+            tensor<int32, [4]> matrix_bd_23_begin_0 = const()[name = tensor<string, []>("matrix_bd_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_23_end_0 = const()[name = tensor<string, []>("matrix_bd_23_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_23_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_23 = slice_by_index(begin = matrix_bd_23_begin_0, end = matrix_bd_23_end_0, end_mask = matrix_bd_23_end_mask_0, x = matrix_bd_21)[name = tensor<string, []>("matrix_bd_23")];
+            tensor<fp32, [1, 8, 2, 72]> var_1549 = add(x = matrix_ac_11, y = matrix_bd_23)[name = tensor<string, []>("op_1549")];
+            tensor<fp32, []> _inversed_scores_21_y_0 = const()[name = tensor<string, []>("_inversed_scores_21_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_21 = mul(x = var_1549, y = _inversed_scores_21_y_0)[name = tensor<string, []>("_inversed_scores_21")];
+            tensor<fp32, [1, 8, 2, 72]> scores_23 = select(a = var_41, b = _inversed_scores_21, cond = mask_11)[name = tensor<string, []>("scores_23")];
+            tensor<fp32, [1, 8, 2, 72]> var_1555 = softmax(axis = var_56, x = scores_23)[name = tensor<string, []>("op_1555")];
+            tensor<fp32, [1, 8, 2, 72]> input_301 = select(a = var_40, b = var_1555, cond = mask_11)[name = tensor<string, []>("input_301")];
+            tensor<bool, []> x_143_transpose_x_0 = const()[name = tensor<string, []>("x_143_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_143_transpose_y_0 = const()[name = tensor<string, []>("x_143_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_19 = transpose(perm = value_19_perm_0, x = v_11)[name = tensor<string, []>("transpose_314")];
+            tensor<fp32, [1, 8, 2, 128]> x_143 = matmul(transpose_x = x_143_transpose_x_0, transpose_y = x_143_transpose_y_0, x = input_301, y = value_19)[name = tensor<string, []>("x_143")];
+            tensor<int32, [4]> var_1559_perm_0 = const()[name = tensor<string, []>("op_1559_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1560 = const()[name = tensor<string, []>("op_1560"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_1559 = transpose(perm = var_1559_perm_0, x = x_143)[name = tensor<string, []>("transpose_313")];
+            tensor<fp32, [1, 2, 1024]> input_303 = reshape(shape = var_1560, x = var_1559)[name = tensor<string, []>("input_303")];
+            tensor<fp32, [1, 2, 1024]> input_305 = linear(bias = linear_2_bias_0, weight = module_layers_5_self_attn_linear_out_weight_quantized, x = input_303)[name = tensor<string, []>("linear_52")];
+            tensor<fp32, [1, 2, 1024]> input_307 = add(x = input_297, y = input_305)[name = tensor<string, []>("input_307")];
+            tensor<int32, [1]> x_147_axes_0 = const()[name = tensor<string, []>("x_147_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_147 = layer_norm(axes = x_147_axes_0, beta = module_layers_5_norm_conv_bias, epsilon = var_38, gamma = module_layers_5_norm_conv_weight, x = input_307)[name = tensor<string, []>("x_147")];
+            tensor<int32, [3]> input_309_perm_0 = const()[name = tensor<string, []>("input_309_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_311_pad_type_0 = const()[name = tensor<string, []>("input_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_311_strides_0 = const()[name = tensor<string, []>("input_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_311_pad_0 = const()[name = tensor<string, []>("input_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_311_dilations_0 = const()[name = tensor<string, []>("input_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_311_groups_0 = const()[name = tensor<string, []>("input_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_309 = transpose(perm = input_309_perm_0, x = x_147)[name = tensor<string, []>("transpose_312")];
+            tensor<fp32, [1, 2048, 2]> input_311 = conv(dilations = input_311_dilations_0, groups = input_311_groups_0, pad = input_311_pad_0, pad_type = input_311_pad_type_0, strides = input_311_strides_0, weight = module_layers_5_conv_pointwise_conv1_weight_quantized, x = input_309)[name = tensor<string, []>("input_311")];
+            tensor<int32, []> x_149_split_num_splits_0 = const()[name = tensor<string, []>("x_149_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_149_split_axis_0 = const()[name = tensor<string, []>("x_149_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_149_split_0, tensor<fp32, [1, 1024, 2]> x_149_split_1 = split(axis = x_149_split_axis_0, num_splits = x_149_split_num_splits_0, x = input_311)[name = tensor<string, []>("x_149_split")];
+            tensor<fp32, [1, 1024, 2]> x_149_split_1_sigmoid = sigmoid(x = x_149_split_1)[name = tensor<string, []>("x_149_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_149 = mul(x = x_149_split_0, y = x_149_split_1_sigmoid)[name = tensor<string, []>("x_149")];
+            tensor<fp32, [1, 1024, 2]> input_313 = select(a = var_40, b = x_149, cond = var_565)[name = tensor<string, []>("input_313")];
+            tensor<bool, []> new_x_23_interleave_0 = const()[name = tensor<string, []>("new_x_23_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_23 = concat(axis = var_56, interleave = new_x_23_interleave_0, values = (cache_23, input_313))[name = tensor<string, []>("new_x_23")];
+            tensor<int32, [3]> var_1598_begin_0 = const()[name = tensor<string, []>("op_1598_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_1598_end_0 = const()[name = tensor<string, []>("op_1598_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_1598_end_mask_0 = const()[name = tensor<string, []>("op_1598_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_1598 = slice_by_index(begin = var_1598_begin_0, end = var_1598_end_0, end_mask = var_1598_end_mask_0, x = new_x_23)[name = tensor<string, []>("op_1598")];
+            tensor<string, []> x_151_pad_type_0 = const()[name = tensor<string, []>("x_151_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_151_groups_0 = const()[name = tensor<string, []>("x_151_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_151_strides_0 = const()[name = tensor<string, []>("x_151_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_151_pad_0 = const()[name = tensor<string, []>("x_151_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_151_dilations_0 = const()[name = tensor<string, []>("x_151_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_151 = conv(dilations = x_151_dilations_0, groups = x_151_groups_0, pad = x_151_pad_0, pad_type = x_151_pad_type_0, strides = x_151_strides_0, weight = module_layers_5_conv_depthwise_conv_weight_quantized, x = new_x_23)[name = tensor<string, []>("x_151")];
+            tensor<int32, [3]> input_315_perm_0 = const()[name = tensor<string, []>("input_315_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_153_axes_0 = const()[name = tensor<string, []>("x_153_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_315 = transpose(perm = input_315_perm_0, x = x_151)[name = tensor<string, []>("transpose_311")];
+            tensor<fp32, [1, 2, 1024]> x_153 = layer_norm(axes = x_153_axes_0, beta = module_layers_5_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_5_conv_batch_norm_weight, x = input_315)[name = tensor<string, []>("x_153")];
+            tensor<int32, [3]> input_317_perm_0 = const()[name = tensor<string, []>("input_317_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_317 = transpose(perm = input_317_perm_0, x = x_153)[name = tensor<string, []>("transpose_310")];
+            tensor<fp32, [1, 1024, 2]> input_319 = silu(x = input_317)[name = tensor<string, []>("input_319")];
+            tensor<string, []> x_155_pad_type_0 = const()[name = tensor<string, []>("x_155_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_155_strides_0 = const()[name = tensor<string, []>("x_155_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_155_pad_0 = const()[name = tensor<string, []>("x_155_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_155_dilations_0 = const()[name = tensor<string, []>("x_155_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_155_groups_0 = const()[name = tensor<string, []>("x_155_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_155 = conv(dilations = x_155_dilations_0, groups = x_155_groups_0, pad = x_155_pad_0, pad_type = x_155_pad_type_0, strides = x_155_strides_0, weight = module_layers_5_conv_pointwise_conv2_weight_quantized, x = input_319)[name = tensor<string, []>("x_155")];
+            tensor<int32, [3]> input_321_perm_0 = const()[name = tensor<string, []>("input_321_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_321 = transpose(perm = input_321_perm_0, x = x_155)[name = tensor<string, []>("transpose_309")];
+            tensor<fp32, [1, 2, 1024]> input_323 = add(x = input_307, y = input_321)[name = tensor<string, []>("input_323")];
+            tensor<int32, [1]> input_325_axes_0 = const()[name = tensor<string, []>("input_325_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_325 = layer_norm(axes = input_325_axes_0, beta = module_layers_5_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_5_norm_feed_forward2_weight, x = input_323)[name = tensor<string, []>("input_325")];
+            tensor<fp32, [1, 2, 4096]> input_327 = linear(bias = linear_1_bias_0, weight = module_layers_5_feed_forward2_linear1_weight_quantized, x = input_325)[name = tensor<string, []>("linear_53")];
+            tensor<fp32, [1, 2, 4096]> input_329 = silu(x = input_327)[name = tensor<string, []>("input_329")];
+            tensor<fp32, [1, 2, 1024]> input_333 = linear(bias = linear_2_bias_0, weight = module_layers_5_feed_forward2_linear2_weight_quantized, x = input_329)[name = tensor<string, []>("linear_54")];
+            tensor<fp32, []> var_1639 = const()[name = tensor<string, []>("op_1639"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1640 = mul(x = input_333, y = var_1639)[name = tensor<string, []>("op_1640")];
+            tensor<fp32, [1, 2, 1024]> input_335 = add(x = input_323, y = var_1640)[name = tensor<string, []>("input_335")];
+            tensor<int32, [1]> input_337_axes_0 = const()[name = tensor<string, []>("input_337_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_337 = layer_norm(axes = input_337_axes_0, beta = module_layers_5_norm_out_bias, epsilon = var_38, gamma = module_layers_5_norm_out_weight, x = input_335)[name = tensor<string, []>("input_337")];
+            tensor<int32, [4]> cache_25_begin_0 = const()[name = tensor<string, []>("cache_25_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_25_end_0 = const()[name = tensor<string, []>("cache_25_end_0"), val = tensor<int32, [4]>([7, 1, 70, 1024])];
+            tensor<bool, [4]> cache_25_end_mask_0 = const()[name = tensor<string, []>("cache_25_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_25_squeeze_mask_0 = const()[name = tensor<string, []>("cache_25_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_25 = slice_by_index(begin = cache_25_begin_0, end = cache_25_end_0, end_mask = cache_25_end_mask_0, squeeze_mask = cache_25_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_25")];
+            tensor<int32, [4]> cache_27_begin_0 = const()[name = tensor<string, []>("cache_27_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
+            tensor<int32, [4]> cache_27_end_0 = const()[name = tensor<string, []>("cache_27_end_0"), val = tensor<int32, [4]>([7, 1, 1024, 8])];
+            tensor<bool, [4]> cache_27_end_mask_0 = const()[name = tensor<string, []>("cache_27_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_27_squeeze_mask_0 = const()[name = tensor<string, []>("cache_27_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_27 = slice_by_index(begin = cache_27_begin_0, end = cache_27_end_0, end_mask = cache_27_end_mask_0, squeeze_mask = cache_27_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_27")];
+            tensor<int32, [1]> input_339_axes_0 = const()[name = tensor<string, []>("input_339_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_339 = layer_norm(axes = input_339_axes_0, beta = module_layers_6_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_6_norm_feed_forward1_weight, x = input_337)[name = tensor<string, []>("input_339")];
+            tensor<fp32, [1, 2, 4096]> input_341 = linear(bias = linear_1_bias_0, weight = module_layers_6_feed_forward1_linear1_weight_quantized, x = input_339)[name = tensor<string, []>("linear_55")];
+            tensor<fp32, [1, 2, 4096]> input_343 = silu(x = input_341)[name = tensor<string, []>("input_343")];
+            tensor<fp32, [1, 2, 1024]> input_347 = linear(bias = linear_2_bias_0, weight = module_layers_6_feed_forward1_linear2_weight_quantized, x = input_343)[name = tensor<string, []>("linear_56")];
+            tensor<fp32, []> var_1674 = const()[name = tensor<string, []>("op_1674"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1675 = mul(x = input_347, y = var_1674)[name = tensor<string, []>("op_1675")];
+            tensor<fp32, [1, 2, 1024]> input_349 = add(x = input_337, y = var_1675)[name = tensor<string, []>("input_349")];
+            tensor<int32, [1]> key_13_axes_0 = const()[name = tensor<string, []>("key_13_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_13 = layer_norm(axes = key_13_axes_0, beta = module_layers_6_norm_self_att_bias, epsilon = var_38, gamma = module_layers_6_norm_self_att_weight, x = input_349)[name = tensor<string, []>("key_13")];
+            tensor<bool, []> input_351_interleave_0 = const()[name = tensor<string, []>("input_351_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_351 = concat(axis = var_65, interleave = input_351_interleave_0, values = (cache_25, key_13))[name = tensor<string, []>("input_351")];
+            tensor<int32, [3]> var_1697_begin_0 = const()[name = tensor<string, []>("op_1697_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_1697_end_0 = const()[name = tensor<string, []>("op_1697_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_1697_end_mask_0 = const()[name = tensor<string, []>("op_1697_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_1697 = slice_by_index(begin = var_1697_begin_0, end = var_1697_end_0, end_mask = var_1697_end_mask_0, x = cache_25)[name = tensor<string, []>("op_1697")];
+            tensor<bool, []> var_1703_interleave_0 = const()[name = tensor<string, []>("op_1703_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_1703 = concat(axis = var_65, interleave = var_1703_interleave_0, values = (var_1697, key_13))[name = tensor<string, []>("op_1703")];
+            tensor<fp32, [1, 2, 1024]> var_1706 = linear(bias = linear_2_bias_0, weight = module_layers_6_self_attn_linear_q_weight_quantized, x = key_13)[name = tensor<string, []>("linear_57")];
+            tensor<int32, [4]> var_1707 = const()[name = tensor<string, []>("op_1707"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_37 = reshape(shape = var_1707, x = var_1706)[name = tensor<string, []>("q_37")];
+            tensor<fp32, [1, 72, 1024]> var_1710 = linear(bias = linear_2_bias_0, weight = module_layers_6_self_attn_linear_k_weight_quantized, x = input_351)[name = tensor<string, []>("linear_58")];
+            tensor<int32, [4]> var_1711 = const()[name = tensor<string, []>("op_1711"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_25 = reshape(shape = var_1711, x = var_1710)[name = tensor<string, []>("k_25")];
+            tensor<fp32, [1, 72, 1024]> var_1714 = linear(bias = linear_2_bias_0, weight = module_layers_6_self_attn_linear_v_weight_quantized, x = input_351)[name = tensor<string, []>("linear_59")];
+            tensor<int32, [4]> var_1715 = const()[name = tensor<string, []>("op_1715"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_13 = reshape(shape = var_1715, x = var_1714)[name = tensor<string, []>("v_13")];
+            tensor<int32, [4]> value_21_perm_0 = const()[name = tensor<string, []>("value_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_1727 = add(x = q_37, y = module_layers_6_self_attn_pos_bias_u)[name = tensor<string, []>("op_1727")];
+            tensor<fp32, [1, 2, 8, 128]> var_1729 = add(x = q_37, y = module_layers_6_self_attn_pos_bias_v)[name = tensor<string, []>("op_1729")];
+            tensor<int32, [4]> q_with_bias_v_13_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_1731_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1731_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588217536))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588364288))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588364032)))];
+            tensor<bool, []> x_163_transpose_x_0 = const()[name = tensor<string, []>("x_163_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_163_transpose_y_0 = const()[name = tensor<string, []>("x_163_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_13 = transpose(perm = q_with_bias_v_13_perm_0, x = var_1729)[name = tensor<string, []>("transpose_308")];
+            tensor<fp32, [1, 8, 2, 143]> x_163 = matmul(transpose_x = x_163_transpose_x_0, transpose_y = x_163_transpose_y_0, x = q_with_bias_v_13, y = op_1731_quantized)[name = tensor<string, []>("x_163")];
+            tensor<fp32, []> const_157 = const()[name = tensor<string, []>("const_157"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_165_pad_0 = const()[name = tensor<string, []>("x_165_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_165_mode_0 = const()[name = tensor<string, []>("x_165_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_165 = pad(constant_val = const_157, mode = x_165_mode_0, pad = x_165_pad_0, x = x_163)[name = tensor<string, []>("x_165")];
+            tensor<int32, [4]> var_1739 = const()[name = tensor<string, []>("op_1739"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_167 = reshape(shape = var_1739, x = x_165)[name = tensor<string, []>("x_167")];
+            tensor<int32, [4]> var_1743_begin_0 = const()[name = tensor<string, []>("op_1743_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1743_end_0 = const()[name = tensor<string, []>("op_1743_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_1743_end_mask_0 = const()[name = tensor<string, []>("op_1743_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_1743 = slice_by_index(begin = var_1743_begin_0, end = var_1743_end_0, end_mask = var_1743_end_mask_0, x = x_167)[name = tensor<string, []>("op_1743")];
+            tensor<int32, [4]> var_1744 = const()[name = tensor<string, []>("op_1744"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_25 = reshape(shape = var_1744, x = var_1743)[name = tensor<string, []>("matrix_bd_25")];
+            tensor<bool, []> matrix_ac_13_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_13_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_13_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_108_perm_0 = const()[name = tensor<string, []>("transpose_108_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_109_perm_0 = const()[name = tensor<string, []>("transpose_109_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_109 = transpose(perm = transpose_109_perm_0, x = k_25)[name = tensor<string, []>("transpose_306")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_108 = transpose(perm = transpose_108_perm_0, x = var_1727)[name = tensor<string, []>("transpose_307")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_13 = matmul(transpose_x = matrix_ac_13_transpose_x_0, transpose_y = matrix_ac_13_transpose_y_0, x = transpose_108, y = transpose_109)[name = tensor<string, []>("matrix_ac_13")];
+            tensor<int32, [4]> matrix_bd_27_begin_0 = const()[name = tensor<string, []>("matrix_bd_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_27_end_0 = const()[name = tensor<string, []>("matrix_bd_27_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_27_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_27 = slice_by_index(begin = matrix_bd_27_begin_0, end = matrix_bd_27_end_0, end_mask = matrix_bd_27_end_mask_0, x = matrix_bd_25)[name = tensor<string, []>("matrix_bd_27")];
+            tensor<fp32, [1, 8, 2, 72]> var_1753 = add(x = matrix_ac_13, y = matrix_bd_27)[name = tensor<string, []>("op_1753")];
+            tensor<fp32, []> _inversed_scores_25_y_0 = const()[name = tensor<string, []>("_inversed_scores_25_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_25 = mul(x = var_1753, y = _inversed_scores_25_y_0)[name = tensor<string, []>("_inversed_scores_25")];
+            tensor<fp32, [1, 8, 2, 72]> scores_27 = select(a = var_41, b = _inversed_scores_25, cond = mask_11)[name = tensor<string, []>("scores_27")];
+            tensor<fp32, [1, 8, 2, 72]> var_1759 = softmax(axis = var_56, x = scores_27)[name = tensor<string, []>("op_1759")];
+            tensor<fp32, [1, 8, 2, 72]> input_353 = select(a = var_40, b = var_1759, cond = mask_11)[name = tensor<string, []>("input_353")];
+            tensor<bool, []> x_169_transpose_x_0 = const()[name = tensor<string, []>("x_169_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_169_transpose_y_0 = const()[name = tensor<string, []>("x_169_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_21 = transpose(perm = value_21_perm_0, x = v_13)[name = tensor<string, []>("transpose_305")];
+            tensor<fp32, [1, 8, 2, 128]> x_169 = matmul(transpose_x = x_169_transpose_x_0, transpose_y = x_169_transpose_y_0, x = input_353, y = value_21)[name = tensor<string, []>("x_169")];
+            tensor<int32, [4]> var_1763_perm_0 = const()[name = tensor<string, []>("op_1763_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1764 = const()[name = tensor<string, []>("op_1764"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_1763 = transpose(perm = var_1763_perm_0, x = x_169)[name = tensor<string, []>("transpose_304")];
+            tensor<fp32, [1, 2, 1024]> input_355 = reshape(shape = var_1764, x = var_1763)[name = tensor<string, []>("input_355")];
+            tensor<fp32, [1, 2, 1024]> input_357 = linear(bias = linear_2_bias_0, weight = module_layers_6_self_attn_linear_out_weight_quantized, x = input_355)[name = tensor<string, []>("linear_61")];
+            tensor<fp32, [1, 2, 1024]> input_359 = add(x = input_349, y = input_357)[name = tensor<string, []>("input_359")];
+            tensor<int32, [1]> x_173_axes_0 = const()[name = tensor<string, []>("x_173_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_173 = layer_norm(axes = x_173_axes_0, beta = module_layers_6_norm_conv_bias, epsilon = var_38, gamma = module_layers_6_norm_conv_weight, x = input_359)[name = tensor<string, []>("x_173")];
+            tensor<int32, [3]> input_361_perm_0 = const()[name = tensor<string, []>("input_361_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_363_pad_type_0 = const()[name = tensor<string, []>("input_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_363_strides_0 = const()[name = tensor<string, []>("input_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_363_pad_0 = const()[name = tensor<string, []>("input_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_363_dilations_0 = const()[name = tensor<string, []>("input_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_363_groups_0 = const()[name = tensor<string, []>("input_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_361 = transpose(perm = input_361_perm_0, x = x_173)[name = tensor<string, []>("transpose_303")];
+            tensor<fp32, [1, 2048, 2]> input_363 = conv(dilations = input_363_dilations_0, groups = input_363_groups_0, pad = input_363_pad_0, pad_type = input_363_pad_type_0, strides = input_363_strides_0, weight = module_layers_6_conv_pointwise_conv1_weight_quantized, x = input_361)[name = tensor<string, []>("input_363")];
+            tensor<int32, []> x_175_split_num_splits_0 = const()[name = tensor<string, []>("x_175_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_175_split_axis_0 = const()[name = tensor<string, []>("x_175_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_175_split_0, tensor<fp32, [1, 1024, 2]> x_175_split_1 = split(axis = x_175_split_axis_0, num_splits = x_175_split_num_splits_0, x = input_363)[name = tensor<string, []>("x_175_split")];
+            tensor<fp32, [1, 1024, 2]> x_175_split_1_sigmoid = sigmoid(x = x_175_split_1)[name = tensor<string, []>("x_175_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_175 = mul(x = x_175_split_0, y = x_175_split_1_sigmoid)[name = tensor<string, []>("x_175")];
+            tensor<fp32, [1, 1024, 2]> input_365 = select(a = var_40, b = x_175, cond = var_565)[name = tensor<string, []>("input_365")];
+            tensor<bool, []> new_x_27_interleave_0 = const()[name = tensor<string, []>("new_x_27_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_27 = concat(axis = var_56, interleave = new_x_27_interleave_0, values = (cache_27, input_365))[name = tensor<string, []>("new_x_27")];
+            tensor<int32, [3]> var_1802_begin_0 = const()[name = tensor<string, []>("op_1802_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_1802_end_0 = const()[name = tensor<string, []>("op_1802_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_1802_end_mask_0 = const()[name = tensor<string, []>("op_1802_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_1802 = slice_by_index(begin = var_1802_begin_0, end = var_1802_end_0, end_mask = var_1802_end_mask_0, x = new_x_27)[name = tensor<string, []>("op_1802")];
+            tensor<string, []> x_177_pad_type_0 = const()[name = tensor<string, []>("x_177_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_177_groups_0 = const()[name = tensor<string, []>("x_177_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_177_strides_0 = const()[name = tensor<string, []>("x_177_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_177_pad_0 = const()[name = tensor<string, []>("x_177_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_177_dilations_0 = const()[name = tensor<string, []>("x_177_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_177 = conv(dilations = x_177_dilations_0, groups = x_177_groups_0, pad = x_177_pad_0, pad_type = x_177_pad_type_0, strides = x_177_strides_0, weight = module_layers_6_conv_depthwise_conv_weight_quantized, x = new_x_27)[name = tensor<string, []>("x_177")];
+            tensor<int32, [3]> input_367_perm_0 = const()[name = tensor<string, []>("input_367_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_179_axes_0 = const()[name = tensor<string, []>("x_179_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_367 = transpose(perm = input_367_perm_0, x = x_177)[name = tensor<string, []>("transpose_302")];
+            tensor<fp32, [1, 2, 1024]> x_179 = layer_norm(axes = x_179_axes_0, beta = module_layers_6_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_6_conv_batch_norm_weight, x = input_367)[name = tensor<string, []>("x_179")];
+            tensor<int32, [3]> input_369_perm_0 = const()[name = tensor<string, []>("input_369_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_369 = transpose(perm = input_369_perm_0, x = x_179)[name = tensor<string, []>("transpose_301")];
+            tensor<fp32, [1, 1024, 2]> input_371 = silu(x = input_369)[name = tensor<string, []>("input_371")];
+            tensor<string, []> x_181_pad_type_0 = const()[name = tensor<string, []>("x_181_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_181_strides_0 = const()[name = tensor<string, []>("x_181_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_181_pad_0 = const()[name = tensor<string, []>("x_181_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_181_dilations_0 = const()[name = tensor<string, []>("x_181_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_181_groups_0 = const()[name = tensor<string, []>("x_181_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_181 = conv(dilations = x_181_dilations_0, groups = x_181_groups_0, pad = x_181_pad_0, pad_type = x_181_pad_type_0, strides = x_181_strides_0, weight = module_layers_6_conv_pointwise_conv2_weight_quantized, x = input_371)[name = tensor<string, []>("x_181")];
+            tensor<int32, [3]> input_373_perm_0 = const()[name = tensor<string, []>("input_373_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_373 = transpose(perm = input_373_perm_0, x = x_181)[name = tensor<string, []>("transpose_300")];
+            tensor<fp32, [1, 2, 1024]> input_375 = add(x = input_359, y = input_373)[name = tensor<string, []>("input_375")];
+            tensor<int32, [1]> input_377_axes_0 = const()[name = tensor<string, []>("input_377_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_377 = layer_norm(axes = input_377_axes_0, beta = module_layers_6_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_6_norm_feed_forward2_weight, x = input_375)[name = tensor<string, []>("input_377")];
+            tensor<fp32, [1, 2, 4096]> input_379 = linear(bias = linear_1_bias_0, weight = module_layers_6_feed_forward2_linear1_weight_quantized, x = input_377)[name = tensor<string, []>("linear_62")];
+            tensor<fp32, [1, 2, 4096]> input_381 = silu(x = input_379)[name = tensor<string, []>("input_381")];
+            tensor<fp32, [1, 2, 1024]> input_385 = linear(bias = linear_2_bias_0, weight = module_layers_6_feed_forward2_linear2_weight_quantized, x = input_381)[name = tensor<string, []>("linear_63")];
+            tensor<fp32, []> var_1843 = const()[name = tensor<string, []>("op_1843"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1844 = mul(x = input_385, y = var_1843)[name = tensor<string, []>("op_1844")];
+            tensor<fp32, [1, 2, 1024]> input_387 = add(x = input_375, y = var_1844)[name = tensor<string, []>("input_387")];
+            tensor<int32, [1]> input_389_axes_0 = const()[name = tensor<string, []>("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_389 = layer_norm(axes = input_389_axes_0, beta = module_layers_6_norm_out_bias, epsilon = var_38, gamma = module_layers_6_norm_out_weight, x = input_387)[name = tensor<string, []>("input_389")];
+            tensor<int32, [4]> cache_29_begin_0 = const()[name = tensor<string, []>("cache_29_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_29_end_0 = const()[name = tensor<string, []>("cache_29_end_0"), val = tensor<int32, [4]>([8, 1, 70, 1024])];
+            tensor<bool, [4]> cache_29_end_mask_0 = const()[name = tensor<string, []>("cache_29_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_29_squeeze_mask_0 = const()[name = tensor<string, []>("cache_29_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_29 = slice_by_index(begin = cache_29_begin_0, end = cache_29_end_0, end_mask = cache_29_end_mask_0, squeeze_mask = cache_29_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_29")];
+            tensor<int32, [4]> cache_31_begin_0 = const()[name = tensor<string, []>("cache_31_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
+            tensor<int32, [4]> cache_31_end_0 = const()[name = tensor<string, []>("cache_31_end_0"), val = tensor<int32, [4]>([8, 1, 1024, 8])];
+            tensor<bool, [4]> cache_31_end_mask_0 = const()[name = tensor<string, []>("cache_31_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_31_squeeze_mask_0 = const()[name = tensor<string, []>("cache_31_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_31 = slice_by_index(begin = cache_31_begin_0, end = cache_31_end_0, end_mask = cache_31_end_mask_0, squeeze_mask = cache_31_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_31")];
+            tensor<int32, [1]> input_391_axes_0 = const()[name = tensor<string, []>("input_391_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_391 = layer_norm(axes = input_391_axes_0, beta = module_layers_7_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_7_norm_feed_forward1_weight, x = input_389)[name = tensor<string, []>("input_391")];
+            tensor<fp32, [1, 2, 4096]> input_393 = linear(bias = linear_1_bias_0, weight = module_layers_7_feed_forward1_linear1_weight_quantized, x = input_391)[name = tensor<string, []>("linear_64")];
+            tensor<fp32, [1, 2, 4096]> input_395 = silu(x = input_393)[name = tensor<string, []>("input_395")];
+            tensor<fp32, [1, 2, 1024]> input_399 = linear(bias = linear_2_bias_0, weight = module_layers_7_feed_forward1_linear2_weight_quantized, x = input_395)[name = tensor<string, []>("linear_65")];
+            tensor<fp32, []> var_1878 = const()[name = tensor<string, []>("op_1878"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_1879 = mul(x = input_399, y = var_1878)[name = tensor<string, []>("op_1879")];
+            tensor<fp32, [1, 2, 1024]> input_401 = add(x = input_389, y = var_1879)[name = tensor<string, []>("input_401")];
+            tensor<int32, [1]> key_15_axes_0 = const()[name = tensor<string, []>("key_15_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_15 = layer_norm(axes = key_15_axes_0, beta = module_layers_7_norm_self_att_bias, epsilon = var_38, gamma = module_layers_7_norm_self_att_weight, x = input_401)[name = tensor<string, []>("key_15")];
+            tensor<bool, []> input_403_interleave_0 = const()[name = tensor<string, []>("input_403_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_403 = concat(axis = var_65, interleave = input_403_interleave_0, values = (cache_29, key_15))[name = tensor<string, []>("input_403")];
+            tensor<int32, [3]> var_1901_begin_0 = const()[name = tensor<string, []>("op_1901_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_1901_end_0 = const()[name = tensor<string, []>("op_1901_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_1901_end_mask_0 = const()[name = tensor<string, []>("op_1901_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_1901 = slice_by_index(begin = var_1901_begin_0, end = var_1901_end_0, end_mask = var_1901_end_mask_0, x = cache_29)[name = tensor<string, []>("op_1901")];
+            tensor<bool, []> var_1907_interleave_0 = const()[name = tensor<string, []>("op_1907_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_1907 = concat(axis = var_65, interleave = var_1907_interleave_0, values = (var_1901, key_15))[name = tensor<string, []>("op_1907")];
+            tensor<fp32, [1, 2, 1024]> var_1910 = linear(bias = linear_2_bias_0, weight = module_layers_7_self_attn_linear_q_weight_quantized, x = key_15)[name = tensor<string, []>("linear_66")];
+            tensor<int32, [4]> var_1911 = const()[name = tensor<string, []>("op_1911"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_43 = reshape(shape = var_1911, x = var_1910)[name = tensor<string, []>("q_43")];
+            tensor<fp32, [1, 72, 1024]> var_1914 = linear(bias = linear_2_bias_0, weight = module_layers_7_self_attn_linear_k_weight_quantized, x = input_403)[name = tensor<string, []>("linear_67")];
+            tensor<int32, [4]> var_1915 = const()[name = tensor<string, []>("op_1915"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_29 = reshape(shape = var_1915, x = var_1914)[name = tensor<string, []>("k_29")];
+            tensor<fp32, [1, 72, 1024]> var_1918 = linear(bias = linear_2_bias_0, weight = module_layers_7_self_attn_linear_v_weight_quantized, x = input_403)[name = tensor<string, []>("linear_68")];
+            tensor<int32, [4]> var_1919 = const()[name = tensor<string, []>("op_1919"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_15 = reshape(shape = var_1919, x = var_1918)[name = tensor<string, []>("v_15")];
+            tensor<int32, [4]> value_23_perm_0 = const()[name = tensor<string, []>("value_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_1931 = add(x = q_43, y = module_layers_7_self_attn_pos_bias_u)[name = tensor<string, []>("op_1931")];
+            tensor<fp32, [1, 2, 8, 128]> var_1933 = add(x = q_43, y = module_layers_7_self_attn_pos_bias_v)[name = tensor<string, []>("op_1933")];
+            tensor<int32, [4]> q_with_bias_v_15_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_1935_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_1935_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588364928))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588511680))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588511424)))];
+            tensor<bool, []> x_189_transpose_x_0 = const()[name = tensor<string, []>("x_189_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_189_transpose_y_0 = const()[name = tensor<string, []>("x_189_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_15 = transpose(perm = q_with_bias_v_15_perm_0, x = var_1933)[name = tensor<string, []>("transpose_299")];
+            tensor<fp32, [1, 8, 2, 143]> x_189 = matmul(transpose_x = x_189_transpose_x_0, transpose_y = x_189_transpose_y_0, x = q_with_bias_v_15, y = op_1935_quantized)[name = tensor<string, []>("x_189")];
+            tensor<fp32, []> const_170 = const()[name = tensor<string, []>("const_170"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_191_pad_0 = const()[name = tensor<string, []>("x_191_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_191_mode_0 = const()[name = tensor<string, []>("x_191_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_191 = pad(constant_val = const_170, mode = x_191_mode_0, pad = x_191_pad_0, x = x_189)[name = tensor<string, []>("x_191")];
+            tensor<int32, [4]> var_1943 = const()[name = tensor<string, []>("op_1943"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_193 = reshape(shape = var_1943, x = x_191)[name = tensor<string, []>("x_193")];
+            tensor<int32, [4]> var_1947_begin_0 = const()[name = tensor<string, []>("op_1947_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_1947_end_0 = const()[name = tensor<string, []>("op_1947_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_1947_end_mask_0 = const()[name = tensor<string, []>("op_1947_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_1947 = slice_by_index(begin = var_1947_begin_0, end = var_1947_end_0, end_mask = var_1947_end_mask_0, x = x_193)[name = tensor<string, []>("op_1947")];
+            tensor<int32, [4]> var_1948 = const()[name = tensor<string, []>("op_1948"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_29 = reshape(shape = var_1948, x = var_1947)[name = tensor<string, []>("matrix_bd_29")];
+            tensor<bool, []> matrix_ac_15_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_15_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_15_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_110_perm_0 = const()[name = tensor<string, []>("transpose_110_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_111_perm_0 = const()[name = tensor<string, []>("transpose_111_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_111 = transpose(perm = transpose_111_perm_0, x = k_29)[name = tensor<string, []>("transpose_297")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_110 = transpose(perm = transpose_110_perm_0, x = var_1931)[name = tensor<string, []>("transpose_298")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_15 = matmul(transpose_x = matrix_ac_15_transpose_x_0, transpose_y = matrix_ac_15_transpose_y_0, x = transpose_110, y = transpose_111)[name = tensor<string, []>("matrix_ac_15")];
+            tensor<int32, [4]> matrix_bd_31_begin_0 = const()[name = tensor<string, []>("matrix_bd_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_31_end_0 = const()[name = tensor<string, []>("matrix_bd_31_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_31_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_31 = slice_by_index(begin = matrix_bd_31_begin_0, end = matrix_bd_31_end_0, end_mask = matrix_bd_31_end_mask_0, x = matrix_bd_29)[name = tensor<string, []>("matrix_bd_31")];
+            tensor<fp32, [1, 8, 2, 72]> var_1957 = add(x = matrix_ac_15, y = matrix_bd_31)[name = tensor<string, []>("op_1957")];
+            tensor<fp32, []> _inversed_scores_29_y_0 = const()[name = tensor<string, []>("_inversed_scores_29_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_29 = mul(x = var_1957, y = _inversed_scores_29_y_0)[name = tensor<string, []>("_inversed_scores_29")];
+            tensor<fp32, [1, 8, 2, 72]> scores_31 = select(a = var_41, b = _inversed_scores_29, cond = mask_11)[name = tensor<string, []>("scores_31")];
+            tensor<fp32, [1, 8, 2, 72]> var_1963 = softmax(axis = var_56, x = scores_31)[name = tensor<string, []>("op_1963")];
+            tensor<fp32, [1, 8, 2, 72]> input_405 = select(a = var_40, b = var_1963, cond = mask_11)[name = tensor<string, []>("input_405")];
+            tensor<bool, []> x_195_transpose_x_0 = const()[name = tensor<string, []>("x_195_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_195_transpose_y_0 = const()[name = tensor<string, []>("x_195_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_23 = transpose(perm = value_23_perm_0, x = v_15)[name = tensor<string, []>("transpose_296")];
+            tensor<fp32, [1, 8, 2, 128]> x_195 = matmul(transpose_x = x_195_transpose_x_0, transpose_y = x_195_transpose_y_0, x = input_405, y = value_23)[name = tensor<string, []>("x_195")];
+            tensor<int32, [4]> var_1967_perm_0 = const()[name = tensor<string, []>("op_1967_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_1968 = const()[name = tensor<string, []>("op_1968"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_1967 = transpose(perm = var_1967_perm_0, x = x_195)[name = tensor<string, []>("transpose_295")];
+            tensor<fp32, [1, 2, 1024]> input_407 = reshape(shape = var_1968, x = var_1967)[name = tensor<string, []>("input_407")];
+            tensor<fp32, [1, 2, 1024]> input_409 = linear(bias = linear_2_bias_0, weight = module_layers_7_self_attn_linear_out_weight_quantized, x = input_407)[name = tensor<string, []>("linear_70")];
+            tensor<fp32, [1, 2, 1024]> input_411 = add(x = input_401, y = input_409)[name = tensor<string, []>("input_411")];
+            tensor<int32, [1]> x_199_axes_0 = const()[name = tensor<string, []>("x_199_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_199 = layer_norm(axes = x_199_axes_0, beta = module_layers_7_norm_conv_bias, epsilon = var_38, gamma = module_layers_7_norm_conv_weight, x = input_411)[name = tensor<string, []>("x_199")];
+            tensor<int32, [3]> input_413_perm_0 = const()[name = tensor<string, []>("input_413_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_415_pad_type_0 = const()[name = tensor<string, []>("input_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_415_strides_0 = const()[name = tensor<string, []>("input_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_415_pad_0 = const()[name = tensor<string, []>("input_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_415_dilations_0 = const()[name = tensor<string, []>("input_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_415_groups_0 = const()[name = tensor<string, []>("input_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_413 = transpose(perm = input_413_perm_0, x = x_199)[name = tensor<string, []>("transpose_294")];
+            tensor<fp32, [1, 2048, 2]> input_415 = conv(dilations = input_415_dilations_0, groups = input_415_groups_0, pad = input_415_pad_0, pad_type = input_415_pad_type_0, strides = input_415_strides_0, weight = module_layers_7_conv_pointwise_conv1_weight_quantized, x = input_413)[name = tensor<string, []>("input_415")];
+            tensor<int32, []> x_201_split_num_splits_0 = const()[name = tensor<string, []>("x_201_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_201_split_axis_0 = const()[name = tensor<string, []>("x_201_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_201_split_0, tensor<fp32, [1, 1024, 2]> x_201_split_1 = split(axis = x_201_split_axis_0, num_splits = x_201_split_num_splits_0, x = input_415)[name = tensor<string, []>("x_201_split")];
+            tensor<fp32, [1, 1024, 2]> x_201_split_1_sigmoid = sigmoid(x = x_201_split_1)[name = tensor<string, []>("x_201_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_201 = mul(x = x_201_split_0, y = x_201_split_1_sigmoid)[name = tensor<string, []>("x_201")];
+            tensor<fp32, [1, 1024, 2]> input_417 = select(a = var_40, b = x_201, cond = var_565)[name = tensor<string, []>("input_417")];
+            tensor<bool, []> new_x_31_interleave_0 = const()[name = tensor<string, []>("new_x_31_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_31 = concat(axis = var_56, interleave = new_x_31_interleave_0, values = (cache_31, input_417))[name = tensor<string, []>("new_x_31")];
+            tensor<int32, [3]> var_2006_begin_0 = const()[name = tensor<string, []>("op_2006_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_2006_end_0 = const()[name = tensor<string, []>("op_2006_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_2006_end_mask_0 = const()[name = tensor<string, []>("op_2006_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_2006 = slice_by_index(begin = var_2006_begin_0, end = var_2006_end_0, end_mask = var_2006_end_mask_0, x = new_x_31)[name = tensor<string, []>("op_2006")];
+            tensor<string, []> x_203_pad_type_0 = const()[name = tensor<string, []>("x_203_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_203_groups_0 = const()[name = tensor<string, []>("x_203_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_203_strides_0 = const()[name = tensor<string, []>("x_203_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_203_pad_0 = const()[name = tensor<string, []>("x_203_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_203_dilations_0 = const()[name = tensor<string, []>("x_203_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_203 = conv(dilations = x_203_dilations_0, groups = x_203_groups_0, pad = x_203_pad_0, pad_type = x_203_pad_type_0, strides = x_203_strides_0, weight = module_layers_7_conv_depthwise_conv_weight_quantized, x = new_x_31)[name = tensor<string, []>("x_203")];
+            tensor<int32, [3]> input_419_perm_0 = const()[name = tensor<string, []>("input_419_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_205_axes_0 = const()[name = tensor<string, []>("x_205_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_419 = transpose(perm = input_419_perm_0, x = x_203)[name = tensor<string, []>("transpose_293")];
+            tensor<fp32, [1, 2, 1024]> x_205 = layer_norm(axes = x_205_axes_0, beta = module_layers_7_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_7_conv_batch_norm_weight, x = input_419)[name = tensor<string, []>("x_205")];
+            tensor<int32, [3]> input_421_perm_0 = const()[name = tensor<string, []>("input_421_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_421 = transpose(perm = input_421_perm_0, x = x_205)[name = tensor<string, []>("transpose_292")];
+            tensor<fp32, [1, 1024, 2]> input_423 = silu(x = input_421)[name = tensor<string, []>("input_423")];
+            tensor<string, []> x_207_pad_type_0 = const()[name = tensor<string, []>("x_207_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_207_strides_0 = const()[name = tensor<string, []>("x_207_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_207_pad_0 = const()[name = tensor<string, []>("x_207_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_207_dilations_0 = const()[name = tensor<string, []>("x_207_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_207_groups_0 = const()[name = tensor<string, []>("x_207_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_207 = conv(dilations = x_207_dilations_0, groups = x_207_groups_0, pad = x_207_pad_0, pad_type = x_207_pad_type_0, strides = x_207_strides_0, weight = module_layers_7_conv_pointwise_conv2_weight_quantized, x = input_423)[name = tensor<string, []>("x_207")];
+            tensor<int32, [3]> input_425_perm_0 = const()[name = tensor<string, []>("input_425_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_425 = transpose(perm = input_425_perm_0, x = x_207)[name = tensor<string, []>("transpose_291")];
+            tensor<fp32, [1, 2, 1024]> input_427 = add(x = input_411, y = input_425)[name = tensor<string, []>("input_427")];
+            tensor<int32, [1]> input_429_axes_0 = const()[name = tensor<string, []>("input_429_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_429 = layer_norm(axes = input_429_axes_0, beta = module_layers_7_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_7_norm_feed_forward2_weight, x = input_427)[name = tensor<string, []>("input_429")];
+            tensor<fp32, [1, 2, 4096]> input_431 = linear(bias = linear_1_bias_0, weight = module_layers_7_feed_forward2_linear1_weight_quantized, x = input_429)[name = tensor<string, []>("linear_71")];
+            tensor<fp32, [1, 2, 4096]> input_433 = silu(x = input_431)[name = tensor<string, []>("input_433")];
+            tensor<fp32, [1, 2, 1024]> input_437 = linear(bias = linear_2_bias_0, weight = module_layers_7_feed_forward2_linear2_weight_quantized, x = input_433)[name = tensor<string, []>("linear_72")];
+            tensor<fp32, []> var_2047 = const()[name = tensor<string, []>("op_2047"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2048 = mul(x = input_437, y = var_2047)[name = tensor<string, []>("op_2048")];
+            tensor<fp32, [1, 2, 1024]> input_439 = add(x = input_427, y = var_2048)[name = tensor<string, []>("input_439")];
+            tensor<int32, [1]> input_441_axes_0 = const()[name = tensor<string, []>("input_441_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_441 = layer_norm(axes = input_441_axes_0, beta = module_layers_7_norm_out_bias, epsilon = var_38, gamma = module_layers_7_norm_out_weight, x = input_439)[name = tensor<string, []>("input_441")];
+            tensor<int32, [4]> cache_33_begin_0 = const()[name = tensor<string, []>("cache_33_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_33_end_0 = const()[name = tensor<string, []>("cache_33_end_0"), val = tensor<int32, [4]>([9, 1, 70, 1024])];
+            tensor<bool, [4]> cache_33_end_mask_0 = const()[name = tensor<string, []>("cache_33_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_33_squeeze_mask_0 = const()[name = tensor<string, []>("cache_33_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_33 = slice_by_index(begin = cache_33_begin_0, end = cache_33_end_0, end_mask = cache_33_end_mask_0, squeeze_mask = cache_33_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_33")];
+            tensor<int32, [4]> cache_35_begin_0 = const()[name = tensor<string, []>("cache_35_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
+            tensor<int32, [4]> cache_35_end_0 = const()[name = tensor<string, []>("cache_35_end_0"), val = tensor<int32, [4]>([9, 1, 1024, 8])];
+            tensor<bool, [4]> cache_35_end_mask_0 = const()[name = tensor<string, []>("cache_35_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_35_squeeze_mask_0 = const()[name = tensor<string, []>("cache_35_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_35 = slice_by_index(begin = cache_35_begin_0, end = cache_35_end_0, end_mask = cache_35_end_mask_0, squeeze_mask = cache_35_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_35")];
+            tensor<int32, [1]> input_443_axes_0 = const()[name = tensor<string, []>("input_443_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_443 = layer_norm(axes = input_443_axes_0, beta = module_layers_8_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_8_norm_feed_forward1_weight, x = input_441)[name = tensor<string, []>("input_443")];
+            tensor<fp32, [1, 2, 4096]> input_445 = linear(bias = linear_1_bias_0, weight = module_layers_8_feed_forward1_linear1_weight_quantized, x = input_443)[name = tensor<string, []>("linear_73")];
+            tensor<fp32, [1, 2, 4096]> input_447 = silu(x = input_445)[name = tensor<string, []>("input_447")];
+            tensor<fp32, [1, 2, 1024]> input_451 = linear(bias = linear_2_bias_0, weight = module_layers_8_feed_forward1_linear2_weight_quantized, x = input_447)[name = tensor<string, []>("linear_74")];
+            tensor<fp32, []> var_2082 = const()[name = tensor<string, []>("op_2082"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2083 = mul(x = input_451, y = var_2082)[name = tensor<string, []>("op_2083")];
+            tensor<fp32, [1, 2, 1024]> input_453 = add(x = input_441, y = var_2083)[name = tensor<string, []>("input_453")];
+            tensor<int32, [1]> key_17_axes_0 = const()[name = tensor<string, []>("key_17_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_17 = layer_norm(axes = key_17_axes_0, beta = module_layers_8_norm_self_att_bias, epsilon = var_38, gamma = module_layers_8_norm_self_att_weight, x = input_453)[name = tensor<string, []>("key_17")];
+            tensor<bool, []> input_455_interleave_0 = const()[name = tensor<string, []>("input_455_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_455 = concat(axis = var_65, interleave = input_455_interleave_0, values = (cache_33, key_17))[name = tensor<string, []>("input_455")];
+            tensor<int32, [3]> var_2105_begin_0 = const()[name = tensor<string, []>("op_2105_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_2105_end_0 = const()[name = tensor<string, []>("op_2105_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_2105_end_mask_0 = const()[name = tensor<string, []>("op_2105_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_2105 = slice_by_index(begin = var_2105_begin_0, end = var_2105_end_0, end_mask = var_2105_end_mask_0, x = cache_33)[name = tensor<string, []>("op_2105")];
+            tensor<bool, []> var_2111_interleave_0 = const()[name = tensor<string, []>("op_2111_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_2111 = concat(axis = var_65, interleave = var_2111_interleave_0, values = (var_2105, key_17))[name = tensor<string, []>("op_2111")];
+            tensor<fp32, [1, 2, 1024]> var_2114 = linear(bias = linear_2_bias_0, weight = module_layers_8_self_attn_linear_q_weight_quantized, x = key_17)[name = tensor<string, []>("linear_75")];
+            tensor<int32, [4]> var_2115 = const()[name = tensor<string, []>("op_2115"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_49 = reshape(shape = var_2115, x = var_2114)[name = tensor<string, []>("q_49")];
+            tensor<fp32, [1, 72, 1024]> var_2118 = linear(bias = linear_2_bias_0, weight = module_layers_8_self_attn_linear_k_weight_quantized, x = input_455)[name = tensor<string, []>("linear_76")];
+            tensor<int32, [4]> var_2119 = const()[name = tensor<string, []>("op_2119"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_33 = reshape(shape = var_2119, x = var_2118)[name = tensor<string, []>("k_33")];
+            tensor<fp32, [1, 72, 1024]> var_2122 = linear(bias = linear_2_bias_0, weight = module_layers_8_self_attn_linear_v_weight_quantized, x = input_455)[name = tensor<string, []>("linear_77")];
+            tensor<int32, [4]> var_2123 = const()[name = tensor<string, []>("op_2123"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_17 = reshape(shape = var_2123, x = var_2122)[name = tensor<string, []>("v_17")];
+            tensor<int32, [4]> value_25_perm_0 = const()[name = tensor<string, []>("value_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_2135 = add(x = q_49, y = module_layers_8_self_attn_pos_bias_u)[name = tensor<string, []>("op_2135")];
+            tensor<fp32, [1, 2, 8, 128]> var_2137 = add(x = q_49, y = module_layers_8_self_attn_pos_bias_v)[name = tensor<string, []>("op_2137")];
+            tensor<int32, [4]> q_with_bias_v_17_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_2139_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2139_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588512320))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588659072))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588658816)))];
+            tensor<bool, []> x_215_transpose_x_0 = const()[name = tensor<string, []>("x_215_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_215_transpose_y_0 = const()[name = tensor<string, []>("x_215_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_17 = transpose(perm = q_with_bias_v_17_perm_0, x = var_2137)[name = tensor<string, []>("transpose_290")];
+            tensor<fp32, [1, 8, 2, 143]> x_215 = matmul(transpose_x = x_215_transpose_x_0, transpose_y = x_215_transpose_y_0, x = q_with_bias_v_17, y = op_2139_quantized)[name = tensor<string, []>("x_215")];
+            tensor<fp32, []> const_183 = const()[name = tensor<string, []>("const_183"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_217_pad_0 = const()[name = tensor<string, []>("x_217_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_217_mode_0 = const()[name = tensor<string, []>("x_217_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_217 = pad(constant_val = const_183, mode = x_217_mode_0, pad = x_217_pad_0, x = x_215)[name = tensor<string, []>("x_217")];
+            tensor<int32, [4]> var_2147 = const()[name = tensor<string, []>("op_2147"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_219 = reshape(shape = var_2147, x = x_217)[name = tensor<string, []>("x_219")];
+            tensor<int32, [4]> var_2151_begin_0 = const()[name = tensor<string, []>("op_2151_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2151_end_0 = const()[name = tensor<string, []>("op_2151_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_2151_end_mask_0 = const()[name = tensor<string, []>("op_2151_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_2151 = slice_by_index(begin = var_2151_begin_0, end = var_2151_end_0, end_mask = var_2151_end_mask_0, x = x_219)[name = tensor<string, []>("op_2151")];
+            tensor<int32, [4]> var_2152 = const()[name = tensor<string, []>("op_2152"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_33 = reshape(shape = var_2152, x = var_2151)[name = tensor<string, []>("matrix_bd_33")];
+            tensor<bool, []> matrix_ac_17_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_17_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_17_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_112_perm_0 = const()[name = tensor<string, []>("transpose_112_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_113_perm_0 = const()[name = tensor<string, []>("transpose_113_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_113 = transpose(perm = transpose_113_perm_0, x = k_33)[name = tensor<string, []>("transpose_288")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_112 = transpose(perm = transpose_112_perm_0, x = var_2135)[name = tensor<string, []>("transpose_289")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_17 = matmul(transpose_x = matrix_ac_17_transpose_x_0, transpose_y = matrix_ac_17_transpose_y_0, x = transpose_112, y = transpose_113)[name = tensor<string, []>("matrix_ac_17")];
+            tensor<int32, [4]> matrix_bd_35_begin_0 = const()[name = tensor<string, []>("matrix_bd_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_35_end_0 = const()[name = tensor<string, []>("matrix_bd_35_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_35_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_35 = slice_by_index(begin = matrix_bd_35_begin_0, end = matrix_bd_35_end_0, end_mask = matrix_bd_35_end_mask_0, x = matrix_bd_33)[name = tensor<string, []>("matrix_bd_35")];
+            tensor<fp32, [1, 8, 2, 72]> var_2161 = add(x = matrix_ac_17, y = matrix_bd_35)[name = tensor<string, []>("op_2161")];
+            tensor<fp32, []> _inversed_scores_33_y_0 = const()[name = tensor<string, []>("_inversed_scores_33_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_33 = mul(x = var_2161, y = _inversed_scores_33_y_0)[name = tensor<string, []>("_inversed_scores_33")];
+            tensor<fp32, [1, 8, 2, 72]> scores_35 = select(a = var_41, b = _inversed_scores_33, cond = mask_11)[name = tensor<string, []>("scores_35")];
+            tensor<fp32, [1, 8, 2, 72]> var_2167 = softmax(axis = var_56, x = scores_35)[name = tensor<string, []>("op_2167")];
+            tensor<fp32, [1, 8, 2, 72]> input_457 = select(a = var_40, b = var_2167, cond = mask_11)[name = tensor<string, []>("input_457")];
+            tensor<bool, []> x_221_transpose_x_0 = const()[name = tensor<string, []>("x_221_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_221_transpose_y_0 = const()[name = tensor<string, []>("x_221_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_25 = transpose(perm = value_25_perm_0, x = v_17)[name = tensor<string, []>("transpose_287")];
+            tensor<fp32, [1, 8, 2, 128]> x_221 = matmul(transpose_x = x_221_transpose_x_0, transpose_y = x_221_transpose_y_0, x = input_457, y = value_25)[name = tensor<string, []>("x_221")];
+            tensor<int32, [4]> var_2171_perm_0 = const()[name = tensor<string, []>("op_2171_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2172 = const()[name = tensor<string, []>("op_2172"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_2171 = transpose(perm = var_2171_perm_0, x = x_221)[name = tensor<string, []>("transpose_286")];
+            tensor<fp32, [1, 2, 1024]> input_459 = reshape(shape = var_2172, x = var_2171)[name = tensor<string, []>("input_459")];
+            tensor<fp32, [1, 2, 1024]> input_461 = linear(bias = linear_2_bias_0, weight = module_layers_8_self_attn_linear_out_weight_quantized, x = input_459)[name = tensor<string, []>("linear_79")];
+            tensor<fp32, [1, 2, 1024]> input_463 = add(x = input_453, y = input_461)[name = tensor<string, []>("input_463")];
+            tensor<int32, [1]> x_225_axes_0 = const()[name = tensor<string, []>("x_225_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_225 = layer_norm(axes = x_225_axes_0, beta = module_layers_8_norm_conv_bias, epsilon = var_38, gamma = module_layers_8_norm_conv_weight, x = input_463)[name = tensor<string, []>("x_225")];
+            tensor<int32, [3]> input_465_perm_0 = const()[name = tensor<string, []>("input_465_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_467_pad_type_0 = const()[name = tensor<string, []>("input_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_467_strides_0 = const()[name = tensor<string, []>("input_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_467_pad_0 = const()[name = tensor<string, []>("input_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_467_dilations_0 = const()[name = tensor<string, []>("input_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_467_groups_0 = const()[name = tensor<string, []>("input_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_465 = transpose(perm = input_465_perm_0, x = x_225)[name = tensor<string, []>("transpose_285")];
+            tensor<fp32, [1, 2048, 2]> input_467 = conv(dilations = input_467_dilations_0, groups = input_467_groups_0, pad = input_467_pad_0, pad_type = input_467_pad_type_0, strides = input_467_strides_0, weight = module_layers_8_conv_pointwise_conv1_weight_quantized, x = input_465)[name = tensor<string, []>("input_467")];
+            tensor<int32, []> x_227_split_num_splits_0 = const()[name = tensor<string, []>("x_227_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_227_split_axis_0 = const()[name = tensor<string, []>("x_227_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_227_split_0, tensor<fp32, [1, 1024, 2]> x_227_split_1 = split(axis = x_227_split_axis_0, num_splits = x_227_split_num_splits_0, x = input_467)[name = tensor<string, []>("x_227_split")];
+            tensor<fp32, [1, 1024, 2]> x_227_split_1_sigmoid = sigmoid(x = x_227_split_1)[name = tensor<string, []>("x_227_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_227 = mul(x = x_227_split_0, y = x_227_split_1_sigmoid)[name = tensor<string, []>("x_227")];
+            tensor<fp32, [1, 1024, 2]> input_469 = select(a = var_40, b = x_227, cond = var_565)[name = tensor<string, []>("input_469")];
+            tensor<bool, []> new_x_35_interleave_0 = const()[name = tensor<string, []>("new_x_35_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_35 = concat(axis = var_56, interleave = new_x_35_interleave_0, values = (cache_35, input_469))[name = tensor<string, []>("new_x_35")];
+            tensor<int32, [3]> var_2210_begin_0 = const()[name = tensor<string, []>("op_2210_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_2210_end_0 = const()[name = tensor<string, []>("op_2210_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_2210_end_mask_0 = const()[name = tensor<string, []>("op_2210_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_2210 = slice_by_index(begin = var_2210_begin_0, end = var_2210_end_0, end_mask = var_2210_end_mask_0, x = new_x_35)[name = tensor<string, []>("op_2210")];
+            tensor<string, []> x_229_pad_type_0 = const()[name = tensor<string, []>("x_229_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_229_groups_0 = const()[name = tensor<string, []>("x_229_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_229_strides_0 = const()[name = tensor<string, []>("x_229_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_229_pad_0 = const()[name = tensor<string, []>("x_229_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_229_dilations_0 = const()[name = tensor<string, []>("x_229_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_229 = conv(dilations = x_229_dilations_0, groups = x_229_groups_0, pad = x_229_pad_0, pad_type = x_229_pad_type_0, strides = x_229_strides_0, weight = module_layers_8_conv_depthwise_conv_weight_quantized, x = new_x_35)[name = tensor<string, []>("x_229")];
+            tensor<int32, [3]> input_471_perm_0 = const()[name = tensor<string, []>("input_471_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_231_axes_0 = const()[name = tensor<string, []>("x_231_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_471 = transpose(perm = input_471_perm_0, x = x_229)[name = tensor<string, []>("transpose_284")];
+            tensor<fp32, [1, 2, 1024]> x_231 = layer_norm(axes = x_231_axes_0, beta = module_layers_8_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_8_conv_batch_norm_weight, x = input_471)[name = tensor<string, []>("x_231")];
+            tensor<int32, [3]> input_473_perm_0 = const()[name = tensor<string, []>("input_473_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_473 = transpose(perm = input_473_perm_0, x = x_231)[name = tensor<string, []>("transpose_283")];
+            tensor<fp32, [1, 1024, 2]> input_475 = silu(x = input_473)[name = tensor<string, []>("input_475")];
+            tensor<string, []> x_233_pad_type_0 = const()[name = tensor<string, []>("x_233_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_233_strides_0 = const()[name = tensor<string, []>("x_233_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_233_pad_0 = const()[name = tensor<string, []>("x_233_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_233_dilations_0 = const()[name = tensor<string, []>("x_233_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_233_groups_0 = const()[name = tensor<string, []>("x_233_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_233 = conv(dilations = x_233_dilations_0, groups = x_233_groups_0, pad = x_233_pad_0, pad_type = x_233_pad_type_0, strides = x_233_strides_0, weight = module_layers_8_conv_pointwise_conv2_weight_quantized, x = input_475)[name = tensor<string, []>("x_233")];
+            tensor<int32, [3]> input_477_perm_0 = const()[name = tensor<string, []>("input_477_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_477 = transpose(perm = input_477_perm_0, x = x_233)[name = tensor<string, []>("transpose_282")];
+            tensor<fp32, [1, 2, 1024]> input_479 = add(x = input_463, y = input_477)[name = tensor<string, []>("input_479")];
+            tensor<int32, [1]> input_481_axes_0 = const()[name = tensor<string, []>("input_481_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_481 = layer_norm(axes = input_481_axes_0, beta = module_layers_8_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_8_norm_feed_forward2_weight, x = input_479)[name = tensor<string, []>("input_481")];
+            tensor<fp32, [1, 2, 4096]> input_483 = linear(bias = linear_1_bias_0, weight = module_layers_8_feed_forward2_linear1_weight_quantized, x = input_481)[name = tensor<string, []>("linear_80")];
+            tensor<fp32, [1, 2, 4096]> input_485 = silu(x = input_483)[name = tensor<string, []>("input_485")];
+            tensor<fp32, [1, 2, 1024]> input_489 = linear(bias = linear_2_bias_0, weight = module_layers_8_feed_forward2_linear2_weight_quantized, x = input_485)[name = tensor<string, []>("linear_81")];
+            tensor<fp32, []> var_2251 = const()[name = tensor<string, []>("op_2251"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2252 = mul(x = input_489, y = var_2251)[name = tensor<string, []>("op_2252")];
+            tensor<fp32, [1, 2, 1024]> input_491 = add(x = input_479, y = var_2252)[name = tensor<string, []>("input_491")];
+            tensor<int32, [1]> input_493_axes_0 = const()[name = tensor<string, []>("input_493_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_493 = layer_norm(axes = input_493_axes_0, beta = module_layers_8_norm_out_bias, epsilon = var_38, gamma = module_layers_8_norm_out_weight, x = input_491)[name = tensor<string, []>("input_493")];
+            tensor<int32, [4]> cache_37_begin_0 = const()[name = tensor<string, []>("cache_37_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_37_end_0 = const()[name = tensor<string, []>("cache_37_end_0"), val = tensor<int32, [4]>([10, 1, 70, 1024])];
+            tensor<bool, [4]> cache_37_end_mask_0 = const()[name = tensor<string, []>("cache_37_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_37_squeeze_mask_0 = const()[name = tensor<string, []>("cache_37_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_37 = slice_by_index(begin = cache_37_begin_0, end = cache_37_end_0, end_mask = cache_37_end_mask_0, squeeze_mask = cache_37_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_37")];
+            tensor<int32, [4]> cache_39_begin_0 = const()[name = tensor<string, []>("cache_39_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
+            tensor<int32, [4]> cache_39_end_0 = const()[name = tensor<string, []>("cache_39_end_0"), val = tensor<int32, [4]>([10, 1, 1024, 8])];
+            tensor<bool, [4]> cache_39_end_mask_0 = const()[name = tensor<string, []>("cache_39_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_39_squeeze_mask_0 = const()[name = tensor<string, []>("cache_39_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_39 = slice_by_index(begin = cache_39_begin_0, end = cache_39_end_0, end_mask = cache_39_end_mask_0, squeeze_mask = cache_39_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_39")];
+            tensor<int32, [1]> input_495_axes_0 = const()[name = tensor<string, []>("input_495_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_495 = layer_norm(axes = input_495_axes_0, beta = module_layers_9_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_9_norm_feed_forward1_weight, x = input_493)[name = tensor<string, []>("input_495")];
+            tensor<fp32, [1, 2, 4096]> input_497 = linear(bias = linear_1_bias_0, weight = module_layers_9_feed_forward1_linear1_weight_quantized, x = input_495)[name = tensor<string, []>("linear_82")];
+            tensor<fp32, [1, 2, 4096]> input_499 = silu(x = input_497)[name = tensor<string, []>("input_499")];
+            tensor<fp32, [1, 2, 1024]> input_503 = linear(bias = linear_2_bias_0, weight = module_layers_9_feed_forward1_linear2_weight_quantized, x = input_499)[name = tensor<string, []>("linear_83")];
+            tensor<fp32, []> var_2286 = const()[name = tensor<string, []>("op_2286"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2287 = mul(x = input_503, y = var_2286)[name = tensor<string, []>("op_2287")];
+            tensor<fp32, [1, 2, 1024]> input_505 = add(x = input_493, y = var_2287)[name = tensor<string, []>("input_505")];
+            tensor<int32, [1]> key_19_axes_0 = const()[name = tensor<string, []>("key_19_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_19 = layer_norm(axes = key_19_axes_0, beta = module_layers_9_norm_self_att_bias, epsilon = var_38, gamma = module_layers_9_norm_self_att_weight, x = input_505)[name = tensor<string, []>("key_19")];
+            tensor<bool, []> input_507_interleave_0 = const()[name = tensor<string, []>("input_507_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_507 = concat(axis = var_65, interleave = input_507_interleave_0, values = (cache_37, key_19))[name = tensor<string, []>("input_507")];
+            tensor<int32, [3]> var_2309_begin_0 = const()[name = tensor<string, []>("op_2309_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_2309_end_0 = const()[name = tensor<string, []>("op_2309_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_2309_end_mask_0 = const()[name = tensor<string, []>("op_2309_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_2309 = slice_by_index(begin = var_2309_begin_0, end = var_2309_end_0, end_mask = var_2309_end_mask_0, x = cache_37)[name = tensor<string, []>("op_2309")];
+            tensor<bool, []> var_2315_interleave_0 = const()[name = tensor<string, []>("op_2315_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_2315 = concat(axis = var_65, interleave = var_2315_interleave_0, values = (var_2309, key_19))[name = tensor<string, []>("op_2315")];
+            tensor<fp32, [1, 2, 1024]> var_2318 = linear(bias = linear_2_bias_0, weight = module_layers_9_self_attn_linear_q_weight_quantized, x = key_19)[name = tensor<string, []>("linear_84")];
+            tensor<int32, [4]> var_2319 = const()[name = tensor<string, []>("op_2319"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_55 = reshape(shape = var_2319, x = var_2318)[name = tensor<string, []>("q_55")];
+            tensor<fp32, [1, 72, 1024]> var_2322 = linear(bias = linear_2_bias_0, weight = module_layers_9_self_attn_linear_k_weight_quantized, x = input_507)[name = tensor<string, []>("linear_85")];
+            tensor<int32, [4]> var_2323 = const()[name = tensor<string, []>("op_2323"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_37 = reshape(shape = var_2323, x = var_2322)[name = tensor<string, []>("k_37")];
+            tensor<fp32, [1, 72, 1024]> var_2326 = linear(bias = linear_2_bias_0, weight = module_layers_9_self_attn_linear_v_weight_quantized, x = input_507)[name = tensor<string, []>("linear_86")];
+            tensor<int32, [4]> var_2327 = const()[name = tensor<string, []>("op_2327"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_19 = reshape(shape = var_2327, x = var_2326)[name = tensor<string, []>("v_19")];
+            tensor<int32, [4]> value_27_perm_0 = const()[name = tensor<string, []>("value_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_2339 = add(x = q_55, y = module_layers_9_self_attn_pos_bias_u)[name = tensor<string, []>("op_2339")];
+            tensor<fp32, [1, 2, 8, 128]> var_2341 = add(x = q_55, y = module_layers_9_self_attn_pos_bias_v)[name = tensor<string, []>("op_2341")];
+            tensor<int32, [4]> q_with_bias_v_19_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_2343_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2343_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588659712))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588806464))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588806208)))];
+            tensor<bool, []> x_241_transpose_x_0 = const()[name = tensor<string, []>("x_241_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_241_transpose_y_0 = const()[name = tensor<string, []>("x_241_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_19 = transpose(perm = q_with_bias_v_19_perm_0, x = var_2341)[name = tensor<string, []>("transpose_281")];
+            tensor<fp32, [1, 8, 2, 143]> x_241 = matmul(transpose_x = x_241_transpose_x_0, transpose_y = x_241_transpose_y_0, x = q_with_bias_v_19, y = op_2343_quantized)[name = tensor<string, []>("x_241")];
+            tensor<fp32, []> const_196 = const()[name = tensor<string, []>("const_196"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_243_pad_0 = const()[name = tensor<string, []>("x_243_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_243_mode_0 = const()[name = tensor<string, []>("x_243_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_243 = pad(constant_val = const_196, mode = x_243_mode_0, pad = x_243_pad_0, x = x_241)[name = tensor<string, []>("x_243")];
+            tensor<int32, [4]> var_2351 = const()[name = tensor<string, []>("op_2351"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_245 = reshape(shape = var_2351, x = x_243)[name = tensor<string, []>("x_245")];
+            tensor<int32, [4]> var_2355_begin_0 = const()[name = tensor<string, []>("op_2355_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2355_end_0 = const()[name = tensor<string, []>("op_2355_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_2355_end_mask_0 = const()[name = tensor<string, []>("op_2355_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_2355 = slice_by_index(begin = var_2355_begin_0, end = var_2355_end_0, end_mask = var_2355_end_mask_0, x = x_245)[name = tensor<string, []>("op_2355")];
+            tensor<int32, [4]> var_2356 = const()[name = tensor<string, []>("op_2356"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_37 = reshape(shape = var_2356, x = var_2355)[name = tensor<string, []>("matrix_bd_37")];
+            tensor<bool, []> matrix_ac_19_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_19_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_19_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_114_perm_0 = const()[name = tensor<string, []>("transpose_114_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_115_perm_0 = const()[name = tensor<string, []>("transpose_115_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_115 = transpose(perm = transpose_115_perm_0, x = k_37)[name = tensor<string, []>("transpose_279")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_114 = transpose(perm = transpose_114_perm_0, x = var_2339)[name = tensor<string, []>("transpose_280")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_19 = matmul(transpose_x = matrix_ac_19_transpose_x_0, transpose_y = matrix_ac_19_transpose_y_0, x = transpose_114, y = transpose_115)[name = tensor<string, []>("matrix_ac_19")];
+            tensor<int32, [4]> matrix_bd_39_begin_0 = const()[name = tensor<string, []>("matrix_bd_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_39_end_0 = const()[name = tensor<string, []>("matrix_bd_39_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_39_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_39 = slice_by_index(begin = matrix_bd_39_begin_0, end = matrix_bd_39_end_0, end_mask = matrix_bd_39_end_mask_0, x = matrix_bd_37)[name = tensor<string, []>("matrix_bd_39")];
+            tensor<fp32, [1, 8, 2, 72]> var_2365 = add(x = matrix_ac_19, y = matrix_bd_39)[name = tensor<string, []>("op_2365")];
+            tensor<fp32, []> _inversed_scores_37_y_0 = const()[name = tensor<string, []>("_inversed_scores_37_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_37 = mul(x = var_2365, y = _inversed_scores_37_y_0)[name = tensor<string, []>("_inversed_scores_37")];
+            tensor<fp32, [1, 8, 2, 72]> scores_39 = select(a = var_41, b = _inversed_scores_37, cond = mask_11)[name = tensor<string, []>("scores_39")];
+            tensor<fp32, [1, 8, 2, 72]> var_2371 = softmax(axis = var_56, x = scores_39)[name = tensor<string, []>("op_2371")];
+            tensor<fp32, [1, 8, 2, 72]> input_509 = select(a = var_40, b = var_2371, cond = mask_11)[name = tensor<string, []>("input_509")];
+            tensor<bool, []> x_247_transpose_x_0 = const()[name = tensor<string, []>("x_247_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_247_transpose_y_0 = const()[name = tensor<string, []>("x_247_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_27 = transpose(perm = value_27_perm_0, x = v_19)[name = tensor<string, []>("transpose_278")];
+            tensor<fp32, [1, 8, 2, 128]> x_247 = matmul(transpose_x = x_247_transpose_x_0, transpose_y = x_247_transpose_y_0, x = input_509, y = value_27)[name = tensor<string, []>("x_247")];
+            tensor<int32, [4]> var_2375_perm_0 = const()[name = tensor<string, []>("op_2375_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2376 = const()[name = tensor<string, []>("op_2376"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_2375 = transpose(perm = var_2375_perm_0, x = x_247)[name = tensor<string, []>("transpose_277")];
+            tensor<fp32, [1, 2, 1024]> input_511 = reshape(shape = var_2376, x = var_2375)[name = tensor<string, []>("input_511")];
+            tensor<fp32, [1, 2, 1024]> input_513 = linear(bias = linear_2_bias_0, weight = module_layers_9_self_attn_linear_out_weight_quantized, x = input_511)[name = tensor<string, []>("linear_88")];
+            tensor<fp32, [1, 2, 1024]> input_515 = add(x = input_505, y = input_513)[name = tensor<string, []>("input_515")];
+            tensor<int32, [1]> x_251_axes_0 = const()[name = tensor<string, []>("x_251_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_251 = layer_norm(axes = x_251_axes_0, beta = module_layers_9_norm_conv_bias, epsilon = var_38, gamma = module_layers_9_norm_conv_weight, x = input_515)[name = tensor<string, []>("x_251")];
+            tensor<int32, [3]> input_517_perm_0 = const()[name = tensor<string, []>("input_517_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_519_pad_type_0 = const()[name = tensor<string, []>("input_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_519_strides_0 = const()[name = tensor<string, []>("input_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_519_pad_0 = const()[name = tensor<string, []>("input_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_519_dilations_0 = const()[name = tensor<string, []>("input_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_519_groups_0 = const()[name = tensor<string, []>("input_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_517 = transpose(perm = input_517_perm_0, x = x_251)[name = tensor<string, []>("transpose_276")];
+            tensor<fp32, [1, 2048, 2]> input_519 = conv(dilations = input_519_dilations_0, groups = input_519_groups_0, pad = input_519_pad_0, pad_type = input_519_pad_type_0, strides = input_519_strides_0, weight = module_layers_9_conv_pointwise_conv1_weight_quantized, x = input_517)[name = tensor<string, []>("input_519")];
+            tensor<int32, []> x_253_split_num_splits_0 = const()[name = tensor<string, []>("x_253_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_253_split_axis_0 = const()[name = tensor<string, []>("x_253_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_253_split_0, tensor<fp32, [1, 1024, 2]> x_253_split_1 = split(axis = x_253_split_axis_0, num_splits = x_253_split_num_splits_0, x = input_519)[name = tensor<string, []>("x_253_split")];
+            tensor<fp32, [1, 1024, 2]> x_253_split_1_sigmoid = sigmoid(x = x_253_split_1)[name = tensor<string, []>("x_253_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_253 = mul(x = x_253_split_0, y = x_253_split_1_sigmoid)[name = tensor<string, []>("x_253")];
+            tensor<fp32, [1, 1024, 2]> input_521 = select(a = var_40, b = x_253, cond = var_565)[name = tensor<string, []>("input_521")];
+            tensor<bool, []> new_x_39_interleave_0 = const()[name = tensor<string, []>("new_x_39_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_39 = concat(axis = var_56, interleave = new_x_39_interleave_0, values = (cache_39, input_521))[name = tensor<string, []>("new_x_39")];
+            tensor<int32, [3]> var_2414_begin_0 = const()[name = tensor<string, []>("op_2414_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_2414_end_0 = const()[name = tensor<string, []>("op_2414_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_2414_end_mask_0 = const()[name = tensor<string, []>("op_2414_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_2414 = slice_by_index(begin = var_2414_begin_0, end = var_2414_end_0, end_mask = var_2414_end_mask_0, x = new_x_39)[name = tensor<string, []>("op_2414")];
+            tensor<string, []> x_255_pad_type_0 = const()[name = tensor<string, []>("x_255_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_255_groups_0 = const()[name = tensor<string, []>("x_255_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_255_strides_0 = const()[name = tensor<string, []>("x_255_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_255_pad_0 = const()[name = tensor<string, []>("x_255_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_255_dilations_0 = const()[name = tensor<string, []>("x_255_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_255 = conv(dilations = x_255_dilations_0, groups = x_255_groups_0, pad = x_255_pad_0, pad_type = x_255_pad_type_0, strides = x_255_strides_0, weight = module_layers_9_conv_depthwise_conv_weight_quantized, x = new_x_39)[name = tensor<string, []>("x_255")];
+            tensor<int32, [3]> input_523_perm_0 = const()[name = tensor<string, []>("input_523_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_257_axes_0 = const()[name = tensor<string, []>("x_257_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_523 = transpose(perm = input_523_perm_0, x = x_255)[name = tensor<string, []>("transpose_275")];
+            tensor<fp32, [1, 2, 1024]> x_257 = layer_norm(axes = x_257_axes_0, beta = module_layers_9_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_9_conv_batch_norm_weight, x = input_523)[name = tensor<string, []>("x_257")];
+            tensor<int32, [3]> input_525_perm_0 = const()[name = tensor<string, []>("input_525_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_525 = transpose(perm = input_525_perm_0, x = x_257)[name = tensor<string, []>("transpose_274")];
+            tensor<fp32, [1, 1024, 2]> input_527 = silu(x = input_525)[name = tensor<string, []>("input_527")];
+            tensor<string, []> x_259_pad_type_0 = const()[name = tensor<string, []>("x_259_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_259_strides_0 = const()[name = tensor<string, []>("x_259_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_259_pad_0 = const()[name = tensor<string, []>("x_259_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_259_dilations_0 = const()[name = tensor<string, []>("x_259_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_259_groups_0 = const()[name = tensor<string, []>("x_259_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_259 = conv(dilations = x_259_dilations_0, groups = x_259_groups_0, pad = x_259_pad_0, pad_type = x_259_pad_type_0, strides = x_259_strides_0, weight = module_layers_9_conv_pointwise_conv2_weight_quantized, x = input_527)[name = tensor<string, []>("x_259")];
+            tensor<int32, [3]> input_529_perm_0 = const()[name = tensor<string, []>("input_529_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_529 = transpose(perm = input_529_perm_0, x = x_259)[name = tensor<string, []>("transpose_273")];
+            tensor<fp32, [1, 2, 1024]> input_531 = add(x = input_515, y = input_529)[name = tensor<string, []>("input_531")];
+            tensor<int32, [1]> input_533_axes_0 = const()[name = tensor<string, []>("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_533 = layer_norm(axes = input_533_axes_0, beta = module_layers_9_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_9_norm_feed_forward2_weight, x = input_531)[name = tensor<string, []>("input_533")];
+            tensor<fp32, [1, 2, 4096]> input_535 = linear(bias = linear_1_bias_0, weight = module_layers_9_feed_forward2_linear1_weight_quantized, x = input_533)[name = tensor<string, []>("linear_89")];
+            tensor<fp32, [1, 2, 4096]> input_537 = silu(x = input_535)[name = tensor<string, []>("input_537")];
+            tensor<fp32, [1, 2, 1024]> input_541 = linear(bias = linear_2_bias_0, weight = module_layers_9_feed_forward2_linear2_weight_quantized, x = input_537)[name = tensor<string, []>("linear_90")];
+            tensor<fp32, []> var_2455 = const()[name = tensor<string, []>("op_2455"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2456 = mul(x = input_541, y = var_2455)[name = tensor<string, []>("op_2456")];
+            tensor<fp32, [1, 2, 1024]> input_543 = add(x = input_531, y = var_2456)[name = tensor<string, []>("input_543")];
+            tensor<int32, [1]> input_545_axes_0 = const()[name = tensor<string, []>("input_545_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_545 = layer_norm(axes = input_545_axes_0, beta = module_layers_9_norm_out_bias, epsilon = var_38, gamma = module_layers_9_norm_out_weight, x = input_543)[name = tensor<string, []>("input_545")];
+            tensor<int32, [4]> cache_41_begin_0 = const()[name = tensor<string, []>("cache_41_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_41_end_0 = const()[name = tensor<string, []>("cache_41_end_0"), val = tensor<int32, [4]>([11, 1, 70, 1024])];
+            tensor<bool, [4]> cache_41_end_mask_0 = const()[name = tensor<string, []>("cache_41_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_41_squeeze_mask_0 = const()[name = tensor<string, []>("cache_41_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_41 = slice_by_index(begin = cache_41_begin_0, end = cache_41_end_0, end_mask = cache_41_end_mask_0, squeeze_mask = cache_41_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_41")];
+            tensor<int32, [4]> cache_43_begin_0 = const()[name = tensor<string, []>("cache_43_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
+            tensor<int32, [4]> cache_43_end_0 = const()[name = tensor<string, []>("cache_43_end_0"), val = tensor<int32, [4]>([11, 1, 1024, 8])];
+            tensor<bool, [4]> cache_43_end_mask_0 = const()[name = tensor<string, []>("cache_43_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_43_squeeze_mask_0 = const()[name = tensor<string, []>("cache_43_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_43 = slice_by_index(begin = cache_43_begin_0, end = cache_43_end_0, end_mask = cache_43_end_mask_0, squeeze_mask = cache_43_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_43")];
+            tensor<int32, [1]> input_547_axes_0 = const()[name = tensor<string, []>("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_547 = layer_norm(axes = input_547_axes_0, beta = module_layers_10_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_10_norm_feed_forward1_weight, x = input_545)[name = tensor<string, []>("input_547")];
+            tensor<fp32, [1, 2, 4096]> input_549 = linear(bias = linear_1_bias_0, weight = module_layers_10_feed_forward1_linear1_weight_quantized, x = input_547)[name = tensor<string, []>("linear_91")];
+            tensor<fp32, [1, 2, 4096]> input_551 = silu(x = input_549)[name = tensor<string, []>("input_551")];
+            tensor<fp32, [1, 2, 1024]> input_555 = linear(bias = linear_2_bias_0, weight = module_layers_10_feed_forward1_linear2_weight_quantized, x = input_551)[name = tensor<string, []>("linear_92")];
+            tensor<fp32, []> var_2490 = const()[name = tensor<string, []>("op_2490"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2491 = mul(x = input_555, y = var_2490)[name = tensor<string, []>("op_2491")];
+            tensor<fp32, [1, 2, 1024]> input_557 = add(x = input_545, y = var_2491)[name = tensor<string, []>("input_557")];
+            tensor<int32, [1]> key_21_axes_0 = const()[name = tensor<string, []>("key_21_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_21 = layer_norm(axes = key_21_axes_0, beta = module_layers_10_norm_self_att_bias, epsilon = var_38, gamma = module_layers_10_norm_self_att_weight, x = input_557)[name = tensor<string, []>("key_21")];
+            tensor<bool, []> input_559_interleave_0 = const()[name = tensor<string, []>("input_559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_559 = concat(axis = var_65, interleave = input_559_interleave_0, values = (cache_41, key_21))[name = tensor<string, []>("input_559")];
+            tensor<int32, [3]> var_2513_begin_0 = const()[name = tensor<string, []>("op_2513_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_2513_end_0 = const()[name = tensor<string, []>("op_2513_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_2513_end_mask_0 = const()[name = tensor<string, []>("op_2513_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_2513 = slice_by_index(begin = var_2513_begin_0, end = var_2513_end_0, end_mask = var_2513_end_mask_0, x = cache_41)[name = tensor<string, []>("op_2513")];
+            tensor<bool, []> var_2519_interleave_0 = const()[name = tensor<string, []>("op_2519_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_2519 = concat(axis = var_65, interleave = var_2519_interleave_0, values = (var_2513, key_21))[name = tensor<string, []>("op_2519")];
+            tensor<fp32, [1, 2, 1024]> var_2522 = linear(bias = linear_2_bias_0, weight = module_layers_10_self_attn_linear_q_weight_quantized, x = key_21)[name = tensor<string, []>("linear_93")];
+            tensor<int32, [4]> var_2523 = const()[name = tensor<string, []>("op_2523"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_61 = reshape(shape = var_2523, x = var_2522)[name = tensor<string, []>("q_61")];
+            tensor<fp32, [1, 72, 1024]> var_2526 = linear(bias = linear_2_bias_0, weight = module_layers_10_self_attn_linear_k_weight_quantized, x = input_559)[name = tensor<string, []>("linear_94")];
+            tensor<int32, [4]> var_2527 = const()[name = tensor<string, []>("op_2527"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_41 = reshape(shape = var_2527, x = var_2526)[name = tensor<string, []>("k_41")];
+            tensor<fp32, [1, 72, 1024]> var_2530 = linear(bias = linear_2_bias_0, weight = module_layers_10_self_attn_linear_v_weight_quantized, x = input_559)[name = tensor<string, []>("linear_95")];
+            tensor<int32, [4]> var_2531 = const()[name = tensor<string, []>("op_2531"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_21 = reshape(shape = var_2531, x = var_2530)[name = tensor<string, []>("v_21")];
+            tensor<int32, [4]> value_29_perm_0 = const()[name = tensor<string, []>("value_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_2543 = add(x = q_61, y = module_layers_10_self_attn_pos_bias_u)[name = tensor<string, []>("op_2543")];
+            tensor<fp32, [1, 2, 8, 128]> var_2545 = add(x = q_61, y = module_layers_10_self_attn_pos_bias_v)[name = tensor<string, []>("op_2545")];
+            tensor<int32, [4]> q_with_bias_v_21_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_2547_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2547_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588807104))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588953856))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588953600)))];
+            tensor<bool, []> x_267_transpose_x_0 = const()[name = tensor<string, []>("x_267_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_267_transpose_y_0 = const()[name = tensor<string, []>("x_267_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_21 = transpose(perm = q_with_bias_v_21_perm_0, x = var_2545)[name = tensor<string, []>("transpose_272")];
+            tensor<fp32, [1, 8, 2, 143]> x_267 = matmul(transpose_x = x_267_transpose_x_0, transpose_y = x_267_transpose_y_0, x = q_with_bias_v_21, y = op_2547_quantized)[name = tensor<string, []>("x_267")];
+            tensor<fp32, []> const_209 = const()[name = tensor<string, []>("const_209"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_269_pad_0 = const()[name = tensor<string, []>("x_269_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_269_mode_0 = const()[name = tensor<string, []>("x_269_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_269 = pad(constant_val = const_209, mode = x_269_mode_0, pad = x_269_pad_0, x = x_267)[name = tensor<string, []>("x_269")];
+            tensor<int32, [4]> var_2555 = const()[name = tensor<string, []>("op_2555"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_271 = reshape(shape = var_2555, x = x_269)[name = tensor<string, []>("x_271")];
+            tensor<int32, [4]> var_2559_begin_0 = const()[name = tensor<string, []>("op_2559_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2559_end_0 = const()[name = tensor<string, []>("op_2559_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_2559_end_mask_0 = const()[name = tensor<string, []>("op_2559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_2559 = slice_by_index(begin = var_2559_begin_0, end = var_2559_end_0, end_mask = var_2559_end_mask_0, x = x_271)[name = tensor<string, []>("op_2559")];
+            tensor<int32, [4]> var_2560 = const()[name = tensor<string, []>("op_2560"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_41 = reshape(shape = var_2560, x = var_2559)[name = tensor<string, []>("matrix_bd_41")];
+            tensor<bool, []> matrix_ac_21_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_21_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_21_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_116_perm_0 = const()[name = tensor<string, []>("transpose_116_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_117_perm_0 = const()[name = tensor<string, []>("transpose_117_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_117 = transpose(perm = transpose_117_perm_0, x = k_41)[name = tensor<string, []>("transpose_270")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_116 = transpose(perm = transpose_116_perm_0, x = var_2543)[name = tensor<string, []>("transpose_271")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_21 = matmul(transpose_x = matrix_ac_21_transpose_x_0, transpose_y = matrix_ac_21_transpose_y_0, x = transpose_116, y = transpose_117)[name = tensor<string, []>("matrix_ac_21")];
+            tensor<int32, [4]> matrix_bd_43_begin_0 = const()[name = tensor<string, []>("matrix_bd_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_43_end_0 = const()[name = tensor<string, []>("matrix_bd_43_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_43_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_43 = slice_by_index(begin = matrix_bd_43_begin_0, end = matrix_bd_43_end_0, end_mask = matrix_bd_43_end_mask_0, x = matrix_bd_41)[name = tensor<string, []>("matrix_bd_43")];
+            tensor<fp32, [1, 8, 2, 72]> var_2569 = add(x = matrix_ac_21, y = matrix_bd_43)[name = tensor<string, []>("op_2569")];
+            tensor<fp32, []> _inversed_scores_41_y_0 = const()[name = tensor<string, []>("_inversed_scores_41_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_41 = mul(x = var_2569, y = _inversed_scores_41_y_0)[name = tensor<string, []>("_inversed_scores_41")];
+            tensor<fp32, [1, 8, 2, 72]> scores_43 = select(a = var_41, b = _inversed_scores_41, cond = mask_11)[name = tensor<string, []>("scores_43")];
+            tensor<fp32, [1, 8, 2, 72]> var_2575 = softmax(axis = var_56, x = scores_43)[name = tensor<string, []>("op_2575")];
+            tensor<fp32, [1, 8, 2, 72]> input_561 = select(a = var_40, b = var_2575, cond = mask_11)[name = tensor<string, []>("input_561")];
+            tensor<bool, []> x_273_transpose_x_0 = const()[name = tensor<string, []>("x_273_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_273_transpose_y_0 = const()[name = tensor<string, []>("x_273_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_29 = transpose(perm = value_29_perm_0, x = v_21)[name = tensor<string, []>("transpose_269")];
+            tensor<fp32, [1, 8, 2, 128]> x_273 = matmul(transpose_x = x_273_transpose_x_0, transpose_y = x_273_transpose_y_0, x = input_561, y = value_29)[name = tensor<string, []>("x_273")];
+            tensor<int32, [4]> var_2579_perm_0 = const()[name = tensor<string, []>("op_2579_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2580 = const()[name = tensor<string, []>("op_2580"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_2579 = transpose(perm = var_2579_perm_0, x = x_273)[name = tensor<string, []>("transpose_268")];
+            tensor<fp32, [1, 2, 1024]> input_563 = reshape(shape = var_2580, x = var_2579)[name = tensor<string, []>("input_563")];
+            tensor<fp32, [1, 2, 1024]> input_565 = linear(bias = linear_2_bias_0, weight = module_layers_10_self_attn_linear_out_weight_quantized, x = input_563)[name = tensor<string, []>("linear_97")];
+            tensor<fp32, [1, 2, 1024]> input_567 = add(x = input_557, y = input_565)[name = tensor<string, []>("input_567")];
+            tensor<int32, [1]> x_277_axes_0 = const()[name = tensor<string, []>("x_277_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_277 = layer_norm(axes = x_277_axes_0, beta = module_layers_10_norm_conv_bias, epsilon = var_38, gamma = module_layers_10_norm_conv_weight, x = input_567)[name = tensor<string, []>("x_277")];
+            tensor<int32, [3]> input_569_perm_0 = const()[name = tensor<string, []>("input_569_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_571_pad_type_0 = const()[name = tensor<string, []>("input_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_571_strides_0 = const()[name = tensor<string, []>("input_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_571_pad_0 = const()[name = tensor<string, []>("input_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_571_dilations_0 = const()[name = tensor<string, []>("input_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_571_groups_0 = const()[name = tensor<string, []>("input_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_569 = transpose(perm = input_569_perm_0, x = x_277)[name = tensor<string, []>("transpose_267")];
+            tensor<fp32, [1, 2048, 2]> input_571 = conv(dilations = input_571_dilations_0, groups = input_571_groups_0, pad = input_571_pad_0, pad_type = input_571_pad_type_0, strides = input_571_strides_0, weight = module_layers_10_conv_pointwise_conv1_weight_quantized, x = input_569)[name = tensor<string, []>("input_571")];
+            tensor<int32, []> x_279_split_num_splits_0 = const()[name = tensor<string, []>("x_279_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_279_split_axis_0 = const()[name = tensor<string, []>("x_279_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_279_split_0, tensor<fp32, [1, 1024, 2]> x_279_split_1 = split(axis = x_279_split_axis_0, num_splits = x_279_split_num_splits_0, x = input_571)[name = tensor<string, []>("x_279_split")];
+            tensor<fp32, [1, 1024, 2]> x_279_split_1_sigmoid = sigmoid(x = x_279_split_1)[name = tensor<string, []>("x_279_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_279 = mul(x = x_279_split_0, y = x_279_split_1_sigmoid)[name = tensor<string, []>("x_279")];
+            tensor<fp32, [1, 1024, 2]> input_573 = select(a = var_40, b = x_279, cond = var_565)[name = tensor<string, []>("input_573")];
+            tensor<bool, []> new_x_43_interleave_0 = const()[name = tensor<string, []>("new_x_43_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_43 = concat(axis = var_56, interleave = new_x_43_interleave_0, values = (cache_43, input_573))[name = tensor<string, []>("new_x_43")];
+            tensor<int32, [3]> var_2618_begin_0 = const()[name = tensor<string, []>("op_2618_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_2618_end_0 = const()[name = tensor<string, []>("op_2618_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_2618_end_mask_0 = const()[name = tensor<string, []>("op_2618_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_2618 = slice_by_index(begin = var_2618_begin_0, end = var_2618_end_0, end_mask = var_2618_end_mask_0, x = new_x_43)[name = tensor<string, []>("op_2618")];
+            tensor<string, []> x_281_pad_type_0 = const()[name = tensor<string, []>("x_281_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_281_groups_0 = const()[name = tensor<string, []>("x_281_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_281_strides_0 = const()[name = tensor<string, []>("x_281_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_281_pad_0 = const()[name = tensor<string, []>("x_281_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_281_dilations_0 = const()[name = tensor<string, []>("x_281_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_281 = conv(dilations = x_281_dilations_0, groups = x_281_groups_0, pad = x_281_pad_0, pad_type = x_281_pad_type_0, strides = x_281_strides_0, weight = module_layers_10_conv_depthwise_conv_weight_quantized, x = new_x_43)[name = tensor<string, []>("x_281")];
+            tensor<int32, [3]> input_575_perm_0 = const()[name = tensor<string, []>("input_575_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_283_axes_0 = const()[name = tensor<string, []>("x_283_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_575 = transpose(perm = input_575_perm_0, x = x_281)[name = tensor<string, []>("transpose_266")];
+            tensor<fp32, [1, 2, 1024]> x_283 = layer_norm(axes = x_283_axes_0, beta = module_layers_10_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_10_conv_batch_norm_weight, x = input_575)[name = tensor<string, []>("x_283")];
+            tensor<int32, [3]> input_577_perm_0 = const()[name = tensor<string, []>("input_577_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_577 = transpose(perm = input_577_perm_0, x = x_283)[name = tensor<string, []>("transpose_265")];
+            tensor<fp32, [1, 1024, 2]> input_579 = silu(x = input_577)[name = tensor<string, []>("input_579")];
+            tensor<string, []> x_285_pad_type_0 = const()[name = tensor<string, []>("x_285_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_285_strides_0 = const()[name = tensor<string, []>("x_285_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_285_pad_0 = const()[name = tensor<string, []>("x_285_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_285_dilations_0 = const()[name = tensor<string, []>("x_285_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_285_groups_0 = const()[name = tensor<string, []>("x_285_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_285 = conv(dilations = x_285_dilations_0, groups = x_285_groups_0, pad = x_285_pad_0, pad_type = x_285_pad_type_0, strides = x_285_strides_0, weight = module_layers_10_conv_pointwise_conv2_weight_quantized, x = input_579)[name = tensor<string, []>("x_285")];
+            tensor<int32, [3]> input_581_perm_0 = const()[name = tensor<string, []>("input_581_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_581 = transpose(perm = input_581_perm_0, x = x_285)[name = tensor<string, []>("transpose_264")];
+            tensor<fp32, [1, 2, 1024]> input_583 = add(x = input_567, y = input_581)[name = tensor<string, []>("input_583")];
+            tensor<int32, [1]> input_585_axes_0 = const()[name = tensor<string, []>("input_585_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_585 = layer_norm(axes = input_585_axes_0, beta = module_layers_10_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_10_norm_feed_forward2_weight, x = input_583)[name = tensor<string, []>("input_585")];
+            tensor<fp32, [1, 2, 4096]> input_587 = linear(bias = linear_1_bias_0, weight = module_layers_10_feed_forward2_linear1_weight_quantized, x = input_585)[name = tensor<string, []>("linear_98")];
+            tensor<fp32, [1, 2, 4096]> input_589 = silu(x = input_587)[name = tensor<string, []>("input_589")];
+            tensor<fp32, [1, 2, 1024]> input_593 = linear(bias = linear_2_bias_0, weight = module_layers_10_feed_forward2_linear2_weight_quantized, x = input_589)[name = tensor<string, []>("linear_99")];
+            tensor<fp32, []> var_2659 = const()[name = tensor<string, []>("op_2659"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2660 = mul(x = input_593, y = var_2659)[name = tensor<string, []>("op_2660")];
+            tensor<fp32, [1, 2, 1024]> input_595 = add(x = input_583, y = var_2660)[name = tensor<string, []>("input_595")];
+            tensor<int32, [1]> input_597_axes_0 = const()[name = tensor<string, []>("input_597_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_597 = layer_norm(axes = input_597_axes_0, beta = module_layers_10_norm_out_bias, epsilon = var_38, gamma = module_layers_10_norm_out_weight, x = input_595)[name = tensor<string, []>("input_597")];
+            tensor<int32, [4]> cache_45_begin_0 = const()[name = tensor<string, []>("cache_45_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_45_end_0 = const()[name = tensor<string, []>("cache_45_end_0"), val = tensor<int32, [4]>([12, 1, 70, 1024])];
+            tensor<bool, [4]> cache_45_end_mask_0 = const()[name = tensor<string, []>("cache_45_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_45_squeeze_mask_0 = const()[name = tensor<string, []>("cache_45_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_45 = slice_by_index(begin = cache_45_begin_0, end = cache_45_end_0, end_mask = cache_45_end_mask_0, squeeze_mask = cache_45_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_45")];
+            tensor<int32, [4]> cache_47_begin_0 = const()[name = tensor<string, []>("cache_47_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
+            tensor<int32, [4]> cache_47_end_0 = const()[name = tensor<string, []>("cache_47_end_0"), val = tensor<int32, [4]>([12, 1, 1024, 8])];
+            tensor<bool, [4]> cache_47_end_mask_0 = const()[name = tensor<string, []>("cache_47_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_47_squeeze_mask_0 = const()[name = tensor<string, []>("cache_47_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_47 = slice_by_index(begin = cache_47_begin_0, end = cache_47_end_0, end_mask = cache_47_end_mask_0, squeeze_mask = cache_47_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_47")];
+            tensor<int32, [1]> input_599_axes_0 = const()[name = tensor<string, []>("input_599_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_599 = layer_norm(axes = input_599_axes_0, beta = module_layers_11_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_11_norm_feed_forward1_weight, x = input_597)[name = tensor<string, []>("input_599")];
+            tensor<fp32, [1, 2, 4096]> input_601 = linear(bias = linear_1_bias_0, weight = module_layers_11_feed_forward1_linear1_weight_quantized, x = input_599)[name = tensor<string, []>("linear_100")];
+            tensor<fp32, [1, 2, 4096]> input_603 = silu(x = input_601)[name = tensor<string, []>("input_603")];
+            tensor<fp32, [1, 2, 1024]> input_607 = linear(bias = linear_2_bias_0, weight = module_layers_11_feed_forward1_linear2_weight_quantized, x = input_603)[name = tensor<string, []>("linear_101")];
+            tensor<fp32, []> var_2694 = const()[name = tensor<string, []>("op_2694"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2695 = mul(x = input_607, y = var_2694)[name = tensor<string, []>("op_2695")];
+            tensor<fp32, [1, 2, 1024]> input_609 = add(x = input_597, y = var_2695)[name = tensor<string, []>("input_609")];
+            tensor<int32, [1]> key_23_axes_0 = const()[name = tensor<string, []>("key_23_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_23 = layer_norm(axes = key_23_axes_0, beta = module_layers_11_norm_self_att_bias, epsilon = var_38, gamma = module_layers_11_norm_self_att_weight, x = input_609)[name = tensor<string, []>("key_23")];
+            tensor<bool, []> input_611_interleave_0 = const()[name = tensor<string, []>("input_611_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_611 = concat(axis = var_65, interleave = input_611_interleave_0, values = (cache_45, key_23))[name = tensor<string, []>("input_611")];
+            tensor<int32, [3]> var_2717_begin_0 = const()[name = tensor<string, []>("op_2717_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_2717_end_0 = const()[name = tensor<string, []>("op_2717_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_2717_end_mask_0 = const()[name = tensor<string, []>("op_2717_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_2717 = slice_by_index(begin = var_2717_begin_0, end = var_2717_end_0, end_mask = var_2717_end_mask_0, x = cache_45)[name = tensor<string, []>("op_2717")];
+            tensor<bool, []> var_2723_interleave_0 = const()[name = tensor<string, []>("op_2723_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_2723 = concat(axis = var_65, interleave = var_2723_interleave_0, values = (var_2717, key_23))[name = tensor<string, []>("op_2723")];
+            tensor<fp32, [1, 2, 1024]> var_2726 = linear(bias = linear_2_bias_0, weight = module_layers_11_self_attn_linear_q_weight_quantized, x = key_23)[name = tensor<string, []>("linear_102")];
+            tensor<int32, [4]> var_2727 = const()[name = tensor<string, []>("op_2727"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_67 = reshape(shape = var_2727, x = var_2726)[name = tensor<string, []>("q_67")];
+            tensor<fp32, [1, 72, 1024]> var_2730 = linear(bias = linear_2_bias_0, weight = module_layers_11_self_attn_linear_k_weight_quantized, x = input_611)[name = tensor<string, []>("linear_103")];
+            tensor<int32, [4]> var_2731 = const()[name = tensor<string, []>("op_2731"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_45 = reshape(shape = var_2731, x = var_2730)[name = tensor<string, []>("k_45")];
+            tensor<fp32, [1, 72, 1024]> var_2734 = linear(bias = linear_2_bias_0, weight = module_layers_11_self_attn_linear_v_weight_quantized, x = input_611)[name = tensor<string, []>("linear_104")];
+            tensor<int32, [4]> var_2735 = const()[name = tensor<string, []>("op_2735"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_23 = reshape(shape = var_2735, x = var_2734)[name = tensor<string, []>("v_23")];
+            tensor<int32, [4]> value_31_perm_0 = const()[name = tensor<string, []>("value_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_2747 = add(x = q_67, y = module_layers_11_self_attn_pos_bias_u)[name = tensor<string, []>("op_2747")];
+            tensor<fp32, [1, 2, 8, 128]> var_2749 = add(x = q_67, y = module_layers_11_self_attn_pos_bias_v)[name = tensor<string, []>("op_2749")];
+            tensor<int32, [4]> q_with_bias_v_23_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_2751_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2751_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(588954496))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589101248))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589100992)))];
+            tensor<bool, []> x_293_transpose_x_0 = const()[name = tensor<string, []>("x_293_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_293_transpose_y_0 = const()[name = tensor<string, []>("x_293_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_23 = transpose(perm = q_with_bias_v_23_perm_0, x = var_2749)[name = tensor<string, []>("transpose_263")];
+            tensor<fp32, [1, 8, 2, 143]> x_293 = matmul(transpose_x = x_293_transpose_x_0, transpose_y = x_293_transpose_y_0, x = q_with_bias_v_23, y = op_2751_quantized)[name = tensor<string, []>("x_293")];
+            tensor<fp32, []> const_222 = const()[name = tensor<string, []>("const_222"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_295_pad_0 = const()[name = tensor<string, []>("x_295_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_295_mode_0 = const()[name = tensor<string, []>("x_295_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_295 = pad(constant_val = const_222, mode = x_295_mode_0, pad = x_295_pad_0, x = x_293)[name = tensor<string, []>("x_295")];
+            tensor<int32, [4]> var_2759 = const()[name = tensor<string, []>("op_2759"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_297 = reshape(shape = var_2759, x = x_295)[name = tensor<string, []>("x_297")];
+            tensor<int32, [4]> var_2763_begin_0 = const()[name = tensor<string, []>("op_2763_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2763_end_0 = const()[name = tensor<string, []>("op_2763_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_2763_end_mask_0 = const()[name = tensor<string, []>("op_2763_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_2763 = slice_by_index(begin = var_2763_begin_0, end = var_2763_end_0, end_mask = var_2763_end_mask_0, x = x_297)[name = tensor<string, []>("op_2763")];
+            tensor<int32, [4]> var_2764 = const()[name = tensor<string, []>("op_2764"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_45 = reshape(shape = var_2764, x = var_2763)[name = tensor<string, []>("matrix_bd_45")];
+            tensor<bool, []> matrix_ac_23_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_23_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_23_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_118_perm_0 = const()[name = tensor<string, []>("transpose_118_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_119_perm_0 = const()[name = tensor<string, []>("transpose_119_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_119 = transpose(perm = transpose_119_perm_0, x = k_45)[name = tensor<string, []>("transpose_261")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_118 = transpose(perm = transpose_118_perm_0, x = var_2747)[name = tensor<string, []>("transpose_262")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_23 = matmul(transpose_x = matrix_ac_23_transpose_x_0, transpose_y = matrix_ac_23_transpose_y_0, x = transpose_118, y = transpose_119)[name = tensor<string, []>("matrix_ac_23")];
+            tensor<int32, [4]> matrix_bd_47_begin_0 = const()[name = tensor<string, []>("matrix_bd_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_47_end_0 = const()[name = tensor<string, []>("matrix_bd_47_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_47_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_47 = slice_by_index(begin = matrix_bd_47_begin_0, end = matrix_bd_47_end_0, end_mask = matrix_bd_47_end_mask_0, x = matrix_bd_45)[name = tensor<string, []>("matrix_bd_47")];
+            tensor<fp32, [1, 8, 2, 72]> var_2773 = add(x = matrix_ac_23, y = matrix_bd_47)[name = tensor<string, []>("op_2773")];
+            tensor<fp32, []> _inversed_scores_45_y_0 = const()[name = tensor<string, []>("_inversed_scores_45_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_45 = mul(x = var_2773, y = _inversed_scores_45_y_0)[name = tensor<string, []>("_inversed_scores_45")];
+            tensor<fp32, [1, 8, 2, 72]> scores_47 = select(a = var_41, b = _inversed_scores_45, cond = mask_11)[name = tensor<string, []>("scores_47")];
+            tensor<fp32, [1, 8, 2, 72]> var_2779 = softmax(axis = var_56, x = scores_47)[name = tensor<string, []>("op_2779")];
+            tensor<fp32, [1, 8, 2, 72]> input_613 = select(a = var_40, b = var_2779, cond = mask_11)[name = tensor<string, []>("input_613")];
+            tensor<bool, []> x_299_transpose_x_0 = const()[name = tensor<string, []>("x_299_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_299_transpose_y_0 = const()[name = tensor<string, []>("x_299_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_31 = transpose(perm = value_31_perm_0, x = v_23)[name = tensor<string, []>("transpose_260")];
+            tensor<fp32, [1, 8, 2, 128]> x_299 = matmul(transpose_x = x_299_transpose_x_0, transpose_y = x_299_transpose_y_0, x = input_613, y = value_31)[name = tensor<string, []>("x_299")];
+            tensor<int32, [4]> var_2783_perm_0 = const()[name = tensor<string, []>("op_2783_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2784 = const()[name = tensor<string, []>("op_2784"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_2783 = transpose(perm = var_2783_perm_0, x = x_299)[name = tensor<string, []>("transpose_259")];
+            tensor<fp32, [1, 2, 1024]> input_615 = reshape(shape = var_2784, x = var_2783)[name = tensor<string, []>("input_615")];
+            tensor<fp32, [1, 2, 1024]> input_617 = linear(bias = linear_2_bias_0, weight = module_layers_11_self_attn_linear_out_weight_quantized, x = input_615)[name = tensor<string, []>("linear_106")];
+            tensor<fp32, [1, 2, 1024]> input_619 = add(x = input_609, y = input_617)[name = tensor<string, []>("input_619")];
+            tensor<int32, [1]> x_303_axes_0 = const()[name = tensor<string, []>("x_303_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_303 = layer_norm(axes = x_303_axes_0, beta = module_layers_11_norm_conv_bias, epsilon = var_38, gamma = module_layers_11_norm_conv_weight, x = input_619)[name = tensor<string, []>("x_303")];
+            tensor<int32, [3]> input_621_perm_0 = const()[name = tensor<string, []>("input_621_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_623_pad_type_0 = const()[name = tensor<string, []>("input_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_623_strides_0 = const()[name = tensor<string, []>("input_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_623_pad_0 = const()[name = tensor<string, []>("input_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_623_dilations_0 = const()[name = tensor<string, []>("input_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_623_groups_0 = const()[name = tensor<string, []>("input_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_621 = transpose(perm = input_621_perm_0, x = x_303)[name = tensor<string, []>("transpose_258")];
+            tensor<fp32, [1, 2048, 2]> input_623 = conv(dilations = input_623_dilations_0, groups = input_623_groups_0, pad = input_623_pad_0, pad_type = input_623_pad_type_0, strides = input_623_strides_0, weight = module_layers_11_conv_pointwise_conv1_weight_quantized, x = input_621)[name = tensor<string, []>("input_623")];
+            tensor<int32, []> x_305_split_num_splits_0 = const()[name = tensor<string, []>("x_305_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_305_split_axis_0 = const()[name = tensor<string, []>("x_305_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_305_split_0, tensor<fp32, [1, 1024, 2]> x_305_split_1 = split(axis = x_305_split_axis_0, num_splits = x_305_split_num_splits_0, x = input_623)[name = tensor<string, []>("x_305_split")];
+            tensor<fp32, [1, 1024, 2]> x_305_split_1_sigmoid = sigmoid(x = x_305_split_1)[name = tensor<string, []>("x_305_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_305 = mul(x = x_305_split_0, y = x_305_split_1_sigmoid)[name = tensor<string, []>("x_305")];
+            tensor<fp32, [1, 1024, 2]> input_625 = select(a = var_40, b = x_305, cond = var_565)[name = tensor<string, []>("input_625")];
+            tensor<bool, []> new_x_47_interleave_0 = const()[name = tensor<string, []>("new_x_47_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_47 = concat(axis = var_56, interleave = new_x_47_interleave_0, values = (cache_47, input_625))[name = tensor<string, []>("new_x_47")];
+            tensor<int32, [3]> var_2822_begin_0 = const()[name = tensor<string, []>("op_2822_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_2822_end_0 = const()[name = tensor<string, []>("op_2822_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_2822_end_mask_0 = const()[name = tensor<string, []>("op_2822_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_2822 = slice_by_index(begin = var_2822_begin_0, end = var_2822_end_0, end_mask = var_2822_end_mask_0, x = new_x_47)[name = tensor<string, []>("op_2822")];
+            tensor<string, []> x_307_pad_type_0 = const()[name = tensor<string, []>("x_307_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_307_groups_0 = const()[name = tensor<string, []>("x_307_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_307_strides_0 = const()[name = tensor<string, []>("x_307_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_307_pad_0 = const()[name = tensor<string, []>("x_307_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_307_dilations_0 = const()[name = tensor<string, []>("x_307_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_307 = conv(dilations = x_307_dilations_0, groups = x_307_groups_0, pad = x_307_pad_0, pad_type = x_307_pad_type_0, strides = x_307_strides_0, weight = module_layers_11_conv_depthwise_conv_weight_quantized, x = new_x_47)[name = tensor<string, []>("x_307")];
+            tensor<int32, [3]> input_627_perm_0 = const()[name = tensor<string, []>("input_627_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_309_axes_0 = const()[name = tensor<string, []>("x_309_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_627 = transpose(perm = input_627_perm_0, x = x_307)[name = tensor<string, []>("transpose_257")];
+            tensor<fp32, [1, 2, 1024]> x_309 = layer_norm(axes = x_309_axes_0, beta = module_layers_11_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_11_conv_batch_norm_weight, x = input_627)[name = tensor<string, []>("x_309")];
+            tensor<int32, [3]> input_629_perm_0 = const()[name = tensor<string, []>("input_629_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_629 = transpose(perm = input_629_perm_0, x = x_309)[name = tensor<string, []>("transpose_256")];
+            tensor<fp32, [1, 1024, 2]> input_631 = silu(x = input_629)[name = tensor<string, []>("input_631")];
+            tensor<string, []> x_311_pad_type_0 = const()[name = tensor<string, []>("x_311_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_311_strides_0 = const()[name = tensor<string, []>("x_311_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_311_pad_0 = const()[name = tensor<string, []>("x_311_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_311_dilations_0 = const()[name = tensor<string, []>("x_311_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_311_groups_0 = const()[name = tensor<string, []>("x_311_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_311 = conv(dilations = x_311_dilations_0, groups = x_311_groups_0, pad = x_311_pad_0, pad_type = x_311_pad_type_0, strides = x_311_strides_0, weight = module_layers_11_conv_pointwise_conv2_weight_quantized, x = input_631)[name = tensor<string, []>("x_311")];
+            tensor<int32, [3]> input_633_perm_0 = const()[name = tensor<string, []>("input_633_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_633 = transpose(perm = input_633_perm_0, x = x_311)[name = tensor<string, []>("transpose_255")];
+            tensor<fp32, [1, 2, 1024]> input_635 = add(x = input_619, y = input_633)[name = tensor<string, []>("input_635")];
+            tensor<int32, [1]> input_637_axes_0 = const()[name = tensor<string, []>("input_637_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_637 = layer_norm(axes = input_637_axes_0, beta = module_layers_11_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_11_norm_feed_forward2_weight, x = input_635)[name = tensor<string, []>("input_637")];
+            tensor<fp32, [1, 2, 4096]> input_639 = linear(bias = linear_1_bias_0, weight = module_layers_11_feed_forward2_linear1_weight_quantized, x = input_637)[name = tensor<string, []>("linear_107")];
+            tensor<fp32, [1, 2, 4096]> input_641 = silu(x = input_639)[name = tensor<string, []>("input_641")];
+            tensor<fp32, [1, 2, 1024]> input_645 = linear(bias = linear_2_bias_0, weight = module_layers_11_feed_forward2_linear2_weight_quantized, x = input_641)[name = tensor<string, []>("linear_108")];
+            tensor<fp32, []> var_2863 = const()[name = tensor<string, []>("op_2863"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2864 = mul(x = input_645, y = var_2863)[name = tensor<string, []>("op_2864")];
+            tensor<fp32, [1, 2, 1024]> input_647 = add(x = input_635, y = var_2864)[name = tensor<string, []>("input_647")];
+            tensor<int32, [1]> input_649_axes_0 = const()[name = tensor<string, []>("input_649_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_649 = layer_norm(axes = input_649_axes_0, beta = module_layers_11_norm_out_bias, epsilon = var_38, gamma = module_layers_11_norm_out_weight, x = input_647)[name = tensor<string, []>("input_649")];
+            tensor<int32, [4]> cache_49_begin_0 = const()[name = tensor<string, []>("cache_49_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_49_end_0 = const()[name = tensor<string, []>("cache_49_end_0"), val = tensor<int32, [4]>([13, 1, 70, 1024])];
+            tensor<bool, [4]> cache_49_end_mask_0 = const()[name = tensor<string, []>("cache_49_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_49_squeeze_mask_0 = const()[name = tensor<string, []>("cache_49_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_49 = slice_by_index(begin = cache_49_begin_0, end = cache_49_end_0, end_mask = cache_49_end_mask_0, squeeze_mask = cache_49_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_49")];
+            tensor<int32, [4]> cache_51_begin_0 = const()[name = tensor<string, []>("cache_51_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
+            tensor<int32, [4]> cache_51_end_0 = const()[name = tensor<string, []>("cache_51_end_0"), val = tensor<int32, [4]>([13, 1, 1024, 8])];
+            tensor<bool, [4]> cache_51_end_mask_0 = const()[name = tensor<string, []>("cache_51_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_51_squeeze_mask_0 = const()[name = tensor<string, []>("cache_51_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_51 = slice_by_index(begin = cache_51_begin_0, end = cache_51_end_0, end_mask = cache_51_end_mask_0, squeeze_mask = cache_51_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_51")];
+            tensor<int32, [1]> input_651_axes_0 = const()[name = tensor<string, []>("input_651_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_651 = layer_norm(axes = input_651_axes_0, beta = module_layers_12_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_12_norm_feed_forward1_weight, x = input_649)[name = tensor<string, []>("input_651")];
+            tensor<fp32, [1, 2, 4096]> input_653 = linear(bias = linear_1_bias_0, weight = module_layers_12_feed_forward1_linear1_weight_quantized, x = input_651)[name = tensor<string, []>("linear_109")];
+            tensor<fp32, [1, 2, 4096]> input_655 = silu(x = input_653)[name = tensor<string, []>("input_655")];
+            tensor<fp32, [1, 2, 1024]> input_659 = linear(bias = linear_2_bias_0, weight = module_layers_12_feed_forward1_linear2_weight_quantized, x = input_655)[name = tensor<string, []>("linear_110")];
+            tensor<fp32, []> var_2898 = const()[name = tensor<string, []>("op_2898"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_2899 = mul(x = input_659, y = var_2898)[name = tensor<string, []>("op_2899")];
+            tensor<fp32, [1, 2, 1024]> input_661 = add(x = input_649, y = var_2899)[name = tensor<string, []>("input_661")];
+            tensor<int32, [1]> key_25_axes_0 = const()[name = tensor<string, []>("key_25_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_25 = layer_norm(axes = key_25_axes_0, beta = module_layers_12_norm_self_att_bias, epsilon = var_38, gamma = module_layers_12_norm_self_att_weight, x = input_661)[name = tensor<string, []>("key_25")];
+            tensor<bool, []> input_663_interleave_0 = const()[name = tensor<string, []>("input_663_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_663 = concat(axis = var_65, interleave = input_663_interleave_0, values = (cache_49, key_25))[name = tensor<string, []>("input_663")];
+            tensor<int32, [3]> var_2921_begin_0 = const()[name = tensor<string, []>("op_2921_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_2921_end_0 = const()[name = tensor<string, []>("op_2921_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_2921_end_mask_0 = const()[name = tensor<string, []>("op_2921_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_2921 = slice_by_index(begin = var_2921_begin_0, end = var_2921_end_0, end_mask = var_2921_end_mask_0, x = cache_49)[name = tensor<string, []>("op_2921")];
+            tensor<bool, []> var_2927_interleave_0 = const()[name = tensor<string, []>("op_2927_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_2927 = concat(axis = var_65, interleave = var_2927_interleave_0, values = (var_2921, key_25))[name = tensor<string, []>("op_2927")];
+            tensor<fp32, [1, 2, 1024]> var_2930 = linear(bias = linear_2_bias_0, weight = module_layers_12_self_attn_linear_q_weight_quantized, x = key_25)[name = tensor<string, []>("linear_111")];
+            tensor<int32, [4]> var_2931 = const()[name = tensor<string, []>("op_2931"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_73 = reshape(shape = var_2931, x = var_2930)[name = tensor<string, []>("q_73")];
+            tensor<fp32, [1, 72, 1024]> var_2934 = linear(bias = linear_2_bias_0, weight = module_layers_12_self_attn_linear_k_weight_quantized, x = input_663)[name = tensor<string, []>("linear_112")];
+            tensor<int32, [4]> var_2935 = const()[name = tensor<string, []>("op_2935"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_49 = reshape(shape = var_2935, x = var_2934)[name = tensor<string, []>("k_49")];
+            tensor<fp32, [1, 72, 1024]> var_2938 = linear(bias = linear_2_bias_0, weight = module_layers_12_self_attn_linear_v_weight_quantized, x = input_663)[name = tensor<string, []>("linear_113")];
+            tensor<int32, [4]> var_2939 = const()[name = tensor<string, []>("op_2939"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_25 = reshape(shape = var_2939, x = var_2938)[name = tensor<string, []>("v_25")];
+            tensor<int32, [4]> value_33_perm_0 = const()[name = tensor<string, []>("value_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_2951 = add(x = q_73, y = module_layers_12_self_attn_pos_bias_u)[name = tensor<string, []>("op_2951")];
+            tensor<fp32, [1, 2, 8, 128]> var_2953 = add(x = q_73, y = module_layers_12_self_attn_pos_bias_v)[name = tensor<string, []>("op_2953")];
+            tensor<int32, [4]> q_with_bias_v_25_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_2955_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_2955_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589101888))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589248640))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589248384)))];
+            tensor<bool, []> x_319_transpose_x_0 = const()[name = tensor<string, []>("x_319_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_319_transpose_y_0 = const()[name = tensor<string, []>("x_319_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_25 = transpose(perm = q_with_bias_v_25_perm_0, x = var_2953)[name = tensor<string, []>("transpose_254")];
+            tensor<fp32, [1, 8, 2, 143]> x_319 = matmul(transpose_x = x_319_transpose_x_0, transpose_y = x_319_transpose_y_0, x = q_with_bias_v_25, y = op_2955_quantized)[name = tensor<string, []>("x_319")];
+            tensor<fp32, []> const_235 = const()[name = tensor<string, []>("const_235"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_321_pad_0 = const()[name = tensor<string, []>("x_321_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_321_mode_0 = const()[name = tensor<string, []>("x_321_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_321 = pad(constant_val = const_235, mode = x_321_mode_0, pad = x_321_pad_0, x = x_319)[name = tensor<string, []>("x_321")];
+            tensor<int32, [4]> var_2963 = const()[name = tensor<string, []>("op_2963"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_323 = reshape(shape = var_2963, x = x_321)[name = tensor<string, []>("x_323")];
+            tensor<int32, [4]> var_2967_begin_0 = const()[name = tensor<string, []>("op_2967_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_2967_end_0 = const()[name = tensor<string, []>("op_2967_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_2967_end_mask_0 = const()[name = tensor<string, []>("op_2967_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_2967 = slice_by_index(begin = var_2967_begin_0, end = var_2967_end_0, end_mask = var_2967_end_mask_0, x = x_323)[name = tensor<string, []>("op_2967")];
+            tensor<int32, [4]> var_2968 = const()[name = tensor<string, []>("op_2968"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_49 = reshape(shape = var_2968, x = var_2967)[name = tensor<string, []>("matrix_bd_49")];
+            tensor<bool, []> matrix_ac_25_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_25_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_25_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_120_perm_0 = const()[name = tensor<string, []>("transpose_120_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_121_perm_0 = const()[name = tensor<string, []>("transpose_121_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_121 = transpose(perm = transpose_121_perm_0, x = k_49)[name = tensor<string, []>("transpose_252")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_120 = transpose(perm = transpose_120_perm_0, x = var_2951)[name = tensor<string, []>("transpose_253")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_25 = matmul(transpose_x = matrix_ac_25_transpose_x_0, transpose_y = matrix_ac_25_transpose_y_0, x = transpose_120, y = transpose_121)[name = tensor<string, []>("matrix_ac_25")];
+            tensor<int32, [4]> matrix_bd_51_begin_0 = const()[name = tensor<string, []>("matrix_bd_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_51_end_0 = const()[name = tensor<string, []>("matrix_bd_51_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_51_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_51 = slice_by_index(begin = matrix_bd_51_begin_0, end = matrix_bd_51_end_0, end_mask = matrix_bd_51_end_mask_0, x = matrix_bd_49)[name = tensor<string, []>("matrix_bd_51")];
+            tensor<fp32, [1, 8, 2, 72]> var_2977 = add(x = matrix_ac_25, y = matrix_bd_51)[name = tensor<string, []>("op_2977")];
+            tensor<fp32, []> _inversed_scores_49_y_0 = const()[name = tensor<string, []>("_inversed_scores_49_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_49 = mul(x = var_2977, y = _inversed_scores_49_y_0)[name = tensor<string, []>("_inversed_scores_49")];
+            tensor<fp32, [1, 8, 2, 72]> scores_51 = select(a = var_41, b = _inversed_scores_49, cond = mask_11)[name = tensor<string, []>("scores_51")];
+            tensor<fp32, [1, 8, 2, 72]> var_2983 = softmax(axis = var_56, x = scores_51)[name = tensor<string, []>("op_2983")];
+            tensor<fp32, [1, 8, 2, 72]> input_665 = select(a = var_40, b = var_2983, cond = mask_11)[name = tensor<string, []>("input_665")];
+            tensor<bool, []> x_325_transpose_x_0 = const()[name = tensor<string, []>("x_325_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_325_transpose_y_0 = const()[name = tensor<string, []>("x_325_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_33 = transpose(perm = value_33_perm_0, x = v_25)[name = tensor<string, []>("transpose_251")];
+            tensor<fp32, [1, 8, 2, 128]> x_325 = matmul(transpose_x = x_325_transpose_x_0, transpose_y = x_325_transpose_y_0, x = input_665, y = value_33)[name = tensor<string, []>("x_325")];
+            tensor<int32, [4]> var_2987_perm_0 = const()[name = tensor<string, []>("op_2987_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_2988 = const()[name = tensor<string, []>("op_2988"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_2987 = transpose(perm = var_2987_perm_0, x = x_325)[name = tensor<string, []>("transpose_250")];
+            tensor<fp32, [1, 2, 1024]> input_667 = reshape(shape = var_2988, x = var_2987)[name = tensor<string, []>("input_667")];
+            tensor<fp32, [1, 2, 1024]> input_669 = linear(bias = linear_2_bias_0, weight = module_layers_12_self_attn_linear_out_weight_quantized, x = input_667)[name = tensor<string, []>("linear_115")];
+            tensor<fp32, [1, 2, 1024]> input_671 = add(x = input_661, y = input_669)[name = tensor<string, []>("input_671")];
+            tensor<int32, [1]> x_329_axes_0 = const()[name = tensor<string, []>("x_329_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_329 = layer_norm(axes = x_329_axes_0, beta = module_layers_12_norm_conv_bias, epsilon = var_38, gamma = module_layers_12_norm_conv_weight, x = input_671)[name = tensor<string, []>("x_329")];
+            tensor<int32, [3]> input_673_perm_0 = const()[name = tensor<string, []>("input_673_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_675_pad_type_0 = const()[name = tensor<string, []>("input_675_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_675_strides_0 = const()[name = tensor<string, []>("input_675_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_675_pad_0 = const()[name = tensor<string, []>("input_675_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_675_dilations_0 = const()[name = tensor<string, []>("input_675_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_675_groups_0 = const()[name = tensor<string, []>("input_675_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_673 = transpose(perm = input_673_perm_0, x = x_329)[name = tensor<string, []>("transpose_249")];
+            tensor<fp32, [1, 2048, 2]> input_675 = conv(dilations = input_675_dilations_0, groups = input_675_groups_0, pad = input_675_pad_0, pad_type = input_675_pad_type_0, strides = input_675_strides_0, weight = module_layers_12_conv_pointwise_conv1_weight_quantized, x = input_673)[name = tensor<string, []>("input_675")];
+            tensor<int32, []> x_331_split_num_splits_0 = const()[name = tensor<string, []>("x_331_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_331_split_axis_0 = const()[name = tensor<string, []>("x_331_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_331_split_0, tensor<fp32, [1, 1024, 2]> x_331_split_1 = split(axis = x_331_split_axis_0, num_splits = x_331_split_num_splits_0, x = input_675)[name = tensor<string, []>("x_331_split")];
+            tensor<fp32, [1, 1024, 2]> x_331_split_1_sigmoid = sigmoid(x = x_331_split_1)[name = tensor<string, []>("x_331_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_331 = mul(x = x_331_split_0, y = x_331_split_1_sigmoid)[name = tensor<string, []>("x_331")];
+            tensor<fp32, [1, 1024, 2]> input_677 = select(a = var_40, b = x_331, cond = var_565)[name = tensor<string, []>("input_677")];
+            tensor<bool, []> new_x_51_interleave_0 = const()[name = tensor<string, []>("new_x_51_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_51 = concat(axis = var_56, interleave = new_x_51_interleave_0, values = (cache_51, input_677))[name = tensor<string, []>("new_x_51")];
+            tensor<int32, [3]> var_3026_begin_0 = const()[name = tensor<string, []>("op_3026_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_3026_end_0 = const()[name = tensor<string, []>("op_3026_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_3026_end_mask_0 = const()[name = tensor<string, []>("op_3026_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_3026 = slice_by_index(begin = var_3026_begin_0, end = var_3026_end_0, end_mask = var_3026_end_mask_0, x = new_x_51)[name = tensor<string, []>("op_3026")];
+            tensor<string, []> x_333_pad_type_0 = const()[name = tensor<string, []>("x_333_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_333_groups_0 = const()[name = tensor<string, []>("x_333_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_333_strides_0 = const()[name = tensor<string, []>("x_333_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_333_pad_0 = const()[name = tensor<string, []>("x_333_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_333_dilations_0 = const()[name = tensor<string, []>("x_333_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_333 = conv(dilations = x_333_dilations_0, groups = x_333_groups_0, pad = x_333_pad_0, pad_type = x_333_pad_type_0, strides = x_333_strides_0, weight = module_layers_12_conv_depthwise_conv_weight_quantized, x = new_x_51)[name = tensor<string, []>("x_333")];
+            tensor<int32, [3]> input_679_perm_0 = const()[name = tensor<string, []>("input_679_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_335_axes_0 = const()[name = tensor<string, []>("x_335_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_679 = transpose(perm = input_679_perm_0, x = x_333)[name = tensor<string, []>("transpose_248")];
+            tensor<fp32, [1, 2, 1024]> x_335 = layer_norm(axes = x_335_axes_0, beta = module_layers_12_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_12_conv_batch_norm_weight, x = input_679)[name = tensor<string, []>("x_335")];
+            tensor<int32, [3]> input_681_perm_0 = const()[name = tensor<string, []>("input_681_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_681 = transpose(perm = input_681_perm_0, x = x_335)[name = tensor<string, []>("transpose_247")];
+            tensor<fp32, [1, 1024, 2]> input_683 = silu(x = input_681)[name = tensor<string, []>("input_683")];
+            tensor<string, []> x_337_pad_type_0 = const()[name = tensor<string, []>("x_337_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_337_strides_0 = const()[name = tensor<string, []>("x_337_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_337_pad_0 = const()[name = tensor<string, []>("x_337_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_337_dilations_0 = const()[name = tensor<string, []>("x_337_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_337_groups_0 = const()[name = tensor<string, []>("x_337_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_337 = conv(dilations = x_337_dilations_0, groups = x_337_groups_0, pad = x_337_pad_0, pad_type = x_337_pad_type_0, strides = x_337_strides_0, weight = module_layers_12_conv_pointwise_conv2_weight_quantized, x = input_683)[name = tensor<string, []>("x_337")];
+            tensor<int32, [3]> input_685_perm_0 = const()[name = tensor<string, []>("input_685_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_685 = transpose(perm = input_685_perm_0, x = x_337)[name = tensor<string, []>("transpose_246")];
+            tensor<fp32, [1, 2, 1024]> input_687 = add(x = input_671, y = input_685)[name = tensor<string, []>("input_687")];
+            tensor<int32, [1]> input_689_axes_0 = const()[name = tensor<string, []>("input_689_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_689 = layer_norm(axes = input_689_axes_0, beta = module_layers_12_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_12_norm_feed_forward2_weight, x = input_687)[name = tensor<string, []>("input_689")];
+            tensor<fp32, [1, 2, 4096]> input_691 = linear(bias = linear_1_bias_0, weight = module_layers_12_feed_forward2_linear1_weight_quantized, x = input_689)[name = tensor<string, []>("linear_116")];
+            tensor<fp32, [1, 2, 4096]> input_693 = silu(x = input_691)[name = tensor<string, []>("input_693")];
+            tensor<fp32, [1, 2, 1024]> input_697 = linear(bias = linear_2_bias_0, weight = module_layers_12_feed_forward2_linear2_weight_quantized, x = input_693)[name = tensor<string, []>("linear_117")];
+            tensor<fp32, []> var_3067 = const()[name = tensor<string, []>("op_3067"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3068 = mul(x = input_697, y = var_3067)[name = tensor<string, []>("op_3068")];
+            tensor<fp32, [1, 2, 1024]> input_699 = add(x = input_687, y = var_3068)[name = tensor<string, []>("input_699")];
+            tensor<int32, [1]> input_701_axes_0 = const()[name = tensor<string, []>("input_701_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_701 = layer_norm(axes = input_701_axes_0, beta = module_layers_12_norm_out_bias, epsilon = var_38, gamma = module_layers_12_norm_out_weight, x = input_699)[name = tensor<string, []>("input_701")];
+            tensor<int32, [4]> cache_53_begin_0 = const()[name = tensor<string, []>("cache_53_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_53_end_0 = const()[name = tensor<string, []>("cache_53_end_0"), val = tensor<int32, [4]>([14, 1, 70, 1024])];
+            tensor<bool, [4]> cache_53_end_mask_0 = const()[name = tensor<string, []>("cache_53_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_53_squeeze_mask_0 = const()[name = tensor<string, []>("cache_53_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_53 = slice_by_index(begin = cache_53_begin_0, end = cache_53_end_0, end_mask = cache_53_end_mask_0, squeeze_mask = cache_53_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_53")];
+            tensor<int32, [4]> cache_55_begin_0 = const()[name = tensor<string, []>("cache_55_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
+            tensor<int32, [4]> cache_55_end_0 = const()[name = tensor<string, []>("cache_55_end_0"), val = tensor<int32, [4]>([14, 1, 1024, 8])];
+            tensor<bool, [4]> cache_55_end_mask_0 = const()[name = tensor<string, []>("cache_55_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_55_squeeze_mask_0 = const()[name = tensor<string, []>("cache_55_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_55 = slice_by_index(begin = cache_55_begin_0, end = cache_55_end_0, end_mask = cache_55_end_mask_0, squeeze_mask = cache_55_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_55")];
+            tensor<int32, [1]> input_703_axes_0 = const()[name = tensor<string, []>("input_703_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_703 = layer_norm(axes = input_703_axes_0, beta = module_layers_13_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_13_norm_feed_forward1_weight, x = input_701)[name = tensor<string, []>("input_703")];
+            tensor<fp32, [1, 2, 4096]> input_705 = linear(bias = linear_1_bias_0, weight = module_layers_13_feed_forward1_linear1_weight_quantized, x = input_703)[name = tensor<string, []>("linear_118")];
+            tensor<fp32, [1, 2, 4096]> input_707 = silu(x = input_705)[name = tensor<string, []>("input_707")];
+            tensor<fp32, [1, 2, 1024]> input_711 = linear(bias = linear_2_bias_0, weight = module_layers_13_feed_forward1_linear2_weight_quantized, x = input_707)[name = tensor<string, []>("linear_119")];
+            tensor<fp32, []> var_3102 = const()[name = tensor<string, []>("op_3102"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3103 = mul(x = input_711, y = var_3102)[name = tensor<string, []>("op_3103")];
+            tensor<fp32, [1, 2, 1024]> input_713 = add(x = input_701, y = var_3103)[name = tensor<string, []>("input_713")];
+            tensor<int32, [1]> key_27_axes_0 = const()[name = tensor<string, []>("key_27_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_27 = layer_norm(axes = key_27_axes_0, beta = module_layers_13_norm_self_att_bias, epsilon = var_38, gamma = module_layers_13_norm_self_att_weight, x = input_713)[name = tensor<string, []>("key_27")];
+            tensor<bool, []> input_715_interleave_0 = const()[name = tensor<string, []>("input_715_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_715 = concat(axis = var_65, interleave = input_715_interleave_0, values = (cache_53, key_27))[name = tensor<string, []>("input_715")];
+            tensor<int32, [3]> var_3125_begin_0 = const()[name = tensor<string, []>("op_3125_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_3125_end_0 = const()[name = tensor<string, []>("op_3125_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_3125_end_mask_0 = const()[name = tensor<string, []>("op_3125_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_3125 = slice_by_index(begin = var_3125_begin_0, end = var_3125_end_0, end_mask = var_3125_end_mask_0, x = cache_53)[name = tensor<string, []>("op_3125")];
+            tensor<bool, []> var_3131_interleave_0 = const()[name = tensor<string, []>("op_3131_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_3131 = concat(axis = var_65, interleave = var_3131_interleave_0, values = (var_3125, key_27))[name = tensor<string, []>("op_3131")];
+            tensor<fp32, [1, 2, 1024]> var_3134 = linear(bias = linear_2_bias_0, weight = module_layers_13_self_attn_linear_q_weight_quantized, x = key_27)[name = tensor<string, []>("linear_120")];
+            tensor<int32, [4]> var_3135 = const()[name = tensor<string, []>("op_3135"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_79 = reshape(shape = var_3135, x = var_3134)[name = tensor<string, []>("q_79")];
+            tensor<fp32, [1, 72, 1024]> var_3138 = linear(bias = linear_2_bias_0, weight = module_layers_13_self_attn_linear_k_weight_quantized, x = input_715)[name = tensor<string, []>("linear_121")];
+            tensor<int32, [4]> var_3139 = const()[name = tensor<string, []>("op_3139"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_53 = reshape(shape = var_3139, x = var_3138)[name = tensor<string, []>("k_53")];
+            tensor<fp32, [1, 72, 1024]> var_3142 = linear(bias = linear_2_bias_0, weight = module_layers_13_self_attn_linear_v_weight_quantized, x = input_715)[name = tensor<string, []>("linear_122")];
+            tensor<int32, [4]> var_3143 = const()[name = tensor<string, []>("op_3143"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_27 = reshape(shape = var_3143, x = var_3142)[name = tensor<string, []>("v_27")];
+            tensor<int32, [4]> value_35_perm_0 = const()[name = tensor<string, []>("value_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_3155 = add(x = q_79, y = module_layers_13_self_attn_pos_bias_u)[name = tensor<string, []>("op_3155")];
+            tensor<fp32, [1, 2, 8, 128]> var_3157 = add(x = q_79, y = module_layers_13_self_attn_pos_bias_v)[name = tensor<string, []>("op_3157")];
+            tensor<int32, [4]> q_with_bias_v_27_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_3159_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3159_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589249280))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589396032))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589395776)))];
+            tensor<bool, []> x_345_transpose_x_0 = const()[name = tensor<string, []>("x_345_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_345_transpose_y_0 = const()[name = tensor<string, []>("x_345_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_27 = transpose(perm = q_with_bias_v_27_perm_0, x = var_3157)[name = tensor<string, []>("transpose_245")];
+            tensor<fp32, [1, 8, 2, 143]> x_345 = matmul(transpose_x = x_345_transpose_x_0, transpose_y = x_345_transpose_y_0, x = q_with_bias_v_27, y = op_3159_quantized)[name = tensor<string, []>("x_345")];
+            tensor<fp32, []> const_248 = const()[name = tensor<string, []>("const_248"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_347_pad_0 = const()[name = tensor<string, []>("x_347_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_347_mode_0 = const()[name = tensor<string, []>("x_347_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_347 = pad(constant_val = const_248, mode = x_347_mode_0, pad = x_347_pad_0, x = x_345)[name = tensor<string, []>("x_347")];
+            tensor<int32, [4]> var_3167 = const()[name = tensor<string, []>("op_3167"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_349 = reshape(shape = var_3167, x = x_347)[name = tensor<string, []>("x_349")];
+            tensor<int32, [4]> var_3171_begin_0 = const()[name = tensor<string, []>("op_3171_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3171_end_0 = const()[name = tensor<string, []>("op_3171_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_3171_end_mask_0 = const()[name = tensor<string, []>("op_3171_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_3171 = slice_by_index(begin = var_3171_begin_0, end = var_3171_end_0, end_mask = var_3171_end_mask_0, x = x_349)[name = tensor<string, []>("op_3171")];
+            tensor<int32, [4]> var_3172 = const()[name = tensor<string, []>("op_3172"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_53 = reshape(shape = var_3172, x = var_3171)[name = tensor<string, []>("matrix_bd_53")];
+            tensor<bool, []> matrix_ac_27_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_27_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_27_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_122_perm_0 = const()[name = tensor<string, []>("transpose_122_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_123_perm_0 = const()[name = tensor<string, []>("transpose_123_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_123 = transpose(perm = transpose_123_perm_0, x = k_53)[name = tensor<string, []>("transpose_243")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_122 = transpose(perm = transpose_122_perm_0, x = var_3155)[name = tensor<string, []>("transpose_244")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_27 = matmul(transpose_x = matrix_ac_27_transpose_x_0, transpose_y = matrix_ac_27_transpose_y_0, x = transpose_122, y = transpose_123)[name = tensor<string, []>("matrix_ac_27")];
+            tensor<int32, [4]> matrix_bd_55_begin_0 = const()[name = tensor<string, []>("matrix_bd_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_55_end_0 = const()[name = tensor<string, []>("matrix_bd_55_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_55_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_55 = slice_by_index(begin = matrix_bd_55_begin_0, end = matrix_bd_55_end_0, end_mask = matrix_bd_55_end_mask_0, x = matrix_bd_53)[name = tensor<string, []>("matrix_bd_55")];
+            tensor<fp32, [1, 8, 2, 72]> var_3181 = add(x = matrix_ac_27, y = matrix_bd_55)[name = tensor<string, []>("op_3181")];
+            tensor<fp32, []> _inversed_scores_53_y_0 = const()[name = tensor<string, []>("_inversed_scores_53_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_53 = mul(x = var_3181, y = _inversed_scores_53_y_0)[name = tensor<string, []>("_inversed_scores_53")];
+            tensor<fp32, [1, 8, 2, 72]> scores_55 = select(a = var_41, b = _inversed_scores_53, cond = mask_11)[name = tensor<string, []>("scores_55")];
+            tensor<fp32, [1, 8, 2, 72]> var_3187 = softmax(axis = var_56, x = scores_55)[name = tensor<string, []>("op_3187")];
+            tensor<fp32, [1, 8, 2, 72]> input_717 = select(a = var_40, b = var_3187, cond = mask_11)[name = tensor<string, []>("input_717")];
+            tensor<bool, []> x_351_transpose_x_0 = const()[name = tensor<string, []>("x_351_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_351_transpose_y_0 = const()[name = tensor<string, []>("x_351_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_35 = transpose(perm = value_35_perm_0, x = v_27)[name = tensor<string, []>("transpose_242")];
+            tensor<fp32, [1, 8, 2, 128]> x_351 = matmul(transpose_x = x_351_transpose_x_0, transpose_y = x_351_transpose_y_0, x = input_717, y = value_35)[name = tensor<string, []>("x_351")];
+            tensor<int32, [4]> var_3191_perm_0 = const()[name = tensor<string, []>("op_3191_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3192 = const()[name = tensor<string, []>("op_3192"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_3191 = transpose(perm = var_3191_perm_0, x = x_351)[name = tensor<string, []>("transpose_241")];
+            tensor<fp32, [1, 2, 1024]> input_719 = reshape(shape = var_3192, x = var_3191)[name = tensor<string, []>("input_719")];
+            tensor<fp32, [1, 2, 1024]> input_721 = linear(bias = linear_2_bias_0, weight = module_layers_13_self_attn_linear_out_weight_quantized, x = input_719)[name = tensor<string, []>("linear_124")];
+            tensor<fp32, [1, 2, 1024]> input_723 = add(x = input_713, y = input_721)[name = tensor<string, []>("input_723")];
+            tensor<int32, [1]> x_355_axes_0 = const()[name = tensor<string, []>("x_355_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_355 = layer_norm(axes = x_355_axes_0, beta = module_layers_13_norm_conv_bias, epsilon = var_38, gamma = module_layers_13_norm_conv_weight, x = input_723)[name = tensor<string, []>("x_355")];
+            tensor<int32, [3]> input_725_perm_0 = const()[name = tensor<string, []>("input_725_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_727_pad_type_0 = const()[name = tensor<string, []>("input_727_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_727_strides_0 = const()[name = tensor<string, []>("input_727_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_727_pad_0 = const()[name = tensor<string, []>("input_727_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_727_dilations_0 = const()[name = tensor<string, []>("input_727_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_727_groups_0 = const()[name = tensor<string, []>("input_727_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_725 = transpose(perm = input_725_perm_0, x = x_355)[name = tensor<string, []>("transpose_240")];
+            tensor<fp32, [1, 2048, 2]> input_727 = conv(dilations = input_727_dilations_0, groups = input_727_groups_0, pad = input_727_pad_0, pad_type = input_727_pad_type_0, strides = input_727_strides_0, weight = module_layers_13_conv_pointwise_conv1_weight_quantized, x = input_725)[name = tensor<string, []>("input_727")];
+            tensor<int32, []> x_357_split_num_splits_0 = const()[name = tensor<string, []>("x_357_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_357_split_axis_0 = const()[name = tensor<string, []>("x_357_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_357_split_0, tensor<fp32, [1, 1024, 2]> x_357_split_1 = split(axis = x_357_split_axis_0, num_splits = x_357_split_num_splits_0, x = input_727)[name = tensor<string, []>("x_357_split")];
+            tensor<fp32, [1, 1024, 2]> x_357_split_1_sigmoid = sigmoid(x = x_357_split_1)[name = tensor<string, []>("x_357_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_357 = mul(x = x_357_split_0, y = x_357_split_1_sigmoid)[name = tensor<string, []>("x_357")];
+            tensor<fp32, [1, 1024, 2]> input_729 = select(a = var_40, b = x_357, cond = var_565)[name = tensor<string, []>("input_729")];
+            tensor<bool, []> new_x_55_interleave_0 = const()[name = tensor<string, []>("new_x_55_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_55 = concat(axis = var_56, interleave = new_x_55_interleave_0, values = (cache_55, input_729))[name = tensor<string, []>("new_x_55")];
+            tensor<int32, [3]> var_3230_begin_0 = const()[name = tensor<string, []>("op_3230_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_3230_end_0 = const()[name = tensor<string, []>("op_3230_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_3230_end_mask_0 = const()[name = tensor<string, []>("op_3230_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_3230 = slice_by_index(begin = var_3230_begin_0, end = var_3230_end_0, end_mask = var_3230_end_mask_0, x = new_x_55)[name = tensor<string, []>("op_3230")];
+            tensor<string, []> x_359_pad_type_0 = const()[name = tensor<string, []>("x_359_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_359_groups_0 = const()[name = tensor<string, []>("x_359_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_359_strides_0 = const()[name = tensor<string, []>("x_359_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_359_pad_0 = const()[name = tensor<string, []>("x_359_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_359_dilations_0 = const()[name = tensor<string, []>("x_359_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_359 = conv(dilations = x_359_dilations_0, groups = x_359_groups_0, pad = x_359_pad_0, pad_type = x_359_pad_type_0, strides = x_359_strides_0, weight = module_layers_13_conv_depthwise_conv_weight_quantized, x = new_x_55)[name = tensor<string, []>("x_359")];
+            tensor<int32, [3]> input_731_perm_0 = const()[name = tensor<string, []>("input_731_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_361_axes_0 = const()[name = tensor<string, []>("x_361_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_731 = transpose(perm = input_731_perm_0, x = x_359)[name = tensor<string, []>("transpose_239")];
+            tensor<fp32, [1, 2, 1024]> x_361 = layer_norm(axes = x_361_axes_0, beta = module_layers_13_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_13_conv_batch_norm_weight, x = input_731)[name = tensor<string, []>("x_361")];
+            tensor<int32, [3]> input_733_perm_0 = const()[name = tensor<string, []>("input_733_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_733 = transpose(perm = input_733_perm_0, x = x_361)[name = tensor<string, []>("transpose_238")];
+            tensor<fp32, [1, 1024, 2]> input_735 = silu(x = input_733)[name = tensor<string, []>("input_735")];
+            tensor<string, []> x_363_pad_type_0 = const()[name = tensor<string, []>("x_363_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_363_strides_0 = const()[name = tensor<string, []>("x_363_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_363_pad_0 = const()[name = tensor<string, []>("x_363_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_363_dilations_0 = const()[name = tensor<string, []>("x_363_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_363_groups_0 = const()[name = tensor<string, []>("x_363_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_363 = conv(dilations = x_363_dilations_0, groups = x_363_groups_0, pad = x_363_pad_0, pad_type = x_363_pad_type_0, strides = x_363_strides_0, weight = module_layers_13_conv_pointwise_conv2_weight_quantized, x = input_735)[name = tensor<string, []>("x_363")];
+            tensor<int32, [3]> input_737_perm_0 = const()[name = tensor<string, []>("input_737_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_737 = transpose(perm = input_737_perm_0, x = x_363)[name = tensor<string, []>("transpose_237")];
+            tensor<fp32, [1, 2, 1024]> input_739 = add(x = input_723, y = input_737)[name = tensor<string, []>("input_739")];
+            tensor<int32, [1]> input_741_axes_0 = const()[name = tensor<string, []>("input_741_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_741 = layer_norm(axes = input_741_axes_0, beta = module_layers_13_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_13_norm_feed_forward2_weight, x = input_739)[name = tensor<string, []>("input_741")];
+            tensor<fp32, [1, 2, 4096]> input_743 = linear(bias = linear_1_bias_0, weight = module_layers_13_feed_forward2_linear1_weight_quantized, x = input_741)[name = tensor<string, []>("linear_125")];
+            tensor<fp32, [1, 2, 4096]> input_745 = silu(x = input_743)[name = tensor<string, []>("input_745")];
+            tensor<fp32, [1, 2, 1024]> input_749 = linear(bias = linear_2_bias_0, weight = module_layers_13_feed_forward2_linear2_weight_quantized, x = input_745)[name = tensor<string, []>("linear_126")];
+            tensor<fp32, []> var_3271 = const()[name = tensor<string, []>("op_3271"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3272 = mul(x = input_749, y = var_3271)[name = tensor<string, []>("op_3272")];
+            tensor<fp32, [1, 2, 1024]> input_751 = add(x = input_739, y = var_3272)[name = tensor<string, []>("input_751")];
+            tensor<int32, [1]> input_753_axes_0 = const()[name = tensor<string, []>("input_753_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_753 = layer_norm(axes = input_753_axes_0, beta = module_layers_13_norm_out_bias, epsilon = var_38, gamma = module_layers_13_norm_out_weight, x = input_751)[name = tensor<string, []>("input_753")];
+            tensor<int32, [4]> cache_57_begin_0 = const()[name = tensor<string, []>("cache_57_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_57_end_0 = const()[name = tensor<string, []>("cache_57_end_0"), val = tensor<int32, [4]>([15, 1, 70, 1024])];
+            tensor<bool, [4]> cache_57_end_mask_0 = const()[name = tensor<string, []>("cache_57_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_57_squeeze_mask_0 = const()[name = tensor<string, []>("cache_57_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_57 = slice_by_index(begin = cache_57_begin_0, end = cache_57_end_0, end_mask = cache_57_end_mask_0, squeeze_mask = cache_57_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_57")];
+            tensor<int32, [4]> cache_59_begin_0 = const()[name = tensor<string, []>("cache_59_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
+            tensor<int32, [4]> cache_59_end_0 = const()[name = tensor<string, []>("cache_59_end_0"), val = tensor<int32, [4]>([15, 1, 1024, 8])];
+            tensor<bool, [4]> cache_59_end_mask_0 = const()[name = tensor<string, []>("cache_59_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_59_squeeze_mask_0 = const()[name = tensor<string, []>("cache_59_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_59 = slice_by_index(begin = cache_59_begin_0, end = cache_59_end_0, end_mask = cache_59_end_mask_0, squeeze_mask = cache_59_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_59")];
+            tensor<int32, [1]> input_755_axes_0 = const()[name = tensor<string, []>("input_755_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_755 = layer_norm(axes = input_755_axes_0, beta = module_layers_14_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_14_norm_feed_forward1_weight, x = input_753)[name = tensor<string, []>("input_755")];
+            tensor<fp32, [1, 2, 4096]> input_757 = linear(bias = linear_1_bias_0, weight = module_layers_14_feed_forward1_linear1_weight_quantized, x = input_755)[name = tensor<string, []>("linear_127")];
+            tensor<fp32, [1, 2, 4096]> input_759 = silu(x = input_757)[name = tensor<string, []>("input_759")];
+            tensor<fp32, [1, 2, 1024]> input_763 = linear(bias = linear_2_bias_0, weight = module_layers_14_feed_forward1_linear2_weight_quantized, x = input_759)[name = tensor<string, []>("linear_128")];
+            tensor<fp32, []> var_3306 = const()[name = tensor<string, []>("op_3306"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3307 = mul(x = input_763, y = var_3306)[name = tensor<string, []>("op_3307")];
+            tensor<fp32, [1, 2, 1024]> input_765 = add(x = input_753, y = var_3307)[name = tensor<string, []>("input_765")];
+            tensor<int32, [1]> key_29_axes_0 = const()[name = tensor<string, []>("key_29_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_29 = layer_norm(axes = key_29_axes_0, beta = module_layers_14_norm_self_att_bias, epsilon = var_38, gamma = module_layers_14_norm_self_att_weight, x = input_765)[name = tensor<string, []>("key_29")];
+            tensor<bool, []> input_767_interleave_0 = const()[name = tensor<string, []>("input_767_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_767 = concat(axis = var_65, interleave = input_767_interleave_0, values = (cache_57, key_29))[name = tensor<string, []>("input_767")];
+            tensor<int32, [3]> var_3329_begin_0 = const()[name = tensor<string, []>("op_3329_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_3329_end_0 = const()[name = tensor<string, []>("op_3329_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_3329_end_mask_0 = const()[name = tensor<string, []>("op_3329_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_3329 = slice_by_index(begin = var_3329_begin_0, end = var_3329_end_0, end_mask = var_3329_end_mask_0, x = cache_57)[name = tensor<string, []>("op_3329")];
+            tensor<bool, []> var_3335_interleave_0 = const()[name = tensor<string, []>("op_3335_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_3335 = concat(axis = var_65, interleave = var_3335_interleave_0, values = (var_3329, key_29))[name = tensor<string, []>("op_3335")];
+            tensor<fp32, [1, 2, 1024]> var_3338 = linear(bias = linear_2_bias_0, weight = module_layers_14_self_attn_linear_q_weight_quantized, x = key_29)[name = tensor<string, []>("linear_129")];
+            tensor<int32, [4]> var_3339 = const()[name = tensor<string, []>("op_3339"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_85 = reshape(shape = var_3339, x = var_3338)[name = tensor<string, []>("q_85")];
+            tensor<fp32, [1, 72, 1024]> var_3342 = linear(bias = linear_2_bias_0, weight = module_layers_14_self_attn_linear_k_weight_quantized, x = input_767)[name = tensor<string, []>("linear_130")];
+            tensor<int32, [4]> var_3343 = const()[name = tensor<string, []>("op_3343"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_57 = reshape(shape = var_3343, x = var_3342)[name = tensor<string, []>("k_57")];
+            tensor<fp32, [1, 72, 1024]> var_3346 = linear(bias = linear_2_bias_0, weight = module_layers_14_self_attn_linear_v_weight_quantized, x = input_767)[name = tensor<string, []>("linear_131")];
+            tensor<int32, [4]> var_3347 = const()[name = tensor<string, []>("op_3347"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_29 = reshape(shape = var_3347, x = var_3346)[name = tensor<string, []>("v_29")];
+            tensor<int32, [4]> value_37_perm_0 = const()[name = tensor<string, []>("value_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_3359 = add(x = q_85, y = module_layers_14_self_attn_pos_bias_u)[name = tensor<string, []>("op_3359")];
+            tensor<fp32, [1, 2, 8, 128]> var_3361 = add(x = q_85, y = module_layers_14_self_attn_pos_bias_v)[name = tensor<string, []>("op_3361")];
+            tensor<int32, [4]> q_with_bias_v_29_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_3363_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3363_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589396672))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589543424))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589543168)))];
+            tensor<bool, []> x_371_transpose_x_0 = const()[name = tensor<string, []>("x_371_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_371_transpose_y_0 = const()[name = tensor<string, []>("x_371_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_29 = transpose(perm = q_with_bias_v_29_perm_0, x = var_3361)[name = tensor<string, []>("transpose_236")];
+            tensor<fp32, [1, 8, 2, 143]> x_371 = matmul(transpose_x = x_371_transpose_x_0, transpose_y = x_371_transpose_y_0, x = q_with_bias_v_29, y = op_3363_quantized)[name = tensor<string, []>("x_371")];
+            tensor<fp32, []> const_261 = const()[name = tensor<string, []>("const_261"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_373_pad_0 = const()[name = tensor<string, []>("x_373_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_373_mode_0 = const()[name = tensor<string, []>("x_373_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_373 = pad(constant_val = const_261, mode = x_373_mode_0, pad = x_373_pad_0, x = x_371)[name = tensor<string, []>("x_373")];
+            tensor<int32, [4]> var_3371 = const()[name = tensor<string, []>("op_3371"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_375 = reshape(shape = var_3371, x = x_373)[name = tensor<string, []>("x_375")];
+            tensor<int32, [4]> var_3375_begin_0 = const()[name = tensor<string, []>("op_3375_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3375_end_0 = const()[name = tensor<string, []>("op_3375_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_3375_end_mask_0 = const()[name = tensor<string, []>("op_3375_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_3375 = slice_by_index(begin = var_3375_begin_0, end = var_3375_end_0, end_mask = var_3375_end_mask_0, x = x_375)[name = tensor<string, []>("op_3375")];
+            tensor<int32, [4]> var_3376 = const()[name = tensor<string, []>("op_3376"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_57 = reshape(shape = var_3376, x = var_3375)[name = tensor<string, []>("matrix_bd_57")];
+            tensor<bool, []> matrix_ac_29_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_29_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_29_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_124_perm_0 = const()[name = tensor<string, []>("transpose_124_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_125_perm_0 = const()[name = tensor<string, []>("transpose_125_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_125 = transpose(perm = transpose_125_perm_0, x = k_57)[name = tensor<string, []>("transpose_234")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_124 = transpose(perm = transpose_124_perm_0, x = var_3359)[name = tensor<string, []>("transpose_235")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_29 = matmul(transpose_x = matrix_ac_29_transpose_x_0, transpose_y = matrix_ac_29_transpose_y_0, x = transpose_124, y = transpose_125)[name = tensor<string, []>("matrix_ac_29")];
+            tensor<int32, [4]> matrix_bd_59_begin_0 = const()[name = tensor<string, []>("matrix_bd_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_59_end_0 = const()[name = tensor<string, []>("matrix_bd_59_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_59_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_59 = slice_by_index(begin = matrix_bd_59_begin_0, end = matrix_bd_59_end_0, end_mask = matrix_bd_59_end_mask_0, x = matrix_bd_57)[name = tensor<string, []>("matrix_bd_59")];
+            tensor<fp32, [1, 8, 2, 72]> var_3385 = add(x = matrix_ac_29, y = matrix_bd_59)[name = tensor<string, []>("op_3385")];
+            tensor<fp32, []> _inversed_scores_57_y_0 = const()[name = tensor<string, []>("_inversed_scores_57_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_57 = mul(x = var_3385, y = _inversed_scores_57_y_0)[name = tensor<string, []>("_inversed_scores_57")];
+            tensor<fp32, [1, 8, 2, 72]> scores_59 = select(a = var_41, b = _inversed_scores_57, cond = mask_11)[name = tensor<string, []>("scores_59")];
+            tensor<fp32, [1, 8, 2, 72]> var_3391 = softmax(axis = var_56, x = scores_59)[name = tensor<string, []>("op_3391")];
+            tensor<fp32, [1, 8, 2, 72]> input_769 = select(a = var_40, b = var_3391, cond = mask_11)[name = tensor<string, []>("input_769")];
+            tensor<bool, []> x_377_transpose_x_0 = const()[name = tensor<string, []>("x_377_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_377_transpose_y_0 = const()[name = tensor<string, []>("x_377_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_37 = transpose(perm = value_37_perm_0, x = v_29)[name = tensor<string, []>("transpose_233")];
+            tensor<fp32, [1, 8, 2, 128]> x_377 = matmul(transpose_x = x_377_transpose_x_0, transpose_y = x_377_transpose_y_0, x = input_769, y = value_37)[name = tensor<string, []>("x_377")];
+            tensor<int32, [4]> var_3395_perm_0 = const()[name = tensor<string, []>("op_3395_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3396 = const()[name = tensor<string, []>("op_3396"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_3395 = transpose(perm = var_3395_perm_0, x = x_377)[name = tensor<string, []>("transpose_232")];
+            tensor<fp32, [1, 2, 1024]> input_771 = reshape(shape = var_3396, x = var_3395)[name = tensor<string, []>("input_771")];
+            tensor<fp32, [1, 2, 1024]> input_773 = linear(bias = linear_2_bias_0, weight = module_layers_14_self_attn_linear_out_weight_quantized, x = input_771)[name = tensor<string, []>("linear_133")];
+            tensor<fp32, [1, 2, 1024]> input_775 = add(x = input_765, y = input_773)[name = tensor<string, []>("input_775")];
+            tensor<int32, [1]> x_381_axes_0 = const()[name = tensor<string, []>("x_381_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_381 = layer_norm(axes = x_381_axes_0, beta = module_layers_14_norm_conv_bias, epsilon = var_38, gamma = module_layers_14_norm_conv_weight, x = input_775)[name = tensor<string, []>("x_381")];
+            tensor<int32, [3]> input_777_perm_0 = const()[name = tensor<string, []>("input_777_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_779_pad_type_0 = const()[name = tensor<string, []>("input_779_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_779_strides_0 = const()[name = tensor<string, []>("input_779_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_779_pad_0 = const()[name = tensor<string, []>("input_779_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_779_dilations_0 = const()[name = tensor<string, []>("input_779_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_779_groups_0 = const()[name = tensor<string, []>("input_779_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_777 = transpose(perm = input_777_perm_0, x = x_381)[name = tensor<string, []>("transpose_231")];
+            tensor<fp32, [1, 2048, 2]> input_779 = conv(dilations = input_779_dilations_0, groups = input_779_groups_0, pad = input_779_pad_0, pad_type = input_779_pad_type_0, strides = input_779_strides_0, weight = module_layers_14_conv_pointwise_conv1_weight_quantized, x = input_777)[name = tensor<string, []>("input_779")];
+            tensor<int32, []> x_383_split_num_splits_0 = const()[name = tensor<string, []>("x_383_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_383_split_axis_0 = const()[name = tensor<string, []>("x_383_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_383_split_0, tensor<fp32, [1, 1024, 2]> x_383_split_1 = split(axis = x_383_split_axis_0, num_splits = x_383_split_num_splits_0, x = input_779)[name = tensor<string, []>("x_383_split")];
+            tensor<fp32, [1, 1024, 2]> x_383_split_1_sigmoid = sigmoid(x = x_383_split_1)[name = tensor<string, []>("x_383_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_383 = mul(x = x_383_split_0, y = x_383_split_1_sigmoid)[name = tensor<string, []>("x_383")];
+            tensor<fp32, [1, 1024, 2]> input_781 = select(a = var_40, b = x_383, cond = var_565)[name = tensor<string, []>("input_781")];
+            tensor<bool, []> new_x_59_interleave_0 = const()[name = tensor<string, []>("new_x_59_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_59 = concat(axis = var_56, interleave = new_x_59_interleave_0, values = (cache_59, input_781))[name = tensor<string, []>("new_x_59")];
+            tensor<int32, [3]> var_3434_begin_0 = const()[name = tensor<string, []>("op_3434_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_3434_end_0 = const()[name = tensor<string, []>("op_3434_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_3434_end_mask_0 = const()[name = tensor<string, []>("op_3434_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_3434 = slice_by_index(begin = var_3434_begin_0, end = var_3434_end_0, end_mask = var_3434_end_mask_0, x = new_x_59)[name = tensor<string, []>("op_3434")];
+            tensor<string, []> x_385_pad_type_0 = const()[name = tensor<string, []>("x_385_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_385_groups_0 = const()[name = tensor<string, []>("x_385_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_385_strides_0 = const()[name = tensor<string, []>("x_385_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_385_pad_0 = const()[name = tensor<string, []>("x_385_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_385_dilations_0 = const()[name = tensor<string, []>("x_385_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_385 = conv(dilations = x_385_dilations_0, groups = x_385_groups_0, pad = x_385_pad_0, pad_type = x_385_pad_type_0, strides = x_385_strides_0, weight = module_layers_14_conv_depthwise_conv_weight_quantized, x = new_x_59)[name = tensor<string, []>("x_385")];
+            tensor<int32, [3]> input_783_perm_0 = const()[name = tensor<string, []>("input_783_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_387_axes_0 = const()[name = tensor<string, []>("x_387_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_783 = transpose(perm = input_783_perm_0, x = x_385)[name = tensor<string, []>("transpose_230")];
+            tensor<fp32, [1, 2, 1024]> x_387 = layer_norm(axes = x_387_axes_0, beta = module_layers_14_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_14_conv_batch_norm_weight, x = input_783)[name = tensor<string, []>("x_387")];
+            tensor<int32, [3]> input_785_perm_0 = const()[name = tensor<string, []>("input_785_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_785 = transpose(perm = input_785_perm_0, x = x_387)[name = tensor<string, []>("transpose_229")];
+            tensor<fp32, [1, 1024, 2]> input_787 = silu(x = input_785)[name = tensor<string, []>("input_787")];
+            tensor<string, []> x_389_pad_type_0 = const()[name = tensor<string, []>("x_389_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_389_strides_0 = const()[name = tensor<string, []>("x_389_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_389_pad_0 = const()[name = tensor<string, []>("x_389_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_389_dilations_0 = const()[name = tensor<string, []>("x_389_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_389_groups_0 = const()[name = tensor<string, []>("x_389_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_389 = conv(dilations = x_389_dilations_0, groups = x_389_groups_0, pad = x_389_pad_0, pad_type = x_389_pad_type_0, strides = x_389_strides_0, weight = module_layers_14_conv_pointwise_conv2_weight_quantized, x = input_787)[name = tensor<string, []>("x_389")];
+            tensor<int32, [3]> input_789_perm_0 = const()[name = tensor<string, []>("input_789_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_789 = transpose(perm = input_789_perm_0, x = x_389)[name = tensor<string, []>("transpose_228")];
+            tensor<fp32, [1, 2, 1024]> input_791 = add(x = input_775, y = input_789)[name = tensor<string, []>("input_791")];
+            tensor<int32, [1]> input_793_axes_0 = const()[name = tensor<string, []>("input_793_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_793 = layer_norm(axes = input_793_axes_0, beta = module_layers_14_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_14_norm_feed_forward2_weight, x = input_791)[name = tensor<string, []>("input_793")];
+            tensor<fp32, [1, 2, 4096]> input_795 = linear(bias = linear_1_bias_0, weight = module_layers_14_feed_forward2_linear1_weight_quantized, x = input_793)[name = tensor<string, []>("linear_134")];
+            tensor<fp32, [1, 2, 4096]> input_797 = silu(x = input_795)[name = tensor<string, []>("input_797")];
+            tensor<fp32, [1, 2, 1024]> input_801 = linear(bias = linear_2_bias_0, weight = module_layers_14_feed_forward2_linear2_weight_quantized, x = input_797)[name = tensor<string, []>("linear_135")];
+            tensor<fp32, []> var_3475 = const()[name = tensor<string, []>("op_3475"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3476 = mul(x = input_801, y = var_3475)[name = tensor<string, []>("op_3476")];
+            tensor<fp32, [1, 2, 1024]> input_803 = add(x = input_791, y = var_3476)[name = tensor<string, []>("input_803")];
+            tensor<int32, [1]> input_805_axes_0 = const()[name = tensor<string, []>("input_805_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_805 = layer_norm(axes = input_805_axes_0, beta = module_layers_14_norm_out_bias, epsilon = var_38, gamma = module_layers_14_norm_out_weight, x = input_803)[name = tensor<string, []>("input_805")];
+            tensor<int32, [4]> cache_61_begin_0 = const()[name = tensor<string, []>("cache_61_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_61_end_0 = const()[name = tensor<string, []>("cache_61_end_0"), val = tensor<int32, [4]>([16, 1, 70, 1024])];
+            tensor<bool, [4]> cache_61_end_mask_0 = const()[name = tensor<string, []>("cache_61_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_61_squeeze_mask_0 = const()[name = tensor<string, []>("cache_61_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_61 = slice_by_index(begin = cache_61_begin_0, end = cache_61_end_0, end_mask = cache_61_end_mask_0, squeeze_mask = cache_61_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_61")];
+            tensor<int32, [4]> cache_63_begin_0 = const()[name = tensor<string, []>("cache_63_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
+            tensor<int32, [4]> cache_63_end_0 = const()[name = tensor<string, []>("cache_63_end_0"), val = tensor<int32, [4]>([16, 1, 1024, 8])];
+            tensor<bool, [4]> cache_63_end_mask_0 = const()[name = tensor<string, []>("cache_63_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_63_squeeze_mask_0 = const()[name = tensor<string, []>("cache_63_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_63 = slice_by_index(begin = cache_63_begin_0, end = cache_63_end_0, end_mask = cache_63_end_mask_0, squeeze_mask = cache_63_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_63")];
+            tensor<int32, [1]> input_807_axes_0 = const()[name = tensor<string, []>("input_807_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_807 = layer_norm(axes = input_807_axes_0, beta = module_layers_15_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_15_norm_feed_forward1_weight, x = input_805)[name = tensor<string, []>("input_807")];
+            tensor<fp32, [1, 2, 4096]> input_809 = linear(bias = linear_1_bias_0, weight = module_layers_15_feed_forward1_linear1_weight_quantized, x = input_807)[name = tensor<string, []>("linear_136")];
+            tensor<fp32, [1, 2, 4096]> input_811 = silu(x = input_809)[name = tensor<string, []>("input_811")];
+            tensor<fp32, [1, 2, 1024]> input_815 = linear(bias = linear_2_bias_0, weight = module_layers_15_feed_forward1_linear2_weight_quantized, x = input_811)[name = tensor<string, []>("linear_137")];
+            tensor<fp32, []> var_3510 = const()[name = tensor<string, []>("op_3510"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3511 = mul(x = input_815, y = var_3510)[name = tensor<string, []>("op_3511")];
+            tensor<fp32, [1, 2, 1024]> input_817 = add(x = input_805, y = var_3511)[name = tensor<string, []>("input_817")];
+            tensor<int32, [1]> key_31_axes_0 = const()[name = tensor<string, []>("key_31_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_31 = layer_norm(axes = key_31_axes_0, beta = module_layers_15_norm_self_att_bias, epsilon = var_38, gamma = module_layers_15_norm_self_att_weight, x = input_817)[name = tensor<string, []>("key_31")];
+            tensor<bool, []> input_819_interleave_0 = const()[name = tensor<string, []>("input_819_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_819 = concat(axis = var_65, interleave = input_819_interleave_0, values = (cache_61, key_31))[name = tensor<string, []>("input_819")];
+            tensor<int32, [3]> var_3533_begin_0 = const()[name = tensor<string, []>("op_3533_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_3533_end_0 = const()[name = tensor<string, []>("op_3533_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_3533_end_mask_0 = const()[name = tensor<string, []>("op_3533_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_3533 = slice_by_index(begin = var_3533_begin_0, end = var_3533_end_0, end_mask = var_3533_end_mask_0, x = cache_61)[name = tensor<string, []>("op_3533")];
+            tensor<bool, []> var_3539_interleave_0 = const()[name = tensor<string, []>("op_3539_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_3539 = concat(axis = var_65, interleave = var_3539_interleave_0, values = (var_3533, key_31))[name = tensor<string, []>("op_3539")];
+            tensor<fp32, [1, 2, 1024]> var_3542 = linear(bias = linear_2_bias_0, weight = module_layers_15_self_attn_linear_q_weight_quantized, x = key_31)[name = tensor<string, []>("linear_138")];
+            tensor<int32, [4]> var_3543 = const()[name = tensor<string, []>("op_3543"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_91 = reshape(shape = var_3543, x = var_3542)[name = tensor<string, []>("q_91")];
+            tensor<fp32, [1, 72, 1024]> var_3546 = linear(bias = linear_2_bias_0, weight = module_layers_15_self_attn_linear_k_weight_quantized, x = input_819)[name = tensor<string, []>("linear_139")];
+            tensor<int32, [4]> var_3547 = const()[name = tensor<string, []>("op_3547"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_61 = reshape(shape = var_3547, x = var_3546)[name = tensor<string, []>("k_61")];
+            tensor<fp32, [1, 72, 1024]> var_3550 = linear(bias = linear_2_bias_0, weight = module_layers_15_self_attn_linear_v_weight_quantized, x = input_819)[name = tensor<string, []>("linear_140")];
+            tensor<int32, [4]> var_3551 = const()[name = tensor<string, []>("op_3551"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_31 = reshape(shape = var_3551, x = var_3550)[name = tensor<string, []>("v_31")];
+            tensor<int32, [4]> value_39_perm_0 = const()[name = tensor<string, []>("value_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_3563 = add(x = q_91, y = module_layers_15_self_attn_pos_bias_u)[name = tensor<string, []>("op_3563")];
+            tensor<fp32, [1, 2, 8, 128]> var_3565 = add(x = q_91, y = module_layers_15_self_attn_pos_bias_v)[name = tensor<string, []>("op_3565")];
+            tensor<int32, [4]> q_with_bias_v_31_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_3567_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3567_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589544064))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589690816))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589690560)))];
+            tensor<bool, []> x_397_transpose_x_0 = const()[name = tensor<string, []>("x_397_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_397_transpose_y_0 = const()[name = tensor<string, []>("x_397_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_31 = transpose(perm = q_with_bias_v_31_perm_0, x = var_3565)[name = tensor<string, []>("transpose_227")];
+            tensor<fp32, [1, 8, 2, 143]> x_397 = matmul(transpose_x = x_397_transpose_x_0, transpose_y = x_397_transpose_y_0, x = q_with_bias_v_31, y = op_3567_quantized)[name = tensor<string, []>("x_397")];
+            tensor<fp32, []> const_274 = const()[name = tensor<string, []>("const_274"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_399_pad_0 = const()[name = tensor<string, []>("x_399_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_399_mode_0 = const()[name = tensor<string, []>("x_399_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_399 = pad(constant_val = const_274, mode = x_399_mode_0, pad = x_399_pad_0, x = x_397)[name = tensor<string, []>("x_399")];
+            tensor<int32, [4]> var_3575 = const()[name = tensor<string, []>("op_3575"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_401 = reshape(shape = var_3575, x = x_399)[name = tensor<string, []>("x_401")];
+            tensor<int32, [4]> var_3579_begin_0 = const()[name = tensor<string, []>("op_3579_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3579_end_0 = const()[name = tensor<string, []>("op_3579_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_3579_end_mask_0 = const()[name = tensor<string, []>("op_3579_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_3579 = slice_by_index(begin = var_3579_begin_0, end = var_3579_end_0, end_mask = var_3579_end_mask_0, x = x_401)[name = tensor<string, []>("op_3579")];
+            tensor<int32, [4]> var_3580 = const()[name = tensor<string, []>("op_3580"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_61 = reshape(shape = var_3580, x = var_3579)[name = tensor<string, []>("matrix_bd_61")];
+            tensor<bool, []> matrix_ac_31_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_31_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_31_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_126_perm_0 = const()[name = tensor<string, []>("transpose_126_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_127_perm_0 = const()[name = tensor<string, []>("transpose_127_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_127 = transpose(perm = transpose_127_perm_0, x = k_61)[name = tensor<string, []>("transpose_225")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_126 = transpose(perm = transpose_126_perm_0, x = var_3563)[name = tensor<string, []>("transpose_226")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_31 = matmul(transpose_x = matrix_ac_31_transpose_x_0, transpose_y = matrix_ac_31_transpose_y_0, x = transpose_126, y = transpose_127)[name = tensor<string, []>("matrix_ac_31")];
+            tensor<int32, [4]> matrix_bd_63_begin_0 = const()[name = tensor<string, []>("matrix_bd_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_63_end_0 = const()[name = tensor<string, []>("matrix_bd_63_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_63_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_63 = slice_by_index(begin = matrix_bd_63_begin_0, end = matrix_bd_63_end_0, end_mask = matrix_bd_63_end_mask_0, x = matrix_bd_61)[name = tensor<string, []>("matrix_bd_63")];
+            tensor<fp32, [1, 8, 2, 72]> var_3589 = add(x = matrix_ac_31, y = matrix_bd_63)[name = tensor<string, []>("op_3589")];
+            tensor<fp32, []> _inversed_scores_61_y_0 = const()[name = tensor<string, []>("_inversed_scores_61_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_61 = mul(x = var_3589, y = _inversed_scores_61_y_0)[name = tensor<string, []>("_inversed_scores_61")];
+            tensor<fp32, [1, 8, 2, 72]> scores_63 = select(a = var_41, b = _inversed_scores_61, cond = mask_11)[name = tensor<string, []>("scores_63")];
+            tensor<fp32, [1, 8, 2, 72]> var_3595 = softmax(axis = var_56, x = scores_63)[name = tensor<string, []>("op_3595")];
+            tensor<fp32, [1, 8, 2, 72]> input_821 = select(a = var_40, b = var_3595, cond = mask_11)[name = tensor<string, []>("input_821")];
+            tensor<bool, []> x_403_transpose_x_0 = const()[name = tensor<string, []>("x_403_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_403_transpose_y_0 = const()[name = tensor<string, []>("x_403_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_39 = transpose(perm = value_39_perm_0, x = v_31)[name = tensor<string, []>("transpose_224")];
+            tensor<fp32, [1, 8, 2, 128]> x_403 = matmul(transpose_x = x_403_transpose_x_0, transpose_y = x_403_transpose_y_0, x = input_821, y = value_39)[name = tensor<string, []>("x_403")];
+            tensor<int32, [4]> var_3599_perm_0 = const()[name = tensor<string, []>("op_3599_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3600 = const()[name = tensor<string, []>("op_3600"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_3599 = transpose(perm = var_3599_perm_0, x = x_403)[name = tensor<string, []>("transpose_223")];
+            tensor<fp32, [1, 2, 1024]> input_823 = reshape(shape = var_3600, x = var_3599)[name = tensor<string, []>("input_823")];
+            tensor<fp32, [1, 2, 1024]> input_825 = linear(bias = linear_2_bias_0, weight = module_layers_15_self_attn_linear_out_weight_quantized, x = input_823)[name = tensor<string, []>("linear_142")];
+            tensor<fp32, [1, 2, 1024]> input_827 = add(x = input_817, y = input_825)[name = tensor<string, []>("input_827")];
+            tensor<int32, [1]> x_407_axes_0 = const()[name = tensor<string, []>("x_407_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_407 = layer_norm(axes = x_407_axes_0, beta = module_layers_15_norm_conv_bias, epsilon = var_38, gamma = module_layers_15_norm_conv_weight, x = input_827)[name = tensor<string, []>("x_407")];
+            tensor<int32, [3]> input_829_perm_0 = const()[name = tensor<string, []>("input_829_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_831_pad_type_0 = const()[name = tensor<string, []>("input_831_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_831_strides_0 = const()[name = tensor<string, []>("input_831_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_831_pad_0 = const()[name = tensor<string, []>("input_831_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_831_dilations_0 = const()[name = tensor<string, []>("input_831_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_831_groups_0 = const()[name = tensor<string, []>("input_831_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_829 = transpose(perm = input_829_perm_0, x = x_407)[name = tensor<string, []>("transpose_222")];
+            tensor<fp32, [1, 2048, 2]> input_831 = conv(dilations = input_831_dilations_0, groups = input_831_groups_0, pad = input_831_pad_0, pad_type = input_831_pad_type_0, strides = input_831_strides_0, weight = module_layers_15_conv_pointwise_conv1_weight_quantized, x = input_829)[name = tensor<string, []>("input_831")];
+            tensor<int32, []> x_409_split_num_splits_0 = const()[name = tensor<string, []>("x_409_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_409_split_axis_0 = const()[name = tensor<string, []>("x_409_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_409_split_0, tensor<fp32, [1, 1024, 2]> x_409_split_1 = split(axis = x_409_split_axis_0, num_splits = x_409_split_num_splits_0, x = input_831)[name = tensor<string, []>("x_409_split")];
+            tensor<fp32, [1, 1024, 2]> x_409_split_1_sigmoid = sigmoid(x = x_409_split_1)[name = tensor<string, []>("x_409_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_409 = mul(x = x_409_split_0, y = x_409_split_1_sigmoid)[name = tensor<string, []>("x_409")];
+            tensor<fp32, [1, 1024, 2]> input_833 = select(a = var_40, b = x_409, cond = var_565)[name = tensor<string, []>("input_833")];
+            tensor<bool, []> new_x_63_interleave_0 = const()[name = tensor<string, []>("new_x_63_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_63 = concat(axis = var_56, interleave = new_x_63_interleave_0, values = (cache_63, input_833))[name = tensor<string, []>("new_x_63")];
+            tensor<int32, [3]> var_3638_begin_0 = const()[name = tensor<string, []>("op_3638_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_3638_end_0 = const()[name = tensor<string, []>("op_3638_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_3638_end_mask_0 = const()[name = tensor<string, []>("op_3638_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_3638 = slice_by_index(begin = var_3638_begin_0, end = var_3638_end_0, end_mask = var_3638_end_mask_0, x = new_x_63)[name = tensor<string, []>("op_3638")];
+            tensor<string, []> x_411_pad_type_0 = const()[name = tensor<string, []>("x_411_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_411_groups_0 = const()[name = tensor<string, []>("x_411_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_411_strides_0 = const()[name = tensor<string, []>("x_411_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_411_pad_0 = const()[name = tensor<string, []>("x_411_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_411_dilations_0 = const()[name = tensor<string, []>("x_411_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_411 = conv(dilations = x_411_dilations_0, groups = x_411_groups_0, pad = x_411_pad_0, pad_type = x_411_pad_type_0, strides = x_411_strides_0, weight = module_layers_15_conv_depthwise_conv_weight_quantized, x = new_x_63)[name = tensor<string, []>("x_411")];
+            tensor<int32, [3]> input_835_perm_0 = const()[name = tensor<string, []>("input_835_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_413_axes_0 = const()[name = tensor<string, []>("x_413_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_835 = transpose(perm = input_835_perm_0, x = x_411)[name = tensor<string, []>("transpose_221")];
+            tensor<fp32, [1, 2, 1024]> x_413 = layer_norm(axes = x_413_axes_0, beta = module_layers_15_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_15_conv_batch_norm_weight, x = input_835)[name = tensor<string, []>("x_413")];
+            tensor<int32, [3]> input_837_perm_0 = const()[name = tensor<string, []>("input_837_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_837 = transpose(perm = input_837_perm_0, x = x_413)[name = tensor<string, []>("transpose_220")];
+            tensor<fp32, [1, 1024, 2]> input_839 = silu(x = input_837)[name = tensor<string, []>("input_839")];
+            tensor<string, []> x_415_pad_type_0 = const()[name = tensor<string, []>("x_415_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_415_strides_0 = const()[name = tensor<string, []>("x_415_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_415_pad_0 = const()[name = tensor<string, []>("x_415_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_415_dilations_0 = const()[name = tensor<string, []>("x_415_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_415_groups_0 = const()[name = tensor<string, []>("x_415_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_415 = conv(dilations = x_415_dilations_0, groups = x_415_groups_0, pad = x_415_pad_0, pad_type = x_415_pad_type_0, strides = x_415_strides_0, weight = module_layers_15_conv_pointwise_conv2_weight_quantized, x = input_839)[name = tensor<string, []>("x_415")];
+            tensor<int32, [3]> input_841_perm_0 = const()[name = tensor<string, []>("input_841_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_841 = transpose(perm = input_841_perm_0, x = x_415)[name = tensor<string, []>("transpose_219")];
+            tensor<fp32, [1, 2, 1024]> input_843 = add(x = input_827, y = input_841)[name = tensor<string, []>("input_843")];
+            tensor<int32, [1]> input_845_axes_0 = const()[name = tensor<string, []>("input_845_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_845 = layer_norm(axes = input_845_axes_0, beta = module_layers_15_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_15_norm_feed_forward2_weight, x = input_843)[name = tensor<string, []>("input_845")];
+            tensor<fp32, [1, 2, 4096]> input_847 = linear(bias = linear_1_bias_0, weight = module_layers_15_feed_forward2_linear1_weight_quantized, x = input_845)[name = tensor<string, []>("linear_143")];
+            tensor<fp32, [1, 2, 4096]> input_849 = silu(x = input_847)[name = tensor<string, []>("input_849")];
+            tensor<fp32, [1, 2, 1024]> input_853 = linear(bias = linear_2_bias_0, weight = module_layers_15_feed_forward2_linear2_weight_quantized, x = input_849)[name = tensor<string, []>("linear_144")];
+            tensor<fp32, []> var_3679 = const()[name = tensor<string, []>("op_3679"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3680 = mul(x = input_853, y = var_3679)[name = tensor<string, []>("op_3680")];
+            tensor<fp32, [1, 2, 1024]> input_855 = add(x = input_843, y = var_3680)[name = tensor<string, []>("input_855")];
+            tensor<int32, [1]> input_857_axes_0 = const()[name = tensor<string, []>("input_857_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_857 = layer_norm(axes = input_857_axes_0, beta = module_layers_15_norm_out_bias, epsilon = var_38, gamma = module_layers_15_norm_out_weight, x = input_855)[name = tensor<string, []>("input_857")];
+            tensor<int32, [4]> cache_65_begin_0 = const()[name = tensor<string, []>("cache_65_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_65_end_0 = const()[name = tensor<string, []>("cache_65_end_0"), val = tensor<int32, [4]>([17, 1, 70, 1024])];
+            tensor<bool, [4]> cache_65_end_mask_0 = const()[name = tensor<string, []>("cache_65_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_65_squeeze_mask_0 = const()[name = tensor<string, []>("cache_65_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_65 = slice_by_index(begin = cache_65_begin_0, end = cache_65_end_0, end_mask = cache_65_end_mask_0, squeeze_mask = cache_65_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_65")];
+            tensor<int32, [4]> cache_67_begin_0 = const()[name = tensor<string, []>("cache_67_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
+            tensor<int32, [4]> cache_67_end_0 = const()[name = tensor<string, []>("cache_67_end_0"), val = tensor<int32, [4]>([17, 1, 1024, 8])];
+            tensor<bool, [4]> cache_67_end_mask_0 = const()[name = tensor<string, []>("cache_67_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_67_squeeze_mask_0 = const()[name = tensor<string, []>("cache_67_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_67 = slice_by_index(begin = cache_67_begin_0, end = cache_67_end_0, end_mask = cache_67_end_mask_0, squeeze_mask = cache_67_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_67")];
+            tensor<int32, [1]> input_859_axes_0 = const()[name = tensor<string, []>("input_859_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_859 = layer_norm(axes = input_859_axes_0, beta = module_layers_16_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_16_norm_feed_forward1_weight, x = input_857)[name = tensor<string, []>("input_859")];
+            tensor<fp32, [1, 2, 4096]> input_861 = linear(bias = linear_1_bias_0, weight = module_layers_16_feed_forward1_linear1_weight_quantized, x = input_859)[name = tensor<string, []>("linear_145")];
+            tensor<fp32, [1, 2, 4096]> input_863 = silu(x = input_861)[name = tensor<string, []>("input_863")];
+            tensor<fp32, [1, 2, 1024]> input_867 = linear(bias = linear_2_bias_0, weight = module_layers_16_feed_forward1_linear2_weight_quantized, x = input_863)[name = tensor<string, []>("linear_146")];
+            tensor<fp32, []> var_3714 = const()[name = tensor<string, []>("op_3714"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3715 = mul(x = input_867, y = var_3714)[name = tensor<string, []>("op_3715")];
+            tensor<fp32, [1, 2, 1024]> input_869 = add(x = input_857, y = var_3715)[name = tensor<string, []>("input_869")];
+            tensor<int32, [1]> key_33_axes_0 = const()[name = tensor<string, []>("key_33_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_33 = layer_norm(axes = key_33_axes_0, beta = module_layers_16_norm_self_att_bias, epsilon = var_38, gamma = module_layers_16_norm_self_att_weight, x = input_869)[name = tensor<string, []>("key_33")];
+            tensor<bool, []> input_871_interleave_0 = const()[name = tensor<string, []>("input_871_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_871 = concat(axis = var_65, interleave = input_871_interleave_0, values = (cache_65, key_33))[name = tensor<string, []>("input_871")];
+            tensor<int32, [3]> var_3737_begin_0 = const()[name = tensor<string, []>("op_3737_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_3737_end_0 = const()[name = tensor<string, []>("op_3737_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_3737_end_mask_0 = const()[name = tensor<string, []>("op_3737_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_3737 = slice_by_index(begin = var_3737_begin_0, end = var_3737_end_0, end_mask = var_3737_end_mask_0, x = cache_65)[name = tensor<string, []>("op_3737")];
+            tensor<bool, []> var_3743_interleave_0 = const()[name = tensor<string, []>("op_3743_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_3743 = concat(axis = var_65, interleave = var_3743_interleave_0, values = (var_3737, key_33))[name = tensor<string, []>("op_3743")];
+            tensor<fp32, [1, 2, 1024]> var_3746 = linear(bias = linear_2_bias_0, weight = module_layers_16_self_attn_linear_q_weight_quantized, x = key_33)[name = tensor<string, []>("linear_147")];
+            tensor<int32, [4]> var_3747 = const()[name = tensor<string, []>("op_3747"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_97 = reshape(shape = var_3747, x = var_3746)[name = tensor<string, []>("q_97")];
+            tensor<fp32, [1, 72, 1024]> var_3750 = linear(bias = linear_2_bias_0, weight = module_layers_16_self_attn_linear_k_weight_quantized, x = input_871)[name = tensor<string, []>("linear_148")];
+            tensor<int32, [4]> var_3751 = const()[name = tensor<string, []>("op_3751"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_65 = reshape(shape = var_3751, x = var_3750)[name = tensor<string, []>("k_65")];
+            tensor<fp32, [1, 72, 1024]> var_3754 = linear(bias = linear_2_bias_0, weight = module_layers_16_self_attn_linear_v_weight_quantized, x = input_871)[name = tensor<string, []>("linear_149")];
+            tensor<int32, [4]> var_3755 = const()[name = tensor<string, []>("op_3755"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_33 = reshape(shape = var_3755, x = var_3754)[name = tensor<string, []>("v_33")];
+            tensor<int32, [4]> value_41_perm_0 = const()[name = tensor<string, []>("value_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_3767 = add(x = q_97, y = module_layers_16_self_attn_pos_bias_u)[name = tensor<string, []>("op_3767")];
+            tensor<fp32, [1, 2, 8, 128]> var_3769 = add(x = q_97, y = module_layers_16_self_attn_pos_bias_v)[name = tensor<string, []>("op_3769")];
+            tensor<int32, [4]> q_with_bias_v_33_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_3771_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3771_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589691456))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589838208))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589837952)))];
+            tensor<bool, []> x_423_transpose_x_0 = const()[name = tensor<string, []>("x_423_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_423_transpose_y_0 = const()[name = tensor<string, []>("x_423_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_33 = transpose(perm = q_with_bias_v_33_perm_0, x = var_3769)[name = tensor<string, []>("transpose_218")];
+            tensor<fp32, [1, 8, 2, 143]> x_423 = matmul(transpose_x = x_423_transpose_x_0, transpose_y = x_423_transpose_y_0, x = q_with_bias_v_33, y = op_3771_quantized)[name = tensor<string, []>("x_423")];
+            tensor<fp32, []> const_287 = const()[name = tensor<string, []>("const_287"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_425_pad_0 = const()[name = tensor<string, []>("x_425_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_425_mode_0 = const()[name = tensor<string, []>("x_425_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_425 = pad(constant_val = const_287, mode = x_425_mode_0, pad = x_425_pad_0, x = x_423)[name = tensor<string, []>("x_425")];
+            tensor<int32, [4]> var_3779 = const()[name = tensor<string, []>("op_3779"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_427 = reshape(shape = var_3779, x = x_425)[name = tensor<string, []>("x_427")];
+            tensor<int32, [4]> var_3783_begin_0 = const()[name = tensor<string, []>("op_3783_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3783_end_0 = const()[name = tensor<string, []>("op_3783_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_3783_end_mask_0 = const()[name = tensor<string, []>("op_3783_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_3783 = slice_by_index(begin = var_3783_begin_0, end = var_3783_end_0, end_mask = var_3783_end_mask_0, x = x_427)[name = tensor<string, []>("op_3783")];
+            tensor<int32, [4]> var_3784 = const()[name = tensor<string, []>("op_3784"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_65 = reshape(shape = var_3784, x = var_3783)[name = tensor<string, []>("matrix_bd_65")];
+            tensor<bool, []> matrix_ac_33_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_33_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_33_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_33_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_128_perm_0 = const()[name = tensor<string, []>("transpose_128_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_129_perm_0 = const()[name = tensor<string, []>("transpose_129_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_129 = transpose(perm = transpose_129_perm_0, x = k_65)[name = tensor<string, []>("transpose_216")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_128 = transpose(perm = transpose_128_perm_0, x = var_3767)[name = tensor<string, []>("transpose_217")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_33 = matmul(transpose_x = matrix_ac_33_transpose_x_0, transpose_y = matrix_ac_33_transpose_y_0, x = transpose_128, y = transpose_129)[name = tensor<string, []>("matrix_ac_33")];
+            tensor<int32, [4]> matrix_bd_67_begin_0 = const()[name = tensor<string, []>("matrix_bd_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_67_end_0 = const()[name = tensor<string, []>("matrix_bd_67_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_67_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_67 = slice_by_index(begin = matrix_bd_67_begin_0, end = matrix_bd_67_end_0, end_mask = matrix_bd_67_end_mask_0, x = matrix_bd_65)[name = tensor<string, []>("matrix_bd_67")];
+            tensor<fp32, [1, 8, 2, 72]> var_3793 = add(x = matrix_ac_33, y = matrix_bd_67)[name = tensor<string, []>("op_3793")];
+            tensor<fp32, []> _inversed_scores_65_y_0 = const()[name = tensor<string, []>("_inversed_scores_65_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_65 = mul(x = var_3793, y = _inversed_scores_65_y_0)[name = tensor<string, []>("_inversed_scores_65")];
+            tensor<fp32, [1, 8, 2, 72]> scores_67 = select(a = var_41, b = _inversed_scores_65, cond = mask_11)[name = tensor<string, []>("scores_67")];
+            tensor<fp32, [1, 8, 2, 72]> var_3799 = softmax(axis = var_56, x = scores_67)[name = tensor<string, []>("op_3799")];
+            tensor<fp32, [1, 8, 2, 72]> input_873 = select(a = var_40, b = var_3799, cond = mask_11)[name = tensor<string, []>("input_873")];
+            tensor<bool, []> x_429_transpose_x_0 = const()[name = tensor<string, []>("x_429_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_429_transpose_y_0 = const()[name = tensor<string, []>("x_429_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_41 = transpose(perm = value_41_perm_0, x = v_33)[name = tensor<string, []>("transpose_215")];
+            tensor<fp32, [1, 8, 2, 128]> x_429 = matmul(transpose_x = x_429_transpose_x_0, transpose_y = x_429_transpose_y_0, x = input_873, y = value_41)[name = tensor<string, []>("x_429")];
+            tensor<int32, [4]> var_3803_perm_0 = const()[name = tensor<string, []>("op_3803_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_3804 = const()[name = tensor<string, []>("op_3804"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_3803 = transpose(perm = var_3803_perm_0, x = x_429)[name = tensor<string, []>("transpose_214")];
+            tensor<fp32, [1, 2, 1024]> input_875 = reshape(shape = var_3804, x = var_3803)[name = tensor<string, []>("input_875")];
+            tensor<fp32, [1, 2, 1024]> input_877 = linear(bias = linear_2_bias_0, weight = module_layers_16_self_attn_linear_out_weight_quantized, x = input_875)[name = tensor<string, []>("linear_151")];
+            tensor<fp32, [1, 2, 1024]> input_879 = add(x = input_869, y = input_877)[name = tensor<string, []>("input_879")];
+            tensor<int32, [1]> x_433_axes_0 = const()[name = tensor<string, []>("x_433_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_433 = layer_norm(axes = x_433_axes_0, beta = module_layers_16_norm_conv_bias, epsilon = var_38, gamma = module_layers_16_norm_conv_weight, x = input_879)[name = tensor<string, []>("x_433")];
+            tensor<int32, [3]> input_881_perm_0 = const()[name = tensor<string, []>("input_881_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_883_pad_type_0 = const()[name = tensor<string, []>("input_883_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_883_strides_0 = const()[name = tensor<string, []>("input_883_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_883_pad_0 = const()[name = tensor<string, []>("input_883_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_883_dilations_0 = const()[name = tensor<string, []>("input_883_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_883_groups_0 = const()[name = tensor<string, []>("input_883_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_881 = transpose(perm = input_881_perm_0, x = x_433)[name = tensor<string, []>("transpose_213")];
+            tensor<fp32, [1, 2048, 2]> input_883 = conv(dilations = input_883_dilations_0, groups = input_883_groups_0, pad = input_883_pad_0, pad_type = input_883_pad_type_0, strides = input_883_strides_0, weight = module_layers_16_conv_pointwise_conv1_weight_quantized, x = input_881)[name = tensor<string, []>("input_883")];
+            tensor<int32, []> x_435_split_num_splits_0 = const()[name = tensor<string, []>("x_435_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_435_split_axis_0 = const()[name = tensor<string, []>("x_435_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_435_split_0, tensor<fp32, [1, 1024, 2]> x_435_split_1 = split(axis = x_435_split_axis_0, num_splits = x_435_split_num_splits_0, x = input_883)[name = tensor<string, []>("x_435_split")];
+            tensor<fp32, [1, 1024, 2]> x_435_split_1_sigmoid = sigmoid(x = x_435_split_1)[name = tensor<string, []>("x_435_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_435 = mul(x = x_435_split_0, y = x_435_split_1_sigmoid)[name = tensor<string, []>("x_435")];
+            tensor<fp32, [1, 1024, 2]> input_885 = select(a = var_40, b = x_435, cond = var_565)[name = tensor<string, []>("input_885")];
+            tensor<bool, []> new_x_67_interleave_0 = const()[name = tensor<string, []>("new_x_67_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_67 = concat(axis = var_56, interleave = new_x_67_interleave_0, values = (cache_67, input_885))[name = tensor<string, []>("new_x_67")];
+            tensor<int32, [3]> var_3842_begin_0 = const()[name = tensor<string, []>("op_3842_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_3842_end_0 = const()[name = tensor<string, []>("op_3842_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_3842_end_mask_0 = const()[name = tensor<string, []>("op_3842_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_3842 = slice_by_index(begin = var_3842_begin_0, end = var_3842_end_0, end_mask = var_3842_end_mask_0, x = new_x_67)[name = tensor<string, []>("op_3842")];
+            tensor<string, []> x_437_pad_type_0 = const()[name = tensor<string, []>("x_437_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_437_groups_0 = const()[name = tensor<string, []>("x_437_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_437_strides_0 = const()[name = tensor<string, []>("x_437_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_437_pad_0 = const()[name = tensor<string, []>("x_437_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_437_dilations_0 = const()[name = tensor<string, []>("x_437_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_437 = conv(dilations = x_437_dilations_0, groups = x_437_groups_0, pad = x_437_pad_0, pad_type = x_437_pad_type_0, strides = x_437_strides_0, weight = module_layers_16_conv_depthwise_conv_weight_quantized, x = new_x_67)[name = tensor<string, []>("x_437")];
+            tensor<int32, [3]> input_887_perm_0 = const()[name = tensor<string, []>("input_887_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_439_axes_0 = const()[name = tensor<string, []>("x_439_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_887 = transpose(perm = input_887_perm_0, x = x_437)[name = tensor<string, []>("transpose_212")];
+            tensor<fp32, [1, 2, 1024]> x_439 = layer_norm(axes = x_439_axes_0, beta = module_layers_16_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_16_conv_batch_norm_weight, x = input_887)[name = tensor<string, []>("x_439")];
+            tensor<int32, [3]> input_889_perm_0 = const()[name = tensor<string, []>("input_889_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_889 = transpose(perm = input_889_perm_0, x = x_439)[name = tensor<string, []>("transpose_211")];
+            tensor<fp32, [1, 1024, 2]> input_891 = silu(x = input_889)[name = tensor<string, []>("input_891")];
+            tensor<string, []> x_441_pad_type_0 = const()[name = tensor<string, []>("x_441_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_441_strides_0 = const()[name = tensor<string, []>("x_441_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_441_pad_0 = const()[name = tensor<string, []>("x_441_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_441_dilations_0 = const()[name = tensor<string, []>("x_441_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_441_groups_0 = const()[name = tensor<string, []>("x_441_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_441 = conv(dilations = x_441_dilations_0, groups = x_441_groups_0, pad = x_441_pad_0, pad_type = x_441_pad_type_0, strides = x_441_strides_0, weight = module_layers_16_conv_pointwise_conv2_weight_quantized, x = input_891)[name = tensor<string, []>("x_441")];
+            tensor<int32, [3]> input_893_perm_0 = const()[name = tensor<string, []>("input_893_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_893 = transpose(perm = input_893_perm_0, x = x_441)[name = tensor<string, []>("transpose_210")];
+            tensor<fp32, [1, 2, 1024]> input_895 = add(x = input_879, y = input_893)[name = tensor<string, []>("input_895")];
+            tensor<int32, [1]> input_897_axes_0 = const()[name = tensor<string, []>("input_897_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_897 = layer_norm(axes = input_897_axes_0, beta = module_layers_16_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_16_norm_feed_forward2_weight, x = input_895)[name = tensor<string, []>("input_897")];
+            tensor<fp32, [1, 2, 4096]> input_899 = linear(bias = linear_1_bias_0, weight = module_layers_16_feed_forward2_linear1_weight_quantized, x = input_897)[name = tensor<string, []>("linear_152")];
+            tensor<fp32, [1, 2, 4096]> input_901 = silu(x = input_899)[name = tensor<string, []>("input_901")];
+            tensor<fp32, [1, 2, 1024]> input_905 = linear(bias = linear_2_bias_0, weight = module_layers_16_feed_forward2_linear2_weight_quantized, x = input_901)[name = tensor<string, []>("linear_153")];
+            tensor<fp32, []> var_3883 = const()[name = tensor<string, []>("op_3883"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3884 = mul(x = input_905, y = var_3883)[name = tensor<string, []>("op_3884")];
+            tensor<fp32, [1, 2, 1024]> input_907 = add(x = input_895, y = var_3884)[name = tensor<string, []>("input_907")];
+            tensor<int32, [1]> input_909_axes_0 = const()[name = tensor<string, []>("input_909_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_909 = layer_norm(axes = input_909_axes_0, beta = module_layers_16_norm_out_bias, epsilon = var_38, gamma = module_layers_16_norm_out_weight, x = input_907)[name = tensor<string, []>("input_909")];
+            tensor<int32, [4]> cache_69_begin_0 = const()[name = tensor<string, []>("cache_69_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_69_end_0 = const()[name = tensor<string, []>("cache_69_end_0"), val = tensor<int32, [4]>([18, 1, 70, 1024])];
+            tensor<bool, [4]> cache_69_end_mask_0 = const()[name = tensor<string, []>("cache_69_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_69_squeeze_mask_0 = const()[name = tensor<string, []>("cache_69_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_69 = slice_by_index(begin = cache_69_begin_0, end = cache_69_end_0, end_mask = cache_69_end_mask_0, squeeze_mask = cache_69_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_69")];
+            tensor<int32, [4]> cache_71_begin_0 = const()[name = tensor<string, []>("cache_71_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
+            tensor<int32, [4]> cache_71_end_0 = const()[name = tensor<string, []>("cache_71_end_0"), val = tensor<int32, [4]>([18, 1, 1024, 8])];
+            tensor<bool, [4]> cache_71_end_mask_0 = const()[name = tensor<string, []>("cache_71_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_71_squeeze_mask_0 = const()[name = tensor<string, []>("cache_71_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_71 = slice_by_index(begin = cache_71_begin_0, end = cache_71_end_0, end_mask = cache_71_end_mask_0, squeeze_mask = cache_71_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_71")];
+            tensor<int32, [1]> input_911_axes_0 = const()[name = tensor<string, []>("input_911_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_911 = layer_norm(axes = input_911_axes_0, beta = module_layers_17_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_17_norm_feed_forward1_weight, x = input_909)[name = tensor<string, []>("input_911")];
+            tensor<fp32, [1, 2, 4096]> input_913 = linear(bias = linear_1_bias_0, weight = module_layers_17_feed_forward1_linear1_weight_quantized, x = input_911)[name = tensor<string, []>("linear_154")];
+            tensor<fp32, [1, 2, 4096]> input_915 = silu(x = input_913)[name = tensor<string, []>("input_915")];
+            tensor<fp32, [1, 2, 1024]> input_919 = linear(bias = linear_2_bias_0, weight = module_layers_17_feed_forward1_linear2_weight_quantized, x = input_915)[name = tensor<string, []>("linear_155")];
+            tensor<fp32, []> var_3918 = const()[name = tensor<string, []>("op_3918"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_3919 = mul(x = input_919, y = var_3918)[name = tensor<string, []>("op_3919")];
+            tensor<fp32, [1, 2, 1024]> input_921 = add(x = input_909, y = var_3919)[name = tensor<string, []>("input_921")];
+            tensor<int32, [1]> key_35_axes_0 = const()[name = tensor<string, []>("key_35_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_35 = layer_norm(axes = key_35_axes_0, beta = module_layers_17_norm_self_att_bias, epsilon = var_38, gamma = module_layers_17_norm_self_att_weight, x = input_921)[name = tensor<string, []>("key_35")];
+            tensor<bool, []> input_923_interleave_0 = const()[name = tensor<string, []>("input_923_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_923 = concat(axis = var_65, interleave = input_923_interleave_0, values = (cache_69, key_35))[name = tensor<string, []>("input_923")];
+            tensor<int32, [3]> var_3941_begin_0 = const()[name = tensor<string, []>("op_3941_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_3941_end_0 = const()[name = tensor<string, []>("op_3941_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_3941_end_mask_0 = const()[name = tensor<string, []>("op_3941_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_3941 = slice_by_index(begin = var_3941_begin_0, end = var_3941_end_0, end_mask = var_3941_end_mask_0, x = cache_69)[name = tensor<string, []>("op_3941")];
+            tensor<bool, []> var_3947_interleave_0 = const()[name = tensor<string, []>("op_3947_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_3947 = concat(axis = var_65, interleave = var_3947_interleave_0, values = (var_3941, key_35))[name = tensor<string, []>("op_3947")];
+            tensor<fp32, [1, 2, 1024]> var_3950 = linear(bias = linear_2_bias_0, weight = module_layers_17_self_attn_linear_q_weight_quantized, x = key_35)[name = tensor<string, []>("linear_156")];
+            tensor<int32, [4]> var_3951 = const()[name = tensor<string, []>("op_3951"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_103 = reshape(shape = var_3951, x = var_3950)[name = tensor<string, []>("q_103")];
+            tensor<fp32, [1, 72, 1024]> var_3954 = linear(bias = linear_2_bias_0, weight = module_layers_17_self_attn_linear_k_weight_quantized, x = input_923)[name = tensor<string, []>("linear_157")];
+            tensor<int32, [4]> var_3955 = const()[name = tensor<string, []>("op_3955"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_69 = reshape(shape = var_3955, x = var_3954)[name = tensor<string, []>("k_69")];
+            tensor<fp32, [1, 72, 1024]> var_3958 = linear(bias = linear_2_bias_0, weight = module_layers_17_self_attn_linear_v_weight_quantized, x = input_923)[name = tensor<string, []>("linear_158")];
+            tensor<int32, [4]> var_3959 = const()[name = tensor<string, []>("op_3959"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_35 = reshape(shape = var_3959, x = var_3958)[name = tensor<string, []>("v_35")];
+            tensor<int32, [4]> value_43_perm_0 = const()[name = tensor<string, []>("value_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_3971 = add(x = q_103, y = module_layers_17_self_attn_pos_bias_u)[name = tensor<string, []>("op_3971")];
+            tensor<fp32, [1, 2, 8, 128]> var_3973 = add(x = q_103, y = module_layers_17_self_attn_pos_bias_v)[name = tensor<string, []>("op_3973")];
+            tensor<int32, [4]> q_with_bias_v_35_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_3975_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_3975_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589838848))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589985600))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589985344)))];
+            tensor<bool, []> x_449_transpose_x_0 = const()[name = tensor<string, []>("x_449_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_449_transpose_y_0 = const()[name = tensor<string, []>("x_449_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_35 = transpose(perm = q_with_bias_v_35_perm_0, x = var_3973)[name = tensor<string, []>("transpose_209")];
+            tensor<fp32, [1, 8, 2, 143]> x_449 = matmul(transpose_x = x_449_transpose_x_0, transpose_y = x_449_transpose_y_0, x = q_with_bias_v_35, y = op_3975_quantized)[name = tensor<string, []>("x_449")];
+            tensor<fp32, []> const_300 = const()[name = tensor<string, []>("const_300"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_451_pad_0 = const()[name = tensor<string, []>("x_451_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_451_mode_0 = const()[name = tensor<string, []>("x_451_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_451 = pad(constant_val = const_300, mode = x_451_mode_0, pad = x_451_pad_0, x = x_449)[name = tensor<string, []>("x_451")];
+            tensor<int32, [4]> var_3983 = const()[name = tensor<string, []>("op_3983"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_453 = reshape(shape = var_3983, x = x_451)[name = tensor<string, []>("x_453")];
+            tensor<int32, [4]> var_3987_begin_0 = const()[name = tensor<string, []>("op_3987_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_3987_end_0 = const()[name = tensor<string, []>("op_3987_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_3987_end_mask_0 = const()[name = tensor<string, []>("op_3987_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_3987 = slice_by_index(begin = var_3987_begin_0, end = var_3987_end_0, end_mask = var_3987_end_mask_0, x = x_453)[name = tensor<string, []>("op_3987")];
+            tensor<int32, [4]> var_3988 = const()[name = tensor<string, []>("op_3988"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_69 = reshape(shape = var_3988, x = var_3987)[name = tensor<string, []>("matrix_bd_69")];
+            tensor<bool, []> matrix_ac_35_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_35_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_35_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_35_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_130_perm_0 = const()[name = tensor<string, []>("transpose_130_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_131_perm_0 = const()[name = tensor<string, []>("transpose_131_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_131 = transpose(perm = transpose_131_perm_0, x = k_69)[name = tensor<string, []>("transpose_207")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_130 = transpose(perm = transpose_130_perm_0, x = var_3971)[name = tensor<string, []>("transpose_208")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_35 = matmul(transpose_x = matrix_ac_35_transpose_x_0, transpose_y = matrix_ac_35_transpose_y_0, x = transpose_130, y = transpose_131)[name = tensor<string, []>("matrix_ac_35")];
+            tensor<int32, [4]> matrix_bd_71_begin_0 = const()[name = tensor<string, []>("matrix_bd_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_71_end_0 = const()[name = tensor<string, []>("matrix_bd_71_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_71_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_71 = slice_by_index(begin = matrix_bd_71_begin_0, end = matrix_bd_71_end_0, end_mask = matrix_bd_71_end_mask_0, x = matrix_bd_69)[name = tensor<string, []>("matrix_bd_71")];
+            tensor<fp32, [1, 8, 2, 72]> var_3997 = add(x = matrix_ac_35, y = matrix_bd_71)[name = tensor<string, []>("op_3997")];
+            tensor<fp32, []> _inversed_scores_69_y_0 = const()[name = tensor<string, []>("_inversed_scores_69_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_69 = mul(x = var_3997, y = _inversed_scores_69_y_0)[name = tensor<string, []>("_inversed_scores_69")];
+            tensor<fp32, [1, 8, 2, 72]> scores_71 = select(a = var_41, b = _inversed_scores_69, cond = mask_11)[name = tensor<string, []>("scores_71")];
+            tensor<fp32, [1, 8, 2, 72]> var_4003 = softmax(axis = var_56, x = scores_71)[name = tensor<string, []>("op_4003")];
+            tensor<fp32, [1, 8, 2, 72]> input_925 = select(a = var_40, b = var_4003, cond = mask_11)[name = tensor<string, []>("input_925")];
+            tensor<bool, []> x_455_transpose_x_0 = const()[name = tensor<string, []>("x_455_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_455_transpose_y_0 = const()[name = tensor<string, []>("x_455_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_43 = transpose(perm = value_43_perm_0, x = v_35)[name = tensor<string, []>("transpose_206")];
+            tensor<fp32, [1, 8, 2, 128]> x_455 = matmul(transpose_x = x_455_transpose_x_0, transpose_y = x_455_transpose_y_0, x = input_925, y = value_43)[name = tensor<string, []>("x_455")];
+            tensor<int32, [4]> var_4007_perm_0 = const()[name = tensor<string, []>("op_4007_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4008 = const()[name = tensor<string, []>("op_4008"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_4007 = transpose(perm = var_4007_perm_0, x = x_455)[name = tensor<string, []>("transpose_205")];
+            tensor<fp32, [1, 2, 1024]> input_927 = reshape(shape = var_4008, x = var_4007)[name = tensor<string, []>("input_927")];
+            tensor<fp32, [1, 2, 1024]> input_929 = linear(bias = linear_2_bias_0, weight = module_layers_17_self_attn_linear_out_weight_quantized, x = input_927)[name = tensor<string, []>("linear_160")];
+            tensor<fp32, [1, 2, 1024]> input_931 = add(x = input_921, y = input_929)[name = tensor<string, []>("input_931")];
+            tensor<int32, [1]> x_459_axes_0 = const()[name = tensor<string, []>("x_459_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_459 = layer_norm(axes = x_459_axes_0, beta = module_layers_17_norm_conv_bias, epsilon = var_38, gamma = module_layers_17_norm_conv_weight, x = input_931)[name = tensor<string, []>("x_459")];
+            tensor<int32, [3]> input_933_perm_0 = const()[name = tensor<string, []>("input_933_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_935_pad_type_0 = const()[name = tensor<string, []>("input_935_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_935_strides_0 = const()[name = tensor<string, []>("input_935_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_935_pad_0 = const()[name = tensor<string, []>("input_935_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_935_dilations_0 = const()[name = tensor<string, []>("input_935_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_935_groups_0 = const()[name = tensor<string, []>("input_935_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_933 = transpose(perm = input_933_perm_0, x = x_459)[name = tensor<string, []>("transpose_204")];
+            tensor<fp32, [1, 2048, 2]> input_935 = conv(dilations = input_935_dilations_0, groups = input_935_groups_0, pad = input_935_pad_0, pad_type = input_935_pad_type_0, strides = input_935_strides_0, weight = module_layers_17_conv_pointwise_conv1_weight_quantized, x = input_933)[name = tensor<string, []>("input_935")];
+            tensor<int32, []> x_461_split_num_splits_0 = const()[name = tensor<string, []>("x_461_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_461_split_axis_0 = const()[name = tensor<string, []>("x_461_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_461_split_0, tensor<fp32, [1, 1024, 2]> x_461_split_1 = split(axis = x_461_split_axis_0, num_splits = x_461_split_num_splits_0, x = input_935)[name = tensor<string, []>("x_461_split")];
+            tensor<fp32, [1, 1024, 2]> x_461_split_1_sigmoid = sigmoid(x = x_461_split_1)[name = tensor<string, []>("x_461_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_461 = mul(x = x_461_split_0, y = x_461_split_1_sigmoid)[name = tensor<string, []>("x_461")];
+            tensor<fp32, [1, 1024, 2]> input_937 = select(a = var_40, b = x_461, cond = var_565)[name = tensor<string, []>("input_937")];
+            tensor<bool, []> new_x_71_interleave_0 = const()[name = tensor<string, []>("new_x_71_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_71 = concat(axis = var_56, interleave = new_x_71_interleave_0, values = (cache_71, input_937))[name = tensor<string, []>("new_x_71")];
+            tensor<int32, [3]> var_4046_begin_0 = const()[name = tensor<string, []>("op_4046_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_4046_end_0 = const()[name = tensor<string, []>("op_4046_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_4046_end_mask_0 = const()[name = tensor<string, []>("op_4046_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_4046 = slice_by_index(begin = var_4046_begin_0, end = var_4046_end_0, end_mask = var_4046_end_mask_0, x = new_x_71)[name = tensor<string, []>("op_4046")];
+            tensor<string, []> x_463_pad_type_0 = const()[name = tensor<string, []>("x_463_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_463_groups_0 = const()[name = tensor<string, []>("x_463_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_463_strides_0 = const()[name = tensor<string, []>("x_463_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_463_pad_0 = const()[name = tensor<string, []>("x_463_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_463_dilations_0 = const()[name = tensor<string, []>("x_463_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_463 = conv(dilations = x_463_dilations_0, groups = x_463_groups_0, pad = x_463_pad_0, pad_type = x_463_pad_type_0, strides = x_463_strides_0, weight = module_layers_17_conv_depthwise_conv_weight_quantized, x = new_x_71)[name = tensor<string, []>("x_463")];
+            tensor<int32, [3]> input_939_perm_0 = const()[name = tensor<string, []>("input_939_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_465_axes_0 = const()[name = tensor<string, []>("x_465_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_939 = transpose(perm = input_939_perm_0, x = x_463)[name = tensor<string, []>("transpose_203")];
+            tensor<fp32, [1, 2, 1024]> x_465 = layer_norm(axes = x_465_axes_0, beta = module_layers_17_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_17_conv_batch_norm_weight, x = input_939)[name = tensor<string, []>("x_465")];
+            tensor<int32, [3]> input_941_perm_0 = const()[name = tensor<string, []>("input_941_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_941 = transpose(perm = input_941_perm_0, x = x_465)[name = tensor<string, []>("transpose_202")];
+            tensor<fp32, [1, 1024, 2]> input_943 = silu(x = input_941)[name = tensor<string, []>("input_943")];
+            tensor<string, []> x_467_pad_type_0 = const()[name = tensor<string, []>("x_467_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_467_strides_0 = const()[name = tensor<string, []>("x_467_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_467_pad_0 = const()[name = tensor<string, []>("x_467_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_467_dilations_0 = const()[name = tensor<string, []>("x_467_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_467_groups_0 = const()[name = tensor<string, []>("x_467_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_467 = conv(dilations = x_467_dilations_0, groups = x_467_groups_0, pad = x_467_pad_0, pad_type = x_467_pad_type_0, strides = x_467_strides_0, weight = module_layers_17_conv_pointwise_conv2_weight_quantized, x = input_943)[name = tensor<string, []>("x_467")];
+            tensor<int32, [3]> input_945_perm_0 = const()[name = tensor<string, []>("input_945_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_945 = transpose(perm = input_945_perm_0, x = x_467)[name = tensor<string, []>("transpose_201")];
+            tensor<fp32, [1, 2, 1024]> input_947 = add(x = input_931, y = input_945)[name = tensor<string, []>("input_947")];
+            tensor<int32, [1]> input_949_axes_0 = const()[name = tensor<string, []>("input_949_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_949 = layer_norm(axes = input_949_axes_0, beta = module_layers_17_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_17_norm_feed_forward2_weight, x = input_947)[name = tensor<string, []>("input_949")];
+            tensor<fp32, [1, 2, 4096]> input_951 = linear(bias = linear_1_bias_0, weight = module_layers_17_feed_forward2_linear1_weight_quantized, x = input_949)[name = tensor<string, []>("linear_161")];
+            tensor<fp32, [1, 2, 4096]> input_953 = silu(x = input_951)[name = tensor<string, []>("input_953")];
+            tensor<fp32, [1, 2, 1024]> input_957 = linear(bias = linear_2_bias_0, weight = module_layers_17_feed_forward2_linear2_weight_quantized, x = input_953)[name = tensor<string, []>("linear_162")];
+            tensor<fp32, []> var_4087 = const()[name = tensor<string, []>("op_4087"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4088 = mul(x = input_957, y = var_4087)[name = tensor<string, []>("op_4088")];
+            tensor<fp32, [1, 2, 1024]> input_959 = add(x = input_947, y = var_4088)[name = tensor<string, []>("input_959")];
+            tensor<int32, [1]> input_961_axes_0 = const()[name = tensor<string, []>("input_961_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_961 = layer_norm(axes = input_961_axes_0, beta = module_layers_17_norm_out_bias, epsilon = var_38, gamma = module_layers_17_norm_out_weight, x = input_959)[name = tensor<string, []>("input_961")];
+            tensor<int32, [4]> cache_73_begin_0 = const()[name = tensor<string, []>("cache_73_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_73_end_0 = const()[name = tensor<string, []>("cache_73_end_0"), val = tensor<int32, [4]>([19, 1, 70, 1024])];
+            tensor<bool, [4]> cache_73_end_mask_0 = const()[name = tensor<string, []>("cache_73_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_73_squeeze_mask_0 = const()[name = tensor<string, []>("cache_73_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_73 = slice_by_index(begin = cache_73_begin_0, end = cache_73_end_0, end_mask = cache_73_end_mask_0, squeeze_mask = cache_73_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_73")];
+            tensor<int32, [4]> cache_75_begin_0 = const()[name = tensor<string, []>("cache_75_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
+            tensor<int32, [4]> cache_75_end_0 = const()[name = tensor<string, []>("cache_75_end_0"), val = tensor<int32, [4]>([19, 1, 1024, 8])];
+            tensor<bool, [4]> cache_75_end_mask_0 = const()[name = tensor<string, []>("cache_75_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_75_squeeze_mask_0 = const()[name = tensor<string, []>("cache_75_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_75 = slice_by_index(begin = cache_75_begin_0, end = cache_75_end_0, end_mask = cache_75_end_mask_0, squeeze_mask = cache_75_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_75")];
+            tensor<int32, [1]> input_963_axes_0 = const()[name = tensor<string, []>("input_963_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_963 = layer_norm(axes = input_963_axes_0, beta = module_layers_18_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_18_norm_feed_forward1_weight, x = input_961)[name = tensor<string, []>("input_963")];
+            tensor<fp32, [1, 2, 4096]> input_965 = linear(bias = linear_1_bias_0, weight = module_layers_18_feed_forward1_linear1_weight_quantized, x = input_963)[name = tensor<string, []>("linear_163")];
+            tensor<fp32, [1, 2, 4096]> input_967 = silu(x = input_965)[name = tensor<string, []>("input_967")];
+            tensor<fp32, [1, 2, 1024]> input_971 = linear(bias = linear_2_bias_0, weight = module_layers_18_feed_forward1_linear2_weight_quantized, x = input_967)[name = tensor<string, []>("linear_164")];
+            tensor<fp32, []> var_4122 = const()[name = tensor<string, []>("op_4122"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4123 = mul(x = input_971, y = var_4122)[name = tensor<string, []>("op_4123")];
+            tensor<fp32, [1, 2, 1024]> input_973 = add(x = input_961, y = var_4123)[name = tensor<string, []>("input_973")];
+            tensor<int32, [1]> key_37_axes_0 = const()[name = tensor<string, []>("key_37_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_37 = layer_norm(axes = key_37_axes_0, beta = module_layers_18_norm_self_att_bias, epsilon = var_38, gamma = module_layers_18_norm_self_att_weight, x = input_973)[name = tensor<string, []>("key_37")];
+            tensor<bool, []> input_975_interleave_0 = const()[name = tensor<string, []>("input_975_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_975 = concat(axis = var_65, interleave = input_975_interleave_0, values = (cache_73, key_37))[name = tensor<string, []>("input_975")];
+            tensor<int32, [3]> var_4145_begin_0 = const()[name = tensor<string, []>("op_4145_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_4145_end_0 = const()[name = tensor<string, []>("op_4145_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_4145_end_mask_0 = const()[name = tensor<string, []>("op_4145_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_4145 = slice_by_index(begin = var_4145_begin_0, end = var_4145_end_0, end_mask = var_4145_end_mask_0, x = cache_73)[name = tensor<string, []>("op_4145")];
+            tensor<bool, []> var_4151_interleave_0 = const()[name = tensor<string, []>("op_4151_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_4151 = concat(axis = var_65, interleave = var_4151_interleave_0, values = (var_4145, key_37))[name = tensor<string, []>("op_4151")];
+            tensor<fp32, [1, 2, 1024]> var_4154 = linear(bias = linear_2_bias_0, weight = module_layers_18_self_attn_linear_q_weight_quantized, x = key_37)[name = tensor<string, []>("linear_165")];
+            tensor<int32, [4]> var_4155 = const()[name = tensor<string, []>("op_4155"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_109 = reshape(shape = var_4155, x = var_4154)[name = tensor<string, []>("q_109")];
+            tensor<fp32, [1, 72, 1024]> var_4158 = linear(bias = linear_2_bias_0, weight = module_layers_18_self_attn_linear_k_weight_quantized, x = input_975)[name = tensor<string, []>("linear_166")];
+            tensor<int32, [4]> var_4159 = const()[name = tensor<string, []>("op_4159"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_73 = reshape(shape = var_4159, x = var_4158)[name = tensor<string, []>("k_73")];
+            tensor<fp32, [1, 72, 1024]> var_4162 = linear(bias = linear_2_bias_0, weight = module_layers_18_self_attn_linear_v_weight_quantized, x = input_975)[name = tensor<string, []>("linear_167")];
+            tensor<int32, [4]> var_4163 = const()[name = tensor<string, []>("op_4163"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_37 = reshape(shape = var_4163, x = var_4162)[name = tensor<string, []>("v_37")];
+            tensor<int32, [4]> value_45_perm_0 = const()[name = tensor<string, []>("value_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_4175 = add(x = q_109, y = module_layers_18_self_attn_pos_bias_u)[name = tensor<string, []>("op_4175")];
+            tensor<fp32, [1, 2, 8, 128]> var_4177 = add(x = q_109, y = module_layers_18_self_attn_pos_bias_v)[name = tensor<string, []>("op_4177")];
+            tensor<int32, [4]> q_with_bias_v_37_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_4179_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4179_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(589986240))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590132992))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590132736)))];
+            tensor<bool, []> x_475_transpose_x_0 = const()[name = tensor<string, []>("x_475_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_475_transpose_y_0 = const()[name = tensor<string, []>("x_475_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_37 = transpose(perm = q_with_bias_v_37_perm_0, x = var_4177)[name = tensor<string, []>("transpose_200")];
+            tensor<fp32, [1, 8, 2, 143]> x_475 = matmul(transpose_x = x_475_transpose_x_0, transpose_y = x_475_transpose_y_0, x = q_with_bias_v_37, y = op_4179_quantized)[name = tensor<string, []>("x_475")];
+            tensor<fp32, []> const_313 = const()[name = tensor<string, []>("const_313"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_477_pad_0 = const()[name = tensor<string, []>("x_477_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_477_mode_0 = const()[name = tensor<string, []>("x_477_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_477 = pad(constant_val = const_313, mode = x_477_mode_0, pad = x_477_pad_0, x = x_475)[name = tensor<string, []>("x_477")];
+            tensor<int32, [4]> var_4187 = const()[name = tensor<string, []>("op_4187"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_479 = reshape(shape = var_4187, x = x_477)[name = tensor<string, []>("x_479")];
+            tensor<int32, [4]> var_4191_begin_0 = const()[name = tensor<string, []>("op_4191_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4191_end_0 = const()[name = tensor<string, []>("op_4191_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_4191_end_mask_0 = const()[name = tensor<string, []>("op_4191_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_4191 = slice_by_index(begin = var_4191_begin_0, end = var_4191_end_0, end_mask = var_4191_end_mask_0, x = x_479)[name = tensor<string, []>("op_4191")];
+            tensor<int32, [4]> var_4192 = const()[name = tensor<string, []>("op_4192"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_73 = reshape(shape = var_4192, x = var_4191)[name = tensor<string, []>("matrix_bd_73")];
+            tensor<bool, []> matrix_ac_37_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_37_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_37_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_37_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_132_perm_0 = const()[name = tensor<string, []>("transpose_132_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_133_perm_0 = const()[name = tensor<string, []>("transpose_133_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_133 = transpose(perm = transpose_133_perm_0, x = k_73)[name = tensor<string, []>("transpose_198")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_132 = transpose(perm = transpose_132_perm_0, x = var_4175)[name = tensor<string, []>("transpose_199")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_37 = matmul(transpose_x = matrix_ac_37_transpose_x_0, transpose_y = matrix_ac_37_transpose_y_0, x = transpose_132, y = transpose_133)[name = tensor<string, []>("matrix_ac_37")];
+            tensor<int32, [4]> matrix_bd_75_begin_0 = const()[name = tensor<string, []>("matrix_bd_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_75_end_0 = const()[name = tensor<string, []>("matrix_bd_75_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_75_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_75 = slice_by_index(begin = matrix_bd_75_begin_0, end = matrix_bd_75_end_0, end_mask = matrix_bd_75_end_mask_0, x = matrix_bd_73)[name = tensor<string, []>("matrix_bd_75")];
+            tensor<fp32, [1, 8, 2, 72]> var_4201 = add(x = matrix_ac_37, y = matrix_bd_75)[name = tensor<string, []>("op_4201")];
+            tensor<fp32, []> _inversed_scores_73_y_0 = const()[name = tensor<string, []>("_inversed_scores_73_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_73 = mul(x = var_4201, y = _inversed_scores_73_y_0)[name = tensor<string, []>("_inversed_scores_73")];
+            tensor<fp32, [1, 8, 2, 72]> scores_75 = select(a = var_41, b = _inversed_scores_73, cond = mask_11)[name = tensor<string, []>("scores_75")];
+            tensor<fp32, [1, 8, 2, 72]> var_4207 = softmax(axis = var_56, x = scores_75)[name = tensor<string, []>("op_4207")];
+            tensor<fp32, [1, 8, 2, 72]> input_977 = select(a = var_40, b = var_4207, cond = mask_11)[name = tensor<string, []>("input_977")];
+            tensor<bool, []> x_481_transpose_x_0 = const()[name = tensor<string, []>("x_481_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_481_transpose_y_0 = const()[name = tensor<string, []>("x_481_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_45 = transpose(perm = value_45_perm_0, x = v_37)[name = tensor<string, []>("transpose_197")];
+            tensor<fp32, [1, 8, 2, 128]> x_481 = matmul(transpose_x = x_481_transpose_x_0, transpose_y = x_481_transpose_y_0, x = input_977, y = value_45)[name = tensor<string, []>("x_481")];
+            tensor<int32, [4]> var_4211_perm_0 = const()[name = tensor<string, []>("op_4211_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4212 = const()[name = tensor<string, []>("op_4212"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_4211 = transpose(perm = var_4211_perm_0, x = x_481)[name = tensor<string, []>("transpose_196")];
+            tensor<fp32, [1, 2, 1024]> input_979 = reshape(shape = var_4212, x = var_4211)[name = tensor<string, []>("input_979")];
+            tensor<fp32, [1, 2, 1024]> input_981 = linear(bias = linear_2_bias_0, weight = module_layers_18_self_attn_linear_out_weight_quantized, x = input_979)[name = tensor<string, []>("linear_169")];
+            tensor<fp32, [1, 2, 1024]> input_983 = add(x = input_973, y = input_981)[name = tensor<string, []>("input_983")];
+            tensor<int32, [1]> x_485_axes_0 = const()[name = tensor<string, []>("x_485_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_485 = layer_norm(axes = x_485_axes_0, beta = module_layers_18_norm_conv_bias, epsilon = var_38, gamma = module_layers_18_norm_conv_weight, x = input_983)[name = tensor<string, []>("x_485")];
+            tensor<int32, [3]> input_985_perm_0 = const()[name = tensor<string, []>("input_985_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_987_pad_type_0 = const()[name = tensor<string, []>("input_987_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_987_strides_0 = const()[name = tensor<string, []>("input_987_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_987_pad_0 = const()[name = tensor<string, []>("input_987_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_987_dilations_0 = const()[name = tensor<string, []>("input_987_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_987_groups_0 = const()[name = tensor<string, []>("input_987_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_985 = transpose(perm = input_985_perm_0, x = x_485)[name = tensor<string, []>("transpose_195")];
+            tensor<fp32, [1, 2048, 2]> input_987 = conv(dilations = input_987_dilations_0, groups = input_987_groups_0, pad = input_987_pad_0, pad_type = input_987_pad_type_0, strides = input_987_strides_0, weight = module_layers_18_conv_pointwise_conv1_weight_quantized, x = input_985)[name = tensor<string, []>("input_987")];
+            tensor<int32, []> x_487_split_num_splits_0 = const()[name = tensor<string, []>("x_487_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_487_split_axis_0 = const()[name = tensor<string, []>("x_487_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_487_split_0, tensor<fp32, [1, 1024, 2]> x_487_split_1 = split(axis = x_487_split_axis_0, num_splits = x_487_split_num_splits_0, x = input_987)[name = tensor<string, []>("x_487_split")];
+            tensor<fp32, [1, 1024, 2]> x_487_split_1_sigmoid = sigmoid(x = x_487_split_1)[name = tensor<string, []>("x_487_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_487 = mul(x = x_487_split_0, y = x_487_split_1_sigmoid)[name = tensor<string, []>("x_487")];
+            tensor<fp32, [1, 1024, 2]> input_989 = select(a = var_40, b = x_487, cond = var_565)[name = tensor<string, []>("input_989")];
+            tensor<bool, []> new_x_75_interleave_0 = const()[name = tensor<string, []>("new_x_75_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_75 = concat(axis = var_56, interleave = new_x_75_interleave_0, values = (cache_75, input_989))[name = tensor<string, []>("new_x_75")];
+            tensor<int32, [3]> var_4250_begin_0 = const()[name = tensor<string, []>("op_4250_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_4250_end_0 = const()[name = tensor<string, []>("op_4250_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_4250_end_mask_0 = const()[name = tensor<string, []>("op_4250_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_4250 = slice_by_index(begin = var_4250_begin_0, end = var_4250_end_0, end_mask = var_4250_end_mask_0, x = new_x_75)[name = tensor<string, []>("op_4250")];
+            tensor<string, []> x_489_pad_type_0 = const()[name = tensor<string, []>("x_489_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_489_groups_0 = const()[name = tensor<string, []>("x_489_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_489_strides_0 = const()[name = tensor<string, []>("x_489_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_489_pad_0 = const()[name = tensor<string, []>("x_489_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_489_dilations_0 = const()[name = tensor<string, []>("x_489_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_489 = conv(dilations = x_489_dilations_0, groups = x_489_groups_0, pad = x_489_pad_0, pad_type = x_489_pad_type_0, strides = x_489_strides_0, weight = module_layers_18_conv_depthwise_conv_weight_quantized, x = new_x_75)[name = tensor<string, []>("x_489")];
+            tensor<int32, [3]> input_991_perm_0 = const()[name = tensor<string, []>("input_991_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_491_axes_0 = const()[name = tensor<string, []>("x_491_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_991 = transpose(perm = input_991_perm_0, x = x_489)[name = tensor<string, []>("transpose_194")];
+            tensor<fp32, [1, 2, 1024]> x_491 = layer_norm(axes = x_491_axes_0, beta = module_layers_18_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_18_conv_batch_norm_weight, x = input_991)[name = tensor<string, []>("x_491")];
+            tensor<int32, [3]> input_993_perm_0 = const()[name = tensor<string, []>("input_993_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_993 = transpose(perm = input_993_perm_0, x = x_491)[name = tensor<string, []>("transpose_193")];
+            tensor<fp32, [1, 1024, 2]> input_995 = silu(x = input_993)[name = tensor<string, []>("input_995")];
+            tensor<string, []> x_493_pad_type_0 = const()[name = tensor<string, []>("x_493_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_493_strides_0 = const()[name = tensor<string, []>("x_493_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_493_pad_0 = const()[name = tensor<string, []>("x_493_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_493_dilations_0 = const()[name = tensor<string, []>("x_493_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_493_groups_0 = const()[name = tensor<string, []>("x_493_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_493 = conv(dilations = x_493_dilations_0, groups = x_493_groups_0, pad = x_493_pad_0, pad_type = x_493_pad_type_0, strides = x_493_strides_0, weight = module_layers_18_conv_pointwise_conv2_weight_quantized, x = input_995)[name = tensor<string, []>("x_493")];
+            tensor<int32, [3]> input_997_perm_0 = const()[name = tensor<string, []>("input_997_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_997 = transpose(perm = input_997_perm_0, x = x_493)[name = tensor<string, []>("transpose_192")];
+            tensor<fp32, [1, 2, 1024]> input_999 = add(x = input_983, y = input_997)[name = tensor<string, []>("input_999")];
+            tensor<int32, [1]> input_1001_axes_0 = const()[name = tensor<string, []>("input_1001_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1001 = layer_norm(axes = input_1001_axes_0, beta = module_layers_18_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_18_norm_feed_forward2_weight, x = input_999)[name = tensor<string, []>("input_1001")];
+            tensor<fp32, [1, 2, 4096]> input_1003 = linear(bias = linear_1_bias_0, weight = module_layers_18_feed_forward2_linear1_weight_quantized, x = input_1001)[name = tensor<string, []>("linear_170")];
+            tensor<fp32, [1, 2, 4096]> input_1005 = silu(x = input_1003)[name = tensor<string, []>("input_1005")];
+            tensor<fp32, [1, 2, 1024]> input_1009 = linear(bias = linear_2_bias_0, weight = module_layers_18_feed_forward2_linear2_weight_quantized, x = input_1005)[name = tensor<string, []>("linear_171")];
+            tensor<fp32, []> var_4291 = const()[name = tensor<string, []>("op_4291"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4292 = mul(x = input_1009, y = var_4291)[name = tensor<string, []>("op_4292")];
+            tensor<fp32, [1, 2, 1024]> input_1011 = add(x = input_999, y = var_4292)[name = tensor<string, []>("input_1011")];
+            tensor<int32, [1]> input_1013_axes_0 = const()[name = tensor<string, []>("input_1013_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1013 = layer_norm(axes = input_1013_axes_0, beta = module_layers_18_norm_out_bias, epsilon = var_38, gamma = module_layers_18_norm_out_weight, x = input_1011)[name = tensor<string, []>("input_1013")];
+            tensor<int32, [4]> cache_77_begin_0 = const()[name = tensor<string, []>("cache_77_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_77_end_0 = const()[name = tensor<string, []>("cache_77_end_0"), val = tensor<int32, [4]>([20, 1, 70, 1024])];
+            tensor<bool, [4]> cache_77_end_mask_0 = const()[name = tensor<string, []>("cache_77_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_77_squeeze_mask_0 = const()[name = tensor<string, []>("cache_77_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_77 = slice_by_index(begin = cache_77_begin_0, end = cache_77_end_0, end_mask = cache_77_end_mask_0, squeeze_mask = cache_77_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_77")];
+            tensor<int32, [4]> cache_79_begin_0 = const()[name = tensor<string, []>("cache_79_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
+            tensor<int32, [4]> cache_79_end_0 = const()[name = tensor<string, []>("cache_79_end_0"), val = tensor<int32, [4]>([20, 1, 1024, 8])];
+            tensor<bool, [4]> cache_79_end_mask_0 = const()[name = tensor<string, []>("cache_79_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_79_squeeze_mask_0 = const()[name = tensor<string, []>("cache_79_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_79 = slice_by_index(begin = cache_79_begin_0, end = cache_79_end_0, end_mask = cache_79_end_mask_0, squeeze_mask = cache_79_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_79")];
+            tensor<int32, [1]> input_1015_axes_0 = const()[name = tensor<string, []>("input_1015_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1015 = layer_norm(axes = input_1015_axes_0, beta = module_layers_19_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_19_norm_feed_forward1_weight, x = input_1013)[name = tensor<string, []>("input_1015")];
+            tensor<fp32, [1, 2, 4096]> input_1017 = linear(bias = linear_1_bias_0, weight = module_layers_19_feed_forward1_linear1_weight_quantized, x = input_1015)[name = tensor<string, []>("linear_172")];
+            tensor<fp32, [1, 2, 4096]> input_1019 = silu(x = input_1017)[name = tensor<string, []>("input_1019")];
+            tensor<fp32, [1, 2, 1024]> input_1023 = linear(bias = linear_2_bias_0, weight = module_layers_19_feed_forward1_linear2_weight_quantized, x = input_1019)[name = tensor<string, []>("linear_173")];
+            tensor<fp32, []> var_4326 = const()[name = tensor<string, []>("op_4326"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4327 = mul(x = input_1023, y = var_4326)[name = tensor<string, []>("op_4327")];
+            tensor<fp32, [1, 2, 1024]> input_1025 = add(x = input_1013, y = var_4327)[name = tensor<string, []>("input_1025")];
+            tensor<int32, [1]> key_39_axes_0 = const()[name = tensor<string, []>("key_39_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_39 = layer_norm(axes = key_39_axes_0, beta = module_layers_19_norm_self_att_bias, epsilon = var_38, gamma = module_layers_19_norm_self_att_weight, x = input_1025)[name = tensor<string, []>("key_39")];
+            tensor<bool, []> input_1027_interleave_0 = const()[name = tensor<string, []>("input_1027_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_1027 = concat(axis = var_65, interleave = input_1027_interleave_0, values = (cache_77, key_39))[name = tensor<string, []>("input_1027")];
+            tensor<int32, [3]> var_4349_begin_0 = const()[name = tensor<string, []>("op_4349_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_4349_end_0 = const()[name = tensor<string, []>("op_4349_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_4349_end_mask_0 = const()[name = tensor<string, []>("op_4349_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_4349 = slice_by_index(begin = var_4349_begin_0, end = var_4349_end_0, end_mask = var_4349_end_mask_0, x = cache_77)[name = tensor<string, []>("op_4349")];
+            tensor<bool, []> var_4355_interleave_0 = const()[name = tensor<string, []>("op_4355_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_4355 = concat(axis = var_65, interleave = var_4355_interleave_0, values = (var_4349, key_39))[name = tensor<string, []>("op_4355")];
+            tensor<fp32, [1, 2, 1024]> var_4358 = linear(bias = linear_2_bias_0, weight = module_layers_19_self_attn_linear_q_weight_quantized, x = key_39)[name = tensor<string, []>("linear_174")];
+            tensor<int32, [4]> var_4359 = const()[name = tensor<string, []>("op_4359"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_115 = reshape(shape = var_4359, x = var_4358)[name = tensor<string, []>("q_115")];
+            tensor<fp32, [1, 72, 1024]> var_4362 = linear(bias = linear_2_bias_0, weight = module_layers_19_self_attn_linear_k_weight_quantized, x = input_1027)[name = tensor<string, []>("linear_175")];
+            tensor<int32, [4]> var_4363 = const()[name = tensor<string, []>("op_4363"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_77 = reshape(shape = var_4363, x = var_4362)[name = tensor<string, []>("k_77")];
+            tensor<fp32, [1, 72, 1024]> var_4366 = linear(bias = linear_2_bias_0, weight = module_layers_19_self_attn_linear_v_weight_quantized, x = input_1027)[name = tensor<string, []>("linear_176")];
+            tensor<int32, [4]> var_4367 = const()[name = tensor<string, []>("op_4367"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_39 = reshape(shape = var_4367, x = var_4366)[name = tensor<string, []>("v_39")];
+            tensor<int32, [4]> value_47_perm_0 = const()[name = tensor<string, []>("value_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_4379 = add(x = q_115, y = module_layers_19_self_attn_pos_bias_u)[name = tensor<string, []>("op_4379")];
+            tensor<fp32, [1, 2, 8, 128]> var_4381 = add(x = q_115, y = module_layers_19_self_attn_pos_bias_v)[name = tensor<string, []>("op_4381")];
+            tensor<int32, [4]> q_with_bias_v_39_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_4383_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4383_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590133632))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590280384))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590280128)))];
+            tensor<bool, []> x_501_transpose_x_0 = const()[name = tensor<string, []>("x_501_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_501_transpose_y_0 = const()[name = tensor<string, []>("x_501_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_39 = transpose(perm = q_with_bias_v_39_perm_0, x = var_4381)[name = tensor<string, []>("transpose_191")];
+            tensor<fp32, [1, 8, 2, 143]> x_501 = matmul(transpose_x = x_501_transpose_x_0, transpose_y = x_501_transpose_y_0, x = q_with_bias_v_39, y = op_4383_quantized)[name = tensor<string, []>("x_501")];
+            tensor<fp32, []> const_326 = const()[name = tensor<string, []>("const_326"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_503_pad_0 = const()[name = tensor<string, []>("x_503_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_503_mode_0 = const()[name = tensor<string, []>("x_503_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_503 = pad(constant_val = const_326, mode = x_503_mode_0, pad = x_503_pad_0, x = x_501)[name = tensor<string, []>("x_503")];
+            tensor<int32, [4]> var_4391 = const()[name = tensor<string, []>("op_4391"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_505 = reshape(shape = var_4391, x = x_503)[name = tensor<string, []>("x_505")];
+            tensor<int32, [4]> var_4395_begin_0 = const()[name = tensor<string, []>("op_4395_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4395_end_0 = const()[name = tensor<string, []>("op_4395_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_4395_end_mask_0 = const()[name = tensor<string, []>("op_4395_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_4395 = slice_by_index(begin = var_4395_begin_0, end = var_4395_end_0, end_mask = var_4395_end_mask_0, x = x_505)[name = tensor<string, []>("op_4395")];
+            tensor<int32, [4]> var_4396 = const()[name = tensor<string, []>("op_4396"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_77 = reshape(shape = var_4396, x = var_4395)[name = tensor<string, []>("matrix_bd_77")];
+            tensor<bool, []> matrix_ac_39_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_39_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_39_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_39_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_134_perm_0 = const()[name = tensor<string, []>("transpose_134_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_135_perm_0 = const()[name = tensor<string, []>("transpose_135_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_135 = transpose(perm = transpose_135_perm_0, x = k_77)[name = tensor<string, []>("transpose_189")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_134 = transpose(perm = transpose_134_perm_0, x = var_4379)[name = tensor<string, []>("transpose_190")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_39 = matmul(transpose_x = matrix_ac_39_transpose_x_0, transpose_y = matrix_ac_39_transpose_y_0, x = transpose_134, y = transpose_135)[name = tensor<string, []>("matrix_ac_39")];
+            tensor<int32, [4]> matrix_bd_79_begin_0 = const()[name = tensor<string, []>("matrix_bd_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_79_end_0 = const()[name = tensor<string, []>("matrix_bd_79_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_79_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_79 = slice_by_index(begin = matrix_bd_79_begin_0, end = matrix_bd_79_end_0, end_mask = matrix_bd_79_end_mask_0, x = matrix_bd_77)[name = tensor<string, []>("matrix_bd_79")];
+            tensor<fp32, [1, 8, 2, 72]> var_4405 = add(x = matrix_ac_39, y = matrix_bd_79)[name = tensor<string, []>("op_4405")];
+            tensor<fp32, []> _inversed_scores_77_y_0 = const()[name = tensor<string, []>("_inversed_scores_77_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_77 = mul(x = var_4405, y = _inversed_scores_77_y_0)[name = tensor<string, []>("_inversed_scores_77")];
+            tensor<fp32, [1, 8, 2, 72]> scores_79 = select(a = var_41, b = _inversed_scores_77, cond = mask_11)[name = tensor<string, []>("scores_79")];
+            tensor<fp32, [1, 8, 2, 72]> var_4411 = softmax(axis = var_56, x = scores_79)[name = tensor<string, []>("op_4411")];
+            tensor<fp32, [1, 8, 2, 72]> input_1029 = select(a = var_40, b = var_4411, cond = mask_11)[name = tensor<string, []>("input_1029")];
+            tensor<bool, []> x_507_transpose_x_0 = const()[name = tensor<string, []>("x_507_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_507_transpose_y_0 = const()[name = tensor<string, []>("x_507_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_47 = transpose(perm = value_47_perm_0, x = v_39)[name = tensor<string, []>("transpose_188")];
+            tensor<fp32, [1, 8, 2, 128]> x_507 = matmul(transpose_x = x_507_transpose_x_0, transpose_y = x_507_transpose_y_0, x = input_1029, y = value_47)[name = tensor<string, []>("x_507")];
+            tensor<int32, [4]> var_4415_perm_0 = const()[name = tensor<string, []>("op_4415_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4416 = const()[name = tensor<string, []>("op_4416"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_4415 = transpose(perm = var_4415_perm_0, x = x_507)[name = tensor<string, []>("transpose_187")];
+            tensor<fp32, [1, 2, 1024]> input_1031 = reshape(shape = var_4416, x = var_4415)[name = tensor<string, []>("input_1031")];
+            tensor<fp32, [1, 2, 1024]> input_1033 = linear(bias = linear_2_bias_0, weight = module_layers_19_self_attn_linear_out_weight_quantized, x = input_1031)[name = tensor<string, []>("linear_178")];
+            tensor<fp32, [1, 2, 1024]> input_1035 = add(x = input_1025, y = input_1033)[name = tensor<string, []>("input_1035")];
+            tensor<int32, [1]> x_511_axes_0 = const()[name = tensor<string, []>("x_511_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_511 = layer_norm(axes = x_511_axes_0, beta = module_layers_19_norm_conv_bias, epsilon = var_38, gamma = module_layers_19_norm_conv_weight, x = input_1035)[name = tensor<string, []>("x_511")];
+            tensor<int32, [3]> input_1037_perm_0 = const()[name = tensor<string, []>("input_1037_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1039_pad_type_0 = const()[name = tensor<string, []>("input_1039_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1039_strides_0 = const()[name = tensor<string, []>("input_1039_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1039_pad_0 = const()[name = tensor<string, []>("input_1039_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1039_dilations_0 = const()[name = tensor<string, []>("input_1039_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1039_groups_0 = const()[name = tensor<string, []>("input_1039_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_1037 = transpose(perm = input_1037_perm_0, x = x_511)[name = tensor<string, []>("transpose_186")];
+            tensor<fp32, [1, 2048, 2]> input_1039 = conv(dilations = input_1039_dilations_0, groups = input_1039_groups_0, pad = input_1039_pad_0, pad_type = input_1039_pad_type_0, strides = input_1039_strides_0, weight = module_layers_19_conv_pointwise_conv1_weight_quantized, x = input_1037)[name = tensor<string, []>("input_1039")];
+            tensor<int32, []> x_513_split_num_splits_0 = const()[name = tensor<string, []>("x_513_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_513_split_axis_0 = const()[name = tensor<string, []>("x_513_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_513_split_0, tensor<fp32, [1, 1024, 2]> x_513_split_1 = split(axis = x_513_split_axis_0, num_splits = x_513_split_num_splits_0, x = input_1039)[name = tensor<string, []>("x_513_split")];
+            tensor<fp32, [1, 1024, 2]> x_513_split_1_sigmoid = sigmoid(x = x_513_split_1)[name = tensor<string, []>("x_513_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_513 = mul(x = x_513_split_0, y = x_513_split_1_sigmoid)[name = tensor<string, []>("x_513")];
+            tensor<fp32, [1, 1024, 2]> input_1041 = select(a = var_40, b = x_513, cond = var_565)[name = tensor<string, []>("input_1041")];
+            tensor<bool, []> new_x_79_interleave_0 = const()[name = tensor<string, []>("new_x_79_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_79 = concat(axis = var_56, interleave = new_x_79_interleave_0, values = (cache_79, input_1041))[name = tensor<string, []>("new_x_79")];
+            tensor<int32, [3]> var_4454_begin_0 = const()[name = tensor<string, []>("op_4454_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_4454_end_0 = const()[name = tensor<string, []>("op_4454_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_4454_end_mask_0 = const()[name = tensor<string, []>("op_4454_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_4454 = slice_by_index(begin = var_4454_begin_0, end = var_4454_end_0, end_mask = var_4454_end_mask_0, x = new_x_79)[name = tensor<string, []>("op_4454")];
+            tensor<string, []> x_515_pad_type_0 = const()[name = tensor<string, []>("x_515_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_515_groups_0 = const()[name = tensor<string, []>("x_515_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_515_strides_0 = const()[name = tensor<string, []>("x_515_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_515_pad_0 = const()[name = tensor<string, []>("x_515_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_515_dilations_0 = const()[name = tensor<string, []>("x_515_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_515 = conv(dilations = x_515_dilations_0, groups = x_515_groups_0, pad = x_515_pad_0, pad_type = x_515_pad_type_0, strides = x_515_strides_0, weight = module_layers_19_conv_depthwise_conv_weight_quantized, x = new_x_79)[name = tensor<string, []>("x_515")];
+            tensor<int32, [3]> input_1043_perm_0 = const()[name = tensor<string, []>("input_1043_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_517_axes_0 = const()[name = tensor<string, []>("x_517_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1043 = transpose(perm = input_1043_perm_0, x = x_515)[name = tensor<string, []>("transpose_185")];
+            tensor<fp32, [1, 2, 1024]> x_517 = layer_norm(axes = x_517_axes_0, beta = module_layers_19_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_19_conv_batch_norm_weight, x = input_1043)[name = tensor<string, []>("x_517")];
+            tensor<int32, [3]> input_1045_perm_0 = const()[name = tensor<string, []>("input_1045_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_1045 = transpose(perm = input_1045_perm_0, x = x_517)[name = tensor<string, []>("transpose_184")];
+            tensor<fp32, [1, 1024, 2]> input_1047 = silu(x = input_1045)[name = tensor<string, []>("input_1047")];
+            tensor<string, []> x_519_pad_type_0 = const()[name = tensor<string, []>("x_519_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_519_strides_0 = const()[name = tensor<string, []>("x_519_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_519_pad_0 = const()[name = tensor<string, []>("x_519_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_519_dilations_0 = const()[name = tensor<string, []>("x_519_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_519_groups_0 = const()[name = tensor<string, []>("x_519_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_519 = conv(dilations = x_519_dilations_0, groups = x_519_groups_0, pad = x_519_pad_0, pad_type = x_519_pad_type_0, strides = x_519_strides_0, weight = module_layers_19_conv_pointwise_conv2_weight_quantized, x = input_1047)[name = tensor<string, []>("x_519")];
+            tensor<int32, [3]> input_1049_perm_0 = const()[name = tensor<string, []>("input_1049_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_1049 = transpose(perm = input_1049_perm_0, x = x_519)[name = tensor<string, []>("transpose_183")];
+            tensor<fp32, [1, 2, 1024]> input_1051 = add(x = input_1035, y = input_1049)[name = tensor<string, []>("input_1051")];
+            tensor<int32, [1]> input_1053_axes_0 = const()[name = tensor<string, []>("input_1053_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1053 = layer_norm(axes = input_1053_axes_0, beta = module_layers_19_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_19_norm_feed_forward2_weight, x = input_1051)[name = tensor<string, []>("input_1053")];
+            tensor<fp32, [1, 2, 4096]> input_1055 = linear(bias = linear_1_bias_0, weight = module_layers_19_feed_forward2_linear1_weight_quantized, x = input_1053)[name = tensor<string, []>("linear_179")];
+            tensor<fp32, [1, 2, 4096]> input_1057 = silu(x = input_1055)[name = tensor<string, []>("input_1057")];
+            tensor<fp32, [1, 2, 1024]> input_1061 = linear(bias = linear_2_bias_0, weight = module_layers_19_feed_forward2_linear2_weight_quantized, x = input_1057)[name = tensor<string, []>("linear_180")];
+            tensor<fp32, []> var_4495 = const()[name = tensor<string, []>("op_4495"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4496 = mul(x = input_1061, y = var_4495)[name = tensor<string, []>("op_4496")];
+            tensor<fp32, [1, 2, 1024]> input_1063 = add(x = input_1051, y = var_4496)[name = tensor<string, []>("input_1063")];
+            tensor<int32, [1]> input_1065_axes_0 = const()[name = tensor<string, []>("input_1065_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1065 = layer_norm(axes = input_1065_axes_0, beta = module_layers_19_norm_out_bias, epsilon = var_38, gamma = module_layers_19_norm_out_weight, x = input_1063)[name = tensor<string, []>("input_1065")];
+            tensor<int32, [4]> cache_81_begin_0 = const()[name = tensor<string, []>("cache_81_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_81_end_0 = const()[name = tensor<string, []>("cache_81_end_0"), val = tensor<int32, [4]>([21, 1, 70, 1024])];
+            tensor<bool, [4]> cache_81_end_mask_0 = const()[name = tensor<string, []>("cache_81_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_81_squeeze_mask_0 = const()[name = tensor<string, []>("cache_81_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_81 = slice_by_index(begin = cache_81_begin_0, end = cache_81_end_0, end_mask = cache_81_end_mask_0, squeeze_mask = cache_81_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_81")];
+            tensor<int32, [4]> cache_83_begin_0 = const()[name = tensor<string, []>("cache_83_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
+            tensor<int32, [4]> cache_83_end_0 = const()[name = tensor<string, []>("cache_83_end_0"), val = tensor<int32, [4]>([21, 1, 1024, 8])];
+            tensor<bool, [4]> cache_83_end_mask_0 = const()[name = tensor<string, []>("cache_83_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_83_squeeze_mask_0 = const()[name = tensor<string, []>("cache_83_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_83 = slice_by_index(begin = cache_83_begin_0, end = cache_83_end_0, end_mask = cache_83_end_mask_0, squeeze_mask = cache_83_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_83")];
+            tensor<int32, [1]> input_1067_axes_0 = const()[name = tensor<string, []>("input_1067_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1067 = layer_norm(axes = input_1067_axes_0, beta = module_layers_20_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_20_norm_feed_forward1_weight, x = input_1065)[name = tensor<string, []>("input_1067")];
+            tensor<fp32, [1, 2, 4096]> input_1069 = linear(bias = linear_1_bias_0, weight = module_layers_20_feed_forward1_linear1_weight_quantized, x = input_1067)[name = tensor<string, []>("linear_181")];
+            tensor<fp32, [1, 2, 4096]> input_1071 = silu(x = input_1069)[name = tensor<string, []>("input_1071")];
+            tensor<fp32, [1, 2, 1024]> input_1075 = linear(bias = linear_2_bias_0, weight = module_layers_20_feed_forward1_linear2_weight_quantized, x = input_1071)[name = tensor<string, []>("linear_182")];
+            tensor<fp32, []> var_4530 = const()[name = tensor<string, []>("op_4530"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4531 = mul(x = input_1075, y = var_4530)[name = tensor<string, []>("op_4531")];
+            tensor<fp32, [1, 2, 1024]> input_1077 = add(x = input_1065, y = var_4531)[name = tensor<string, []>("input_1077")];
+            tensor<int32, [1]> key_41_axes_0 = const()[name = tensor<string, []>("key_41_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_41 = layer_norm(axes = key_41_axes_0, beta = module_layers_20_norm_self_att_bias, epsilon = var_38, gamma = module_layers_20_norm_self_att_weight, x = input_1077)[name = tensor<string, []>("key_41")];
+            tensor<bool, []> input_1079_interleave_0 = const()[name = tensor<string, []>("input_1079_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_1079 = concat(axis = var_65, interleave = input_1079_interleave_0, values = (cache_81, key_41))[name = tensor<string, []>("input_1079")];
+            tensor<int32, [3]> var_4553_begin_0 = const()[name = tensor<string, []>("op_4553_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_4553_end_0 = const()[name = tensor<string, []>("op_4553_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_4553_end_mask_0 = const()[name = tensor<string, []>("op_4553_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_4553 = slice_by_index(begin = var_4553_begin_0, end = var_4553_end_0, end_mask = var_4553_end_mask_0, x = cache_81)[name = tensor<string, []>("op_4553")];
+            tensor<bool, []> var_4559_interleave_0 = const()[name = tensor<string, []>("op_4559_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_4559 = concat(axis = var_65, interleave = var_4559_interleave_0, values = (var_4553, key_41))[name = tensor<string, []>("op_4559")];
+            tensor<fp32, [1, 2, 1024]> var_4562 = linear(bias = linear_2_bias_0, weight = module_layers_20_self_attn_linear_q_weight_quantized, x = key_41)[name = tensor<string, []>("linear_183")];
+            tensor<int32, [4]> var_4563 = const()[name = tensor<string, []>("op_4563"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_121 = reshape(shape = var_4563, x = var_4562)[name = tensor<string, []>("q_121")];
+            tensor<fp32, [1, 72, 1024]> var_4566 = linear(bias = linear_2_bias_0, weight = module_layers_20_self_attn_linear_k_weight_quantized, x = input_1079)[name = tensor<string, []>("linear_184")];
+            tensor<int32, [4]> var_4567 = const()[name = tensor<string, []>("op_4567"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_81 = reshape(shape = var_4567, x = var_4566)[name = tensor<string, []>("k_81")];
+            tensor<fp32, [1, 72, 1024]> var_4570 = linear(bias = linear_2_bias_0, weight = module_layers_20_self_attn_linear_v_weight_quantized, x = input_1079)[name = tensor<string, []>("linear_185")];
+            tensor<int32, [4]> var_4571 = const()[name = tensor<string, []>("op_4571"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_41 = reshape(shape = var_4571, x = var_4570)[name = tensor<string, []>("v_41")];
+            tensor<int32, [4]> value_49_perm_0 = const()[name = tensor<string, []>("value_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_4583 = add(x = q_121, y = module_layers_20_self_attn_pos_bias_u)[name = tensor<string, []>("op_4583")];
+            tensor<fp32, [1, 2, 8, 128]> var_4585 = add(x = q_121, y = module_layers_20_self_attn_pos_bias_v)[name = tensor<string, []>("op_4585")];
+            tensor<int32, [4]> q_with_bias_v_41_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_4587_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4587_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590281024))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590427776))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590427520)))];
+            tensor<bool, []> x_527_transpose_x_0 = const()[name = tensor<string, []>("x_527_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_527_transpose_y_0 = const()[name = tensor<string, []>("x_527_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_41 = transpose(perm = q_with_bias_v_41_perm_0, x = var_4585)[name = tensor<string, []>("transpose_182")];
+            tensor<fp32, [1, 8, 2, 143]> x_527 = matmul(transpose_x = x_527_transpose_x_0, transpose_y = x_527_transpose_y_0, x = q_with_bias_v_41, y = op_4587_quantized)[name = tensor<string, []>("x_527")];
+            tensor<fp32, []> const_339 = const()[name = tensor<string, []>("const_339"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_529_pad_0 = const()[name = tensor<string, []>("x_529_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_529_mode_0 = const()[name = tensor<string, []>("x_529_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_529 = pad(constant_val = const_339, mode = x_529_mode_0, pad = x_529_pad_0, x = x_527)[name = tensor<string, []>("x_529")];
+            tensor<int32, [4]> var_4595 = const()[name = tensor<string, []>("op_4595"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_531 = reshape(shape = var_4595, x = x_529)[name = tensor<string, []>("x_531")];
+            tensor<int32, [4]> var_4599_begin_0 = const()[name = tensor<string, []>("op_4599_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4599_end_0 = const()[name = tensor<string, []>("op_4599_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_4599_end_mask_0 = const()[name = tensor<string, []>("op_4599_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_4599 = slice_by_index(begin = var_4599_begin_0, end = var_4599_end_0, end_mask = var_4599_end_mask_0, x = x_531)[name = tensor<string, []>("op_4599")];
+            tensor<int32, [4]> var_4600 = const()[name = tensor<string, []>("op_4600"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_81 = reshape(shape = var_4600, x = var_4599)[name = tensor<string, []>("matrix_bd_81")];
+            tensor<bool, []> matrix_ac_41_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_41_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_41_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_41_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_136_perm_0 = const()[name = tensor<string, []>("transpose_136_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_137_perm_0 = const()[name = tensor<string, []>("transpose_137_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_137 = transpose(perm = transpose_137_perm_0, x = k_81)[name = tensor<string, []>("transpose_180")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_136 = transpose(perm = transpose_136_perm_0, x = var_4583)[name = tensor<string, []>("transpose_181")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_41 = matmul(transpose_x = matrix_ac_41_transpose_x_0, transpose_y = matrix_ac_41_transpose_y_0, x = transpose_136, y = transpose_137)[name = tensor<string, []>("matrix_ac_41")];
+            tensor<int32, [4]> matrix_bd_83_begin_0 = const()[name = tensor<string, []>("matrix_bd_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_83_end_0 = const()[name = tensor<string, []>("matrix_bd_83_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_83_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_83 = slice_by_index(begin = matrix_bd_83_begin_0, end = matrix_bd_83_end_0, end_mask = matrix_bd_83_end_mask_0, x = matrix_bd_81)[name = tensor<string, []>("matrix_bd_83")];
+            tensor<fp32, [1, 8, 2, 72]> var_4609 = add(x = matrix_ac_41, y = matrix_bd_83)[name = tensor<string, []>("op_4609")];
+            tensor<fp32, []> _inversed_scores_81_y_0 = const()[name = tensor<string, []>("_inversed_scores_81_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_81 = mul(x = var_4609, y = _inversed_scores_81_y_0)[name = tensor<string, []>("_inversed_scores_81")];
+            tensor<fp32, [1, 8, 2, 72]> scores_83 = select(a = var_41, b = _inversed_scores_81, cond = mask_11)[name = tensor<string, []>("scores_83")];
+            tensor<fp32, [1, 8, 2, 72]> var_4615 = softmax(axis = var_56, x = scores_83)[name = tensor<string, []>("op_4615")];
+            tensor<fp32, [1, 8, 2, 72]> input_1081 = select(a = var_40, b = var_4615, cond = mask_11)[name = tensor<string, []>("input_1081")];
+            tensor<bool, []> x_533_transpose_x_0 = const()[name = tensor<string, []>("x_533_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_533_transpose_y_0 = const()[name = tensor<string, []>("x_533_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_49 = transpose(perm = value_49_perm_0, x = v_41)[name = tensor<string, []>("transpose_179")];
+            tensor<fp32, [1, 8, 2, 128]> x_533 = matmul(transpose_x = x_533_transpose_x_0, transpose_y = x_533_transpose_y_0, x = input_1081, y = value_49)[name = tensor<string, []>("x_533")];
+            tensor<int32, [4]> var_4619_perm_0 = const()[name = tensor<string, []>("op_4619_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4620 = const()[name = tensor<string, []>("op_4620"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_4619 = transpose(perm = var_4619_perm_0, x = x_533)[name = tensor<string, []>("transpose_178")];
+            tensor<fp32, [1, 2, 1024]> input_1083 = reshape(shape = var_4620, x = var_4619)[name = tensor<string, []>("input_1083")];
+            tensor<fp32, [1, 2, 1024]> input_1085 = linear(bias = linear_2_bias_0, weight = module_layers_20_self_attn_linear_out_weight_quantized, x = input_1083)[name = tensor<string, []>("linear_187")];
+            tensor<fp32, [1, 2, 1024]> input_1087 = add(x = input_1077, y = input_1085)[name = tensor<string, []>("input_1087")];
+            tensor<int32, [1]> x_537_axes_0 = const()[name = tensor<string, []>("x_537_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_537 = layer_norm(axes = x_537_axes_0, beta = module_layers_20_norm_conv_bias, epsilon = var_38, gamma = module_layers_20_norm_conv_weight, x = input_1087)[name = tensor<string, []>("x_537")];
+            tensor<int32, [3]> input_1089_perm_0 = const()[name = tensor<string, []>("input_1089_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1091_pad_type_0 = const()[name = tensor<string, []>("input_1091_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1091_strides_0 = const()[name = tensor<string, []>("input_1091_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1091_pad_0 = const()[name = tensor<string, []>("input_1091_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1091_dilations_0 = const()[name = tensor<string, []>("input_1091_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1091_groups_0 = const()[name = tensor<string, []>("input_1091_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_1089 = transpose(perm = input_1089_perm_0, x = x_537)[name = tensor<string, []>("transpose_177")];
+            tensor<fp32, [1, 2048, 2]> input_1091 = conv(dilations = input_1091_dilations_0, groups = input_1091_groups_0, pad = input_1091_pad_0, pad_type = input_1091_pad_type_0, strides = input_1091_strides_0, weight = module_layers_20_conv_pointwise_conv1_weight_quantized, x = input_1089)[name = tensor<string, []>("input_1091")];
+            tensor<int32, []> x_539_split_num_splits_0 = const()[name = tensor<string, []>("x_539_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_539_split_axis_0 = const()[name = tensor<string, []>("x_539_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_539_split_0, tensor<fp32, [1, 1024, 2]> x_539_split_1 = split(axis = x_539_split_axis_0, num_splits = x_539_split_num_splits_0, x = input_1091)[name = tensor<string, []>("x_539_split")];
+            tensor<fp32, [1, 1024, 2]> x_539_split_1_sigmoid = sigmoid(x = x_539_split_1)[name = tensor<string, []>("x_539_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_539 = mul(x = x_539_split_0, y = x_539_split_1_sigmoid)[name = tensor<string, []>("x_539")];
+            tensor<fp32, [1, 1024, 2]> input_1093 = select(a = var_40, b = x_539, cond = var_565)[name = tensor<string, []>("input_1093")];
+            tensor<bool, []> new_x_83_interleave_0 = const()[name = tensor<string, []>("new_x_83_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_83 = concat(axis = var_56, interleave = new_x_83_interleave_0, values = (cache_83, input_1093))[name = tensor<string, []>("new_x_83")];
+            tensor<int32, [3]> var_4658_begin_0 = const()[name = tensor<string, []>("op_4658_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_4658_end_0 = const()[name = tensor<string, []>("op_4658_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_4658_end_mask_0 = const()[name = tensor<string, []>("op_4658_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_4658 = slice_by_index(begin = var_4658_begin_0, end = var_4658_end_0, end_mask = var_4658_end_mask_0, x = new_x_83)[name = tensor<string, []>("op_4658")];
+            tensor<string, []> x_541_pad_type_0 = const()[name = tensor<string, []>("x_541_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_541_groups_0 = const()[name = tensor<string, []>("x_541_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_541_strides_0 = const()[name = tensor<string, []>("x_541_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_541_pad_0 = const()[name = tensor<string, []>("x_541_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_541_dilations_0 = const()[name = tensor<string, []>("x_541_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_541 = conv(dilations = x_541_dilations_0, groups = x_541_groups_0, pad = x_541_pad_0, pad_type = x_541_pad_type_0, strides = x_541_strides_0, weight = module_layers_20_conv_depthwise_conv_weight_quantized, x = new_x_83)[name = tensor<string, []>("x_541")];
+            tensor<int32, [3]> input_1095_perm_0 = const()[name = tensor<string, []>("input_1095_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_543_axes_0 = const()[name = tensor<string, []>("x_543_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1095 = transpose(perm = input_1095_perm_0, x = x_541)[name = tensor<string, []>("transpose_176")];
+            tensor<fp32, [1, 2, 1024]> x_543 = layer_norm(axes = x_543_axes_0, beta = module_layers_20_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_20_conv_batch_norm_weight, x = input_1095)[name = tensor<string, []>("x_543")];
+            tensor<int32, [3]> input_1097_perm_0 = const()[name = tensor<string, []>("input_1097_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_1097 = transpose(perm = input_1097_perm_0, x = x_543)[name = tensor<string, []>("transpose_175")];
+            tensor<fp32, [1, 1024, 2]> input_1099 = silu(x = input_1097)[name = tensor<string, []>("input_1099")];
+            tensor<string, []> x_545_pad_type_0 = const()[name = tensor<string, []>("x_545_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_545_strides_0 = const()[name = tensor<string, []>("x_545_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_545_pad_0 = const()[name = tensor<string, []>("x_545_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_545_dilations_0 = const()[name = tensor<string, []>("x_545_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_545_groups_0 = const()[name = tensor<string, []>("x_545_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_545 = conv(dilations = x_545_dilations_0, groups = x_545_groups_0, pad = x_545_pad_0, pad_type = x_545_pad_type_0, strides = x_545_strides_0, weight = module_layers_20_conv_pointwise_conv2_weight_quantized, x = input_1099)[name = tensor<string, []>("x_545")];
+            tensor<int32, [3]> input_1101_perm_0 = const()[name = tensor<string, []>("input_1101_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_1101 = transpose(perm = input_1101_perm_0, x = x_545)[name = tensor<string, []>("transpose_174")];
+            tensor<fp32, [1, 2, 1024]> input_1103 = add(x = input_1087, y = input_1101)[name = tensor<string, []>("input_1103")];
+            tensor<int32, [1]> input_1105_axes_0 = const()[name = tensor<string, []>("input_1105_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1105 = layer_norm(axes = input_1105_axes_0, beta = module_layers_20_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_20_norm_feed_forward2_weight, x = input_1103)[name = tensor<string, []>("input_1105")];
+            tensor<fp32, [1, 2, 4096]> input_1107 = linear(bias = linear_1_bias_0, weight = module_layers_20_feed_forward2_linear1_weight_quantized, x = input_1105)[name = tensor<string, []>("linear_188")];
+            tensor<fp32, [1, 2, 4096]> input_1109 = silu(x = input_1107)[name = tensor<string, []>("input_1109")];
+            tensor<fp32, [1, 2, 1024]> input_1113 = linear(bias = linear_2_bias_0, weight = module_layers_20_feed_forward2_linear2_weight_quantized, x = input_1109)[name = tensor<string, []>("linear_189")];
+            tensor<fp32, []> var_4699 = const()[name = tensor<string, []>("op_4699"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4700 = mul(x = input_1113, y = var_4699)[name = tensor<string, []>("op_4700")];
+            tensor<fp32, [1, 2, 1024]> input_1115 = add(x = input_1103, y = var_4700)[name = tensor<string, []>("input_1115")];
+            tensor<int32, [1]> input_1117_axes_0 = const()[name = tensor<string, []>("input_1117_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1117 = layer_norm(axes = input_1117_axes_0, beta = module_layers_20_norm_out_bias, epsilon = var_38, gamma = module_layers_20_norm_out_weight, x = input_1115)[name = tensor<string, []>("input_1117")];
+            tensor<int32, [4]> cache_85_begin_0 = const()[name = tensor<string, []>("cache_85_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_85_end_0 = const()[name = tensor<string, []>("cache_85_end_0"), val = tensor<int32, [4]>([22, 1, 70, 1024])];
+            tensor<bool, [4]> cache_85_end_mask_0 = const()[name = tensor<string, []>("cache_85_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_85_squeeze_mask_0 = const()[name = tensor<string, []>("cache_85_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_85 = slice_by_index(begin = cache_85_begin_0, end = cache_85_end_0, end_mask = cache_85_end_mask_0, squeeze_mask = cache_85_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_85")];
+            tensor<int32, [4]> cache_87_begin_0 = const()[name = tensor<string, []>("cache_87_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
+            tensor<int32, [4]> cache_87_end_0 = const()[name = tensor<string, []>("cache_87_end_0"), val = tensor<int32, [4]>([22, 1, 1024, 8])];
+            tensor<bool, [4]> cache_87_end_mask_0 = const()[name = tensor<string, []>("cache_87_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_87_squeeze_mask_0 = const()[name = tensor<string, []>("cache_87_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_87 = slice_by_index(begin = cache_87_begin_0, end = cache_87_end_0, end_mask = cache_87_end_mask_0, squeeze_mask = cache_87_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_87")];
+            tensor<int32, [1]> input_1119_axes_0 = const()[name = tensor<string, []>("input_1119_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1119 = layer_norm(axes = input_1119_axes_0, beta = module_layers_21_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_21_norm_feed_forward1_weight, x = input_1117)[name = tensor<string, []>("input_1119")];
+            tensor<fp32, [1, 2, 4096]> input_1121 = linear(bias = linear_1_bias_0, weight = module_layers_21_feed_forward1_linear1_weight_quantized, x = input_1119)[name = tensor<string, []>("linear_190")];
+            tensor<fp32, [1, 2, 4096]> input_1123 = silu(x = input_1121)[name = tensor<string, []>("input_1123")];
+            tensor<fp32, [1, 2, 1024]> input_1127 = linear(bias = linear_2_bias_0, weight = module_layers_21_feed_forward1_linear2_weight_quantized, x = input_1123)[name = tensor<string, []>("linear_191")];
+            tensor<fp32, []> var_4734 = const()[name = tensor<string, []>("op_4734"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4735 = mul(x = input_1127, y = var_4734)[name = tensor<string, []>("op_4735")];
+            tensor<fp32, [1, 2, 1024]> input_1129 = add(x = input_1117, y = var_4735)[name = tensor<string, []>("input_1129")];
+            tensor<int32, [1]> key_43_axes_0 = const()[name = tensor<string, []>("key_43_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_43 = layer_norm(axes = key_43_axes_0, beta = module_layers_21_norm_self_att_bias, epsilon = var_38, gamma = module_layers_21_norm_self_att_weight, x = input_1129)[name = tensor<string, []>("key_43")];
+            tensor<bool, []> input_1131_interleave_0 = const()[name = tensor<string, []>("input_1131_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_1131 = concat(axis = var_65, interleave = input_1131_interleave_0, values = (cache_85, key_43))[name = tensor<string, []>("input_1131")];
+            tensor<int32, [3]> var_4757_begin_0 = const()[name = tensor<string, []>("op_4757_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_4757_end_0 = const()[name = tensor<string, []>("op_4757_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_4757_end_mask_0 = const()[name = tensor<string, []>("op_4757_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_4757 = slice_by_index(begin = var_4757_begin_0, end = var_4757_end_0, end_mask = var_4757_end_mask_0, x = cache_85)[name = tensor<string, []>("op_4757")];
+            tensor<bool, []> var_4763_interleave_0 = const()[name = tensor<string, []>("op_4763_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_4763 = concat(axis = var_65, interleave = var_4763_interleave_0, values = (var_4757, key_43))[name = tensor<string, []>("op_4763")];
+            tensor<fp32, [1, 2, 1024]> var_4766 = linear(bias = linear_2_bias_0, weight = module_layers_21_self_attn_linear_q_weight_quantized, x = key_43)[name = tensor<string, []>("linear_192")];
+            tensor<int32, [4]> var_4767 = const()[name = tensor<string, []>("op_4767"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_127 = reshape(shape = var_4767, x = var_4766)[name = tensor<string, []>("q_127")];
+            tensor<fp32, [1, 72, 1024]> var_4770 = linear(bias = linear_2_bias_0, weight = module_layers_21_self_attn_linear_k_weight_quantized, x = input_1131)[name = tensor<string, []>("linear_193")];
+            tensor<int32, [4]> var_4771 = const()[name = tensor<string, []>("op_4771"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_85 = reshape(shape = var_4771, x = var_4770)[name = tensor<string, []>("k_85")];
+            tensor<fp32, [1, 72, 1024]> var_4774 = linear(bias = linear_2_bias_0, weight = module_layers_21_self_attn_linear_v_weight_quantized, x = input_1131)[name = tensor<string, []>("linear_194")];
+            tensor<int32, [4]> var_4775 = const()[name = tensor<string, []>("op_4775"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_43 = reshape(shape = var_4775, x = var_4774)[name = tensor<string, []>("v_43")];
+            tensor<int32, [4]> value_51_perm_0 = const()[name = tensor<string, []>("value_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_4787 = add(x = q_127, y = module_layers_21_self_attn_pos_bias_u)[name = tensor<string, []>("op_4787")];
+            tensor<fp32, [1, 2, 8, 128]> var_4789 = add(x = q_127, y = module_layers_21_self_attn_pos_bias_v)[name = tensor<string, []>("op_4789")];
+            tensor<int32, [4]> q_with_bias_v_43_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_4791_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4791_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590428416))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590575168))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590574912)))];
+            tensor<bool, []> x_553_transpose_x_0 = const()[name = tensor<string, []>("x_553_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_553_transpose_y_0 = const()[name = tensor<string, []>("x_553_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_43 = transpose(perm = q_with_bias_v_43_perm_0, x = var_4789)[name = tensor<string, []>("transpose_173")];
+            tensor<fp32, [1, 8, 2, 143]> x_553 = matmul(transpose_x = x_553_transpose_x_0, transpose_y = x_553_transpose_y_0, x = q_with_bias_v_43, y = op_4791_quantized)[name = tensor<string, []>("x_553")];
+            tensor<fp32, []> const_352 = const()[name = tensor<string, []>("const_352"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_555_pad_0 = const()[name = tensor<string, []>("x_555_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_555_mode_0 = const()[name = tensor<string, []>("x_555_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_555 = pad(constant_val = const_352, mode = x_555_mode_0, pad = x_555_pad_0, x = x_553)[name = tensor<string, []>("x_555")];
+            tensor<int32, [4]> var_4799 = const()[name = tensor<string, []>("op_4799"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_557 = reshape(shape = var_4799, x = x_555)[name = tensor<string, []>("x_557")];
+            tensor<int32, [4]> var_4803_begin_0 = const()[name = tensor<string, []>("op_4803_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_4803_end_0 = const()[name = tensor<string, []>("op_4803_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_4803_end_mask_0 = const()[name = tensor<string, []>("op_4803_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_4803 = slice_by_index(begin = var_4803_begin_0, end = var_4803_end_0, end_mask = var_4803_end_mask_0, x = x_557)[name = tensor<string, []>("op_4803")];
+            tensor<int32, [4]> var_4804 = const()[name = tensor<string, []>("op_4804"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_85 = reshape(shape = var_4804, x = var_4803)[name = tensor<string, []>("matrix_bd_85")];
+            tensor<bool, []> matrix_ac_43_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_43_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_43_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_43_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_138_perm_0 = const()[name = tensor<string, []>("transpose_138_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_139_perm_0 = const()[name = tensor<string, []>("transpose_139_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_139 = transpose(perm = transpose_139_perm_0, x = k_85)[name = tensor<string, []>("transpose_171")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_138 = transpose(perm = transpose_138_perm_0, x = var_4787)[name = tensor<string, []>("transpose_172")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_43 = matmul(transpose_x = matrix_ac_43_transpose_x_0, transpose_y = matrix_ac_43_transpose_y_0, x = transpose_138, y = transpose_139)[name = tensor<string, []>("matrix_ac_43")];
+            tensor<int32, [4]> matrix_bd_87_begin_0 = const()[name = tensor<string, []>("matrix_bd_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_87_end_0 = const()[name = tensor<string, []>("matrix_bd_87_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_87_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_87 = slice_by_index(begin = matrix_bd_87_begin_0, end = matrix_bd_87_end_0, end_mask = matrix_bd_87_end_mask_0, x = matrix_bd_85)[name = tensor<string, []>("matrix_bd_87")];
+            tensor<fp32, [1, 8, 2, 72]> var_4813 = add(x = matrix_ac_43, y = matrix_bd_87)[name = tensor<string, []>("op_4813")];
+            tensor<fp32, []> _inversed_scores_85_y_0 = const()[name = tensor<string, []>("_inversed_scores_85_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_85 = mul(x = var_4813, y = _inversed_scores_85_y_0)[name = tensor<string, []>("_inversed_scores_85")];
+            tensor<fp32, [1, 8, 2, 72]> scores_87 = select(a = var_41, b = _inversed_scores_85, cond = mask_11)[name = tensor<string, []>("scores_87")];
+            tensor<fp32, [1, 8, 2, 72]> var_4819 = softmax(axis = var_56, x = scores_87)[name = tensor<string, []>("op_4819")];
+            tensor<fp32, [1, 8, 2, 72]> input_1133 = select(a = var_40, b = var_4819, cond = mask_11)[name = tensor<string, []>("input_1133")];
+            tensor<bool, []> x_559_transpose_x_0 = const()[name = tensor<string, []>("x_559_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_559_transpose_y_0 = const()[name = tensor<string, []>("x_559_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_51 = transpose(perm = value_51_perm_0, x = v_43)[name = tensor<string, []>("transpose_170")];
+            tensor<fp32, [1, 8, 2, 128]> x_559 = matmul(transpose_x = x_559_transpose_x_0, transpose_y = x_559_transpose_y_0, x = input_1133, y = value_51)[name = tensor<string, []>("x_559")];
+            tensor<int32, [4]> var_4823_perm_0 = const()[name = tensor<string, []>("op_4823_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_4824 = const()[name = tensor<string, []>("op_4824"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_4823 = transpose(perm = var_4823_perm_0, x = x_559)[name = tensor<string, []>("transpose_169")];
+            tensor<fp32, [1, 2, 1024]> input_1135 = reshape(shape = var_4824, x = var_4823)[name = tensor<string, []>("input_1135")];
+            tensor<fp32, [1, 2, 1024]> input_1137 = linear(bias = linear_2_bias_0, weight = module_layers_21_self_attn_linear_out_weight_quantized, x = input_1135)[name = tensor<string, []>("linear_196")];
+            tensor<fp32, [1, 2, 1024]> input_1139 = add(x = input_1129, y = input_1137)[name = tensor<string, []>("input_1139")];
+            tensor<int32, [1]> x_563_axes_0 = const()[name = tensor<string, []>("x_563_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_563 = layer_norm(axes = x_563_axes_0, beta = module_layers_21_norm_conv_bias, epsilon = var_38, gamma = module_layers_21_norm_conv_weight, x = input_1139)[name = tensor<string, []>("x_563")];
+            tensor<int32, [3]> input_1141_perm_0 = const()[name = tensor<string, []>("input_1141_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1143_pad_type_0 = const()[name = tensor<string, []>("input_1143_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1143_strides_0 = const()[name = tensor<string, []>("input_1143_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1143_pad_0 = const()[name = tensor<string, []>("input_1143_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1143_dilations_0 = const()[name = tensor<string, []>("input_1143_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1143_groups_0 = const()[name = tensor<string, []>("input_1143_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_1141 = transpose(perm = input_1141_perm_0, x = x_563)[name = tensor<string, []>("transpose_168")];
+            tensor<fp32, [1, 2048, 2]> input_1143 = conv(dilations = input_1143_dilations_0, groups = input_1143_groups_0, pad = input_1143_pad_0, pad_type = input_1143_pad_type_0, strides = input_1143_strides_0, weight = module_layers_21_conv_pointwise_conv1_weight_quantized, x = input_1141)[name = tensor<string, []>("input_1143")];
+            tensor<int32, []> x_565_split_num_splits_0 = const()[name = tensor<string, []>("x_565_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_565_split_axis_0 = const()[name = tensor<string, []>("x_565_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_565_split_0, tensor<fp32, [1, 1024, 2]> x_565_split_1 = split(axis = x_565_split_axis_0, num_splits = x_565_split_num_splits_0, x = input_1143)[name = tensor<string, []>("x_565_split")];
+            tensor<fp32, [1, 1024, 2]> x_565_split_1_sigmoid = sigmoid(x = x_565_split_1)[name = tensor<string, []>("x_565_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_565 = mul(x = x_565_split_0, y = x_565_split_1_sigmoid)[name = tensor<string, []>("x_565")];
+            tensor<fp32, [1, 1024, 2]> input_1145 = select(a = var_40, b = x_565, cond = var_565)[name = tensor<string, []>("input_1145")];
+            tensor<bool, []> new_x_87_interleave_0 = const()[name = tensor<string, []>("new_x_87_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_87 = concat(axis = var_56, interleave = new_x_87_interleave_0, values = (cache_87, input_1145))[name = tensor<string, []>("new_x_87")];
+            tensor<int32, [3]> var_4862_begin_0 = const()[name = tensor<string, []>("op_4862_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_4862_end_0 = const()[name = tensor<string, []>("op_4862_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_4862_end_mask_0 = const()[name = tensor<string, []>("op_4862_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_4862 = slice_by_index(begin = var_4862_begin_0, end = var_4862_end_0, end_mask = var_4862_end_mask_0, x = new_x_87)[name = tensor<string, []>("op_4862")];
+            tensor<string, []> x_567_pad_type_0 = const()[name = tensor<string, []>("x_567_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_567_groups_0 = const()[name = tensor<string, []>("x_567_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_567_strides_0 = const()[name = tensor<string, []>("x_567_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_567_pad_0 = const()[name = tensor<string, []>("x_567_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_567_dilations_0 = const()[name = tensor<string, []>("x_567_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_567 = conv(dilations = x_567_dilations_0, groups = x_567_groups_0, pad = x_567_pad_0, pad_type = x_567_pad_type_0, strides = x_567_strides_0, weight = module_layers_21_conv_depthwise_conv_weight_quantized, x = new_x_87)[name = tensor<string, []>("x_567")];
+            tensor<int32, [3]> input_1147_perm_0 = const()[name = tensor<string, []>("input_1147_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_569_axes_0 = const()[name = tensor<string, []>("x_569_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1147 = transpose(perm = input_1147_perm_0, x = x_567)[name = tensor<string, []>("transpose_167")];
+            tensor<fp32, [1, 2, 1024]> x_569 = layer_norm(axes = x_569_axes_0, beta = module_layers_21_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_21_conv_batch_norm_weight, x = input_1147)[name = tensor<string, []>("x_569")];
+            tensor<int32, [3]> input_1149_perm_0 = const()[name = tensor<string, []>("input_1149_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_1149 = transpose(perm = input_1149_perm_0, x = x_569)[name = tensor<string, []>("transpose_166")];
+            tensor<fp32, [1, 1024, 2]> input_1151 = silu(x = input_1149)[name = tensor<string, []>("input_1151")];
+            tensor<string, []> x_571_pad_type_0 = const()[name = tensor<string, []>("x_571_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_571_strides_0 = const()[name = tensor<string, []>("x_571_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_571_pad_0 = const()[name = tensor<string, []>("x_571_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_571_dilations_0 = const()[name = tensor<string, []>("x_571_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_571_groups_0 = const()[name = tensor<string, []>("x_571_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_571 = conv(dilations = x_571_dilations_0, groups = x_571_groups_0, pad = x_571_pad_0, pad_type = x_571_pad_type_0, strides = x_571_strides_0, weight = module_layers_21_conv_pointwise_conv2_weight_quantized, x = input_1151)[name = tensor<string, []>("x_571")];
+            tensor<int32, [3]> input_1153_perm_0 = const()[name = tensor<string, []>("input_1153_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_1153 = transpose(perm = input_1153_perm_0, x = x_571)[name = tensor<string, []>("transpose_165")];
+            tensor<fp32, [1, 2, 1024]> input_1155 = add(x = input_1139, y = input_1153)[name = tensor<string, []>("input_1155")];
+            tensor<int32, [1]> input_1157_axes_0 = const()[name = tensor<string, []>("input_1157_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1157 = layer_norm(axes = input_1157_axes_0, beta = module_layers_21_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_21_norm_feed_forward2_weight, x = input_1155)[name = tensor<string, []>("input_1157")];
+            tensor<fp32, [1, 2, 4096]> input_1159 = linear(bias = linear_1_bias_0, weight = module_layers_21_feed_forward2_linear1_weight_quantized, x = input_1157)[name = tensor<string, []>("linear_197")];
+            tensor<fp32, [1, 2, 4096]> input_1161 = silu(x = input_1159)[name = tensor<string, []>("input_1161")];
+            tensor<fp32, [1, 2, 1024]> input_1165 = linear(bias = linear_2_bias_0, weight = module_layers_21_feed_forward2_linear2_weight_quantized, x = input_1161)[name = tensor<string, []>("linear_198")];
+            tensor<fp32, []> var_4903 = const()[name = tensor<string, []>("op_4903"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4904 = mul(x = input_1165, y = var_4903)[name = tensor<string, []>("op_4904")];
+            tensor<fp32, [1, 2, 1024]> input_1167 = add(x = input_1155, y = var_4904)[name = tensor<string, []>("input_1167")];
+            tensor<int32, [1]> input_1169_axes_0 = const()[name = tensor<string, []>("input_1169_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1169 = layer_norm(axes = input_1169_axes_0, beta = module_layers_21_norm_out_bias, epsilon = var_38, gamma = module_layers_21_norm_out_weight, x = input_1167)[name = tensor<string, []>("input_1169")];
+            tensor<int32, [4]> cache_89_begin_0 = const()[name = tensor<string, []>("cache_89_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_89_end_0 = const()[name = tensor<string, []>("cache_89_end_0"), val = tensor<int32, [4]>([23, 1, 70, 1024])];
+            tensor<bool, [4]> cache_89_end_mask_0 = const()[name = tensor<string, []>("cache_89_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_89_squeeze_mask_0 = const()[name = tensor<string, []>("cache_89_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_89 = slice_by_index(begin = cache_89_begin_0, end = cache_89_end_0, end_mask = cache_89_end_mask_0, squeeze_mask = cache_89_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_89")];
+            tensor<int32, [4]> cache_91_begin_0 = const()[name = tensor<string, []>("cache_91_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
+            tensor<int32, [4]> cache_91_end_0 = const()[name = tensor<string, []>("cache_91_end_0"), val = tensor<int32, [4]>([23, 1, 1024, 8])];
+            tensor<bool, [4]> cache_91_end_mask_0 = const()[name = tensor<string, []>("cache_91_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_91_squeeze_mask_0 = const()[name = tensor<string, []>("cache_91_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache_91 = slice_by_index(begin = cache_91_begin_0, end = cache_91_end_0, end_mask = cache_91_end_mask_0, squeeze_mask = cache_91_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache_91")];
+            tensor<int32, [1]> input_1171_axes_0 = const()[name = tensor<string, []>("input_1171_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1171 = layer_norm(axes = input_1171_axes_0, beta = module_layers_22_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_22_norm_feed_forward1_weight, x = input_1169)[name = tensor<string, []>("input_1171")];
+            tensor<fp32, [1, 2, 4096]> input_1173 = linear(bias = linear_1_bias_0, weight = module_layers_22_feed_forward1_linear1_weight_quantized, x = input_1171)[name = tensor<string, []>("linear_199")];
+            tensor<fp32, [1, 2, 4096]> input_1175 = silu(x = input_1173)[name = tensor<string, []>("input_1175")];
+            tensor<fp32, [1, 2, 1024]> input_1179 = linear(bias = linear_2_bias_0, weight = module_layers_22_feed_forward1_linear2_weight_quantized, x = input_1175)[name = tensor<string, []>("linear_200")];
+            tensor<fp32, []> var_4938 = const()[name = tensor<string, []>("op_4938"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_4939 = mul(x = input_1179, y = var_4938)[name = tensor<string, []>("op_4939")];
+            tensor<fp32, [1, 2, 1024]> input_1181 = add(x = input_1169, y = var_4939)[name = tensor<string, []>("input_1181")];
+            tensor<int32, [1]> key_45_axes_0 = const()[name = tensor<string, []>("key_45_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key_45 = layer_norm(axes = key_45_axes_0, beta = module_layers_22_norm_self_att_bias, epsilon = var_38, gamma = module_layers_22_norm_self_att_weight, x = input_1181)[name = tensor<string, []>("key_45")];
+            tensor<bool, []> input_1183_interleave_0 = const()[name = tensor<string, []>("input_1183_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_1183 = concat(axis = var_65, interleave = input_1183_interleave_0, values = (cache_89, key_45))[name = tensor<string, []>("input_1183")];
+            tensor<int32, [3]> var_4961_begin_0 = const()[name = tensor<string, []>("op_4961_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_4961_end_0 = const()[name = tensor<string, []>("op_4961_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_4961_end_mask_0 = const()[name = tensor<string, []>("op_4961_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_4961 = slice_by_index(begin = var_4961_begin_0, end = var_4961_end_0, end_mask = var_4961_end_mask_0, x = cache_89)[name = tensor<string, []>("op_4961")];
+            tensor<bool, []> var_4967_interleave_0 = const()[name = tensor<string, []>("op_4967_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> var_4967 = concat(axis = var_65, interleave = var_4967_interleave_0, values = (var_4961, key_45))[name = tensor<string, []>("op_4967")];
+            tensor<fp32, [1, 2, 1024]> var_4970 = linear(bias = linear_2_bias_0, weight = module_layers_22_self_attn_linear_q_weight_quantized, x = key_45)[name = tensor<string, []>("linear_201")];
+            tensor<int32, [4]> var_4971 = const()[name = tensor<string, []>("op_4971"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_133 = reshape(shape = var_4971, x = var_4970)[name = tensor<string, []>("q_133")];
+            tensor<fp32, [1, 72, 1024]> var_4974 = linear(bias = linear_2_bias_0, weight = module_layers_22_self_attn_linear_k_weight_quantized, x = input_1183)[name = tensor<string, []>("linear_202")];
+            tensor<int32, [4]> var_4975 = const()[name = tensor<string, []>("op_4975"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_89 = reshape(shape = var_4975, x = var_4974)[name = tensor<string, []>("k_89")];
+            tensor<fp32, [1, 72, 1024]> var_4978 = linear(bias = linear_2_bias_0, weight = module_layers_22_self_attn_linear_v_weight_quantized, x = input_1183)[name = tensor<string, []>("linear_203")];
+            tensor<int32, [4]> var_4979 = const()[name = tensor<string, []>("op_4979"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v_45 = reshape(shape = var_4979, x = var_4978)[name = tensor<string, []>("v_45")];
+            tensor<int32, [4]> value_53_perm_0 = const()[name = tensor<string, []>("value_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_4991 = add(x = q_133, y = module_layers_22_self_attn_pos_bias_u)[name = tensor<string, []>("op_4991")];
+            tensor<fp32, [1, 2, 8, 128]> var_4993 = add(x = q_133, y = module_layers_22_self_attn_pos_bias_v)[name = tensor<string, []>("op_4993")];
+            tensor<int32, [4]> q_with_bias_v_45_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_4995_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_4995_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590575808))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590722560))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590722304)))];
+            tensor<bool, []> x_579_transpose_x_0 = const()[name = tensor<string, []>("x_579_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_579_transpose_y_0 = const()[name = tensor<string, []>("x_579_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v_45 = transpose(perm = q_with_bias_v_45_perm_0, x = var_4993)[name = tensor<string, []>("transpose_164")];
+            tensor<fp32, [1, 8, 2, 143]> x_579 = matmul(transpose_x = x_579_transpose_x_0, transpose_y = x_579_transpose_y_0, x = q_with_bias_v_45, y = op_4995_quantized)[name = tensor<string, []>("x_579")];
+            tensor<fp32, []> const_365 = const()[name = tensor<string, []>("const_365"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_581_pad_0 = const()[name = tensor<string, []>("x_581_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_581_mode_0 = const()[name = tensor<string, []>("x_581_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_581 = pad(constant_val = const_365, mode = x_581_mode_0, pad = x_581_pad_0, x = x_579)[name = tensor<string, []>("x_581")];
+            tensor<int32, [4]> var_5003 = const()[name = tensor<string, []>("op_5003"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_583 = reshape(shape = var_5003, x = x_581)[name = tensor<string, []>("x_583")];
+            tensor<int32, [4]> var_5007_begin_0 = const()[name = tensor<string, []>("op_5007_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5007_end_0 = const()[name = tensor<string, []>("op_5007_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_5007_end_mask_0 = const()[name = tensor<string, []>("op_5007_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_5007 = slice_by_index(begin = var_5007_begin_0, end = var_5007_end_0, end_mask = var_5007_end_mask_0, x = x_583)[name = tensor<string, []>("op_5007")];
+            tensor<int32, [4]> var_5008 = const()[name = tensor<string, []>("op_5008"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_89 = reshape(shape = var_5008, x = var_5007)[name = tensor<string, []>("matrix_bd_89")];
+            tensor<bool, []> matrix_ac_45_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_45_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_45_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_45_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_140_perm_0 = const()[name = tensor<string, []>("transpose_140_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_141_perm_0 = const()[name = tensor<string, []>("transpose_141_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_141 = transpose(perm = transpose_141_perm_0, x = k_89)[name = tensor<string, []>("transpose_162")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_140 = transpose(perm = transpose_140_perm_0, x = var_4991)[name = tensor<string, []>("transpose_163")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac_45 = matmul(transpose_x = matrix_ac_45_transpose_x_0, transpose_y = matrix_ac_45_transpose_y_0, x = transpose_140, y = transpose_141)[name = tensor<string, []>("matrix_ac_45")];
+            tensor<int32, [4]> matrix_bd_91_begin_0 = const()[name = tensor<string, []>("matrix_bd_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_91_end_0 = const()[name = tensor<string, []>("matrix_bd_91_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_91_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd_91 = slice_by_index(begin = matrix_bd_91_begin_0, end = matrix_bd_91_end_0, end_mask = matrix_bd_91_end_mask_0, x = matrix_bd_89)[name = tensor<string, []>("matrix_bd_91")];
+            tensor<fp32, [1, 8, 2, 72]> var_5017 = add(x = matrix_ac_45, y = matrix_bd_91)[name = tensor<string, []>("op_5017")];
+            tensor<fp32, []> _inversed_scores_89_y_0 = const()[name = tensor<string, []>("_inversed_scores_89_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_89 = mul(x = var_5017, y = _inversed_scores_89_y_0)[name = tensor<string, []>("_inversed_scores_89")];
+            tensor<fp32, [1, 8, 2, 72]> scores_91 = select(a = var_41, b = _inversed_scores_89, cond = mask_11)[name = tensor<string, []>("scores_91")];
+            tensor<fp32, [1, 8, 2, 72]> var_5023 = softmax(axis = var_56, x = scores_91)[name = tensor<string, []>("op_5023")];
+            tensor<fp32, [1, 8, 2, 72]> input_1185 = select(a = var_40, b = var_5023, cond = mask_11)[name = tensor<string, []>("input_1185")];
+            tensor<bool, []> x_585_transpose_x_0 = const()[name = tensor<string, []>("x_585_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_585_transpose_y_0 = const()[name = tensor<string, []>("x_585_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value_53 = transpose(perm = value_53_perm_0, x = v_45)[name = tensor<string, []>("transpose_161")];
+            tensor<fp32, [1, 8, 2, 128]> x_585 = matmul(transpose_x = x_585_transpose_x_0, transpose_y = x_585_transpose_y_0, x = input_1185, y = value_53)[name = tensor<string, []>("x_585")];
+            tensor<int32, [4]> var_5027_perm_0 = const()[name = tensor<string, []>("op_5027_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5028 = const()[name = tensor<string, []>("op_5028"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_5027 = transpose(perm = var_5027_perm_0, x = x_585)[name = tensor<string, []>("transpose_160")];
+            tensor<fp32, [1, 2, 1024]> input_1187 = reshape(shape = var_5028, x = var_5027)[name = tensor<string, []>("input_1187")];
+            tensor<fp32, [1, 2, 1024]> input_1189 = linear(bias = linear_2_bias_0, weight = module_layers_22_self_attn_linear_out_weight_quantized, x = input_1187)[name = tensor<string, []>("linear_205")];
+            tensor<fp32, [1, 2, 1024]> input_1191 = add(x = input_1181, y = input_1189)[name = tensor<string, []>("input_1191")];
+            tensor<int32, [1]> x_589_axes_0 = const()[name = tensor<string, []>("x_589_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_589 = layer_norm(axes = x_589_axes_0, beta = module_layers_22_norm_conv_bias, epsilon = var_38, gamma = module_layers_22_norm_conv_weight, x = input_1191)[name = tensor<string, []>("x_589")];
+            tensor<int32, [3]> input_1193_perm_0 = const()[name = tensor<string, []>("input_1193_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1195_pad_type_0 = const()[name = tensor<string, []>("input_1195_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1195_strides_0 = const()[name = tensor<string, []>("input_1195_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1195_pad_0 = const()[name = tensor<string, []>("input_1195_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1195_dilations_0 = const()[name = tensor<string, []>("input_1195_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1195_groups_0 = const()[name = tensor<string, []>("input_1195_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_1193 = transpose(perm = input_1193_perm_0, x = x_589)[name = tensor<string, []>("transpose_159")];
+            tensor<fp32, [1, 2048, 2]> input_1195 = conv(dilations = input_1195_dilations_0, groups = input_1195_groups_0, pad = input_1195_pad_0, pad_type = input_1195_pad_type_0, strides = input_1195_strides_0, weight = module_layers_22_conv_pointwise_conv1_weight_quantized, x = input_1193)[name = tensor<string, []>("input_1195")];
+            tensor<int32, []> x_591_split_num_splits_0 = const()[name = tensor<string, []>("x_591_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_591_split_axis_0 = const()[name = tensor<string, []>("x_591_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_591_split_0, tensor<fp32, [1, 1024, 2]> x_591_split_1 = split(axis = x_591_split_axis_0, num_splits = x_591_split_num_splits_0, x = input_1195)[name = tensor<string, []>("x_591_split")];
+            tensor<fp32, [1, 1024, 2]> x_591_split_1_sigmoid = sigmoid(x = x_591_split_1)[name = tensor<string, []>("x_591_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_591 = mul(x = x_591_split_0, y = x_591_split_1_sigmoid)[name = tensor<string, []>("x_591")];
+            tensor<fp32, [1, 1024, 2]> input_1197 = select(a = var_40, b = x_591, cond = var_565)[name = tensor<string, []>("input_1197")];
+            tensor<bool, []> new_x_91_interleave_0 = const()[name = tensor<string, []>("new_x_91_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x_91 = concat(axis = var_56, interleave = new_x_91_interleave_0, values = (cache_91, input_1197))[name = tensor<string, []>("new_x_91")];
+            tensor<int32, [3]> var_5066_begin_0 = const()[name = tensor<string, []>("op_5066_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> var_5066_end_0 = const()[name = tensor<string, []>("op_5066_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> var_5066_end_mask_0 = const()[name = tensor<string, []>("op_5066_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> var_5066 = slice_by_index(begin = var_5066_begin_0, end = var_5066_end_0, end_mask = var_5066_end_mask_0, x = new_x_91)[name = tensor<string, []>("op_5066")];
+            tensor<string, []> x_593_pad_type_0 = const()[name = tensor<string, []>("x_593_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_593_groups_0 = const()[name = tensor<string, []>("x_593_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_593_strides_0 = const()[name = tensor<string, []>("x_593_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_593_pad_0 = const()[name = tensor<string, []>("x_593_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_593_dilations_0 = const()[name = tensor<string, []>("x_593_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_593 = conv(dilations = x_593_dilations_0, groups = x_593_groups_0, pad = x_593_pad_0, pad_type = x_593_pad_type_0, strides = x_593_strides_0, weight = module_layers_22_conv_depthwise_conv_weight_quantized, x = new_x_91)[name = tensor<string, []>("x_593")];
+            tensor<int32, [3]> input_1199_perm_0 = const()[name = tensor<string, []>("input_1199_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_595_axes_0 = const()[name = tensor<string, []>("x_595_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1199 = transpose(perm = input_1199_perm_0, x = x_593)[name = tensor<string, []>("transpose_158")];
+            tensor<fp32, [1, 2, 1024]> x_595 = layer_norm(axes = x_595_axes_0, beta = module_layers_22_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_22_conv_batch_norm_weight, x = input_1199)[name = tensor<string, []>("x_595")];
+            tensor<int32, [3]> input_1201_perm_0 = const()[name = tensor<string, []>("input_1201_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_1201 = transpose(perm = input_1201_perm_0, x = x_595)[name = tensor<string, []>("transpose_157")];
+            tensor<fp32, [1, 1024, 2]> input_1203 = silu(x = input_1201)[name = tensor<string, []>("input_1203")];
+            tensor<string, []> x_597_pad_type_0 = const()[name = tensor<string, []>("x_597_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_597_strides_0 = const()[name = tensor<string, []>("x_597_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_597_pad_0 = const()[name = tensor<string, []>("x_597_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_597_dilations_0 = const()[name = tensor<string, []>("x_597_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_597_groups_0 = const()[name = tensor<string, []>("x_597_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_597 = conv(dilations = x_597_dilations_0, groups = x_597_groups_0, pad = x_597_pad_0, pad_type = x_597_pad_type_0, strides = x_597_strides_0, weight = module_layers_22_conv_pointwise_conv2_weight_quantized, x = input_1203)[name = tensor<string, []>("x_597")];
+            tensor<int32, [3]> input_1205_perm_0 = const()[name = tensor<string, []>("input_1205_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_1205 = transpose(perm = input_1205_perm_0, x = x_597)[name = tensor<string, []>("transpose_156")];
+            tensor<fp32, [1, 2, 1024]> input_1207 = add(x = input_1191, y = input_1205)[name = tensor<string, []>("input_1207")];
+            tensor<int32, [1]> input_1209_axes_0 = const()[name = tensor<string, []>("input_1209_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1209 = layer_norm(axes = input_1209_axes_0, beta = module_layers_22_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_22_norm_feed_forward2_weight, x = input_1207)[name = tensor<string, []>("input_1209")];
+            tensor<fp32, [1, 2, 4096]> input_1211 = linear(bias = linear_1_bias_0, weight = module_layers_22_feed_forward2_linear1_weight_quantized, x = input_1209)[name = tensor<string, []>("linear_206")];
+            tensor<fp32, [1, 2, 4096]> input_1213 = silu(x = input_1211)[name = tensor<string, []>("input_1213")];
+            tensor<fp32, [1, 2, 1024]> input_1217 = linear(bias = linear_2_bias_0, weight = module_layers_22_feed_forward2_linear2_weight_quantized, x = input_1213)[name = tensor<string, []>("linear_207")];
+            tensor<fp32, []> var_5107 = const()[name = tensor<string, []>("op_5107"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_5108 = mul(x = input_1217, y = var_5107)[name = tensor<string, []>("op_5108")];
+            tensor<fp32, [1, 2, 1024]> input_1219 = add(x = input_1207, y = var_5108)[name = tensor<string, []>("input_1219")];
+            tensor<int32, [1]> input_1221_axes_0 = const()[name = tensor<string, []>("input_1221_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1221 = layer_norm(axes = input_1221_axes_0, beta = module_layers_22_norm_out_bias, epsilon = var_38, gamma = module_layers_22_norm_out_weight, x = input_1219)[name = tensor<string, []>("input_1221")];
+            tensor<int32, [4]> cache_93_begin_0 = const()[name = tensor<string, []>("cache_93_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_93_end_0 = const()[name = tensor<string, []>("cache_93_end_0"), val = tensor<int32, [4]>([24, 1, 70, 1024])];
+            tensor<bool, [4]> cache_93_end_mask_0 = const()[name = tensor<string, []>("cache_93_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_93_squeeze_mask_0 = const()[name = tensor<string, []>("cache_93_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 70, 1024]> cache_93 = slice_by_index(begin = cache_93_begin_0, end = cache_93_end_0, end_mask = cache_93_end_mask_0, squeeze_mask = cache_93_squeeze_mask_0, x = value_3)[name = tensor<string, []>("cache_93")];
+            tensor<int32, [4]> cache_begin_0 = const()[name = tensor<string, []>("cache_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
+            tensor<int32, [4]> cache_end_0 = const()[name = tensor<string, []>("cache_end_0"), val = tensor<int32, [4]>([24, 1, 1024, 8])];
+            tensor<bool, [4]> cache_end_mask_0 = const()[name = tensor<string, []>("cache_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
+            tensor<bool, [4]> cache_squeeze_mask_0 = const()[name = tensor<string, []>("cache_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
+            tensor<fp32, [1, 1024, 8]> cache = slice_by_index(begin = cache_begin_0, end = cache_end_0, end_mask = cache_end_mask_0, squeeze_mask = cache_squeeze_mask_0, x = value_5)[name = tensor<string, []>("cache")];
+            tensor<int32, [1]> input_1223_axes_0 = const()[name = tensor<string, []>("input_1223_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1223 = layer_norm(axes = input_1223_axes_0, beta = module_layers_23_norm_feed_forward1_bias, epsilon = var_38, gamma = module_layers_23_norm_feed_forward1_weight, x = input_1221)[name = tensor<string, []>("input_1223")];
+            tensor<fp32, [1, 2, 4096]> input_1225 = linear(bias = linear_1_bias_0, weight = module_layers_23_feed_forward1_linear1_weight_quantized, x = input_1223)[name = tensor<string, []>("linear_208")];
+            tensor<fp32, [1, 2, 4096]> input_1227 = silu(x = input_1225)[name = tensor<string, []>("input_1227")];
+            tensor<fp32, [1, 2, 1024]> input_1231 = linear(bias = linear_2_bias_0, weight = module_layers_23_feed_forward1_linear2_weight_quantized, x = input_1227)[name = tensor<string, []>("linear_209")];
+            tensor<fp32, []> var_5142 = const()[name = tensor<string, []>("op_5142"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_5143 = mul(x = input_1231, y = var_5142)[name = tensor<string, []>("op_5143")];
+            tensor<fp32, [1, 2, 1024]> input_1233 = add(x = input_1221, y = var_5143)[name = tensor<string, []>("input_1233")];
+            tensor<int32, [1]> key_axes_0 = const()[name = tensor<string, []>("key_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> key = layer_norm(axes = key_axes_0, beta = module_layers_23_norm_self_att_bias, epsilon = var_38, gamma = module_layers_23_norm_self_att_weight, x = input_1233)[name = tensor<string, []>("key")];
+            tensor<bool, []> input_1235_interleave_0 = const()[name = tensor<string, []>("input_1235_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 72, 1024]> input_1235 = concat(axis = var_65, interleave = input_1235_interleave_0, values = (cache_93, key))[name = tensor<string, []>("input_1235")];
+            tensor<int32, [3]> var_5165_begin_0 = const()[name = tensor<string, []>("op_5165_begin_0"), val = tensor<int32, [3]>([0, 2, 0])];
+            tensor<int32, [3]> var_5165_end_0 = const()[name = tensor<string, []>("op_5165_end_0"), val = tensor<int32, [3]>([1, 70, 1024])];
+            tensor<bool, [3]> var_5165_end_mask_0 = const()[name = tensor<string, []>("op_5165_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 68, 1024]> var_5165 = slice_by_index(begin = var_5165_begin_0, end = var_5165_end_0, end_mask = var_5165_end_mask_0, x = cache_93)[name = tensor<string, []>("op_5165")];
+            tensor<bool, []> cache_last_channel_cur_interleave_0 = const()[name = tensor<string, []>("cache_last_channel_cur_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 70, 1024]> cache_last_channel_cur = concat(axis = var_65, interleave = cache_last_channel_cur_interleave_0, values = (var_5165, key))[name = tensor<string, []>("cache_last_channel_cur")];
+            tensor<fp32, [1, 2, 1024]> var_5174 = linear(bias = linear_2_bias_0, weight = module_layers_23_self_attn_linear_q_weight_quantized, x = key)[name = tensor<string, []>("linear_210")];
+            tensor<int32, [4]> var_5175 = const()[name = tensor<string, []>("op_5175"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 2, 8, 128]> q_139 = reshape(shape = var_5175, x = var_5174)[name = tensor<string, []>("q_139")];
+            tensor<fp32, [1, 72, 1024]> var_5178 = linear(bias = linear_2_bias_0, weight = module_layers_23_self_attn_linear_k_weight_quantized, x = input_1235)[name = tensor<string, []>("linear_211")];
+            tensor<int32, [4]> var_5179 = const()[name = tensor<string, []>("op_5179"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> k_93 = reshape(shape = var_5179, x = var_5178)[name = tensor<string, []>("k_93")];
+            tensor<fp32, [1, 72, 1024]> var_5182 = linear(bias = linear_2_bias_0, weight = module_layers_23_self_attn_linear_v_weight_quantized, x = input_1235)[name = tensor<string, []>("linear_212")];
+            tensor<int32, [4]> var_5183 = const()[name = tensor<string, []>("op_5183"), val = tensor<int32, [4]>([1, -1, 8, 128])];
+            tensor<fp32, [1, 72, 8, 128]> v = reshape(shape = var_5183, x = var_5182)[name = tensor<string, []>("v")];
+            tensor<int32, [4]> value_perm_0 = const()[name = tensor<string, []>("value_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 2, 8, 128]> var_5195 = add(x = q_139, y = module_layers_23_self_attn_pos_bias_u)[name = tensor<string, []>("op_5195")];
+            tensor<fp32, [1, 2, 8, 128]> var_5197 = add(x = q_139, y = module_layers_23_self_attn_pos_bias_v)[name = tensor<string, []>("op_5197")];
+            tensor<int32, [4]> q_with_bias_v_perm_0 = const()[name = tensor<string, []>("q_with_bias_v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<fp32, [1, 8, 128, 143]> op_5199_quantized = constexpr_affine_dequantize()[axis = tensor<int32, []>(3), name = tensor<string, []>("op_5199_quantized"), quantized_data = tensor<int8, [1, 8, 128, 143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590723200))), scale = tensor<fp32, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590869952))), zero_point = tensor<int8, [143]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(590869696)))];
+            tensor<bool, []> x_605_transpose_x_0 = const()[name = tensor<string, []>("x_605_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_605_transpose_y_0 = const()[name = tensor<string, []>("x_605_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 2, 128]> q_with_bias_v = transpose(perm = q_with_bias_v_perm_0, x = var_5197)[name = tensor<string, []>("transpose_155")];
+            tensor<fp32, [1, 8, 2, 143]> x_605 = matmul(transpose_x = x_605_transpose_x_0, transpose_y = x_605_transpose_y_0, x = q_with_bias_v, y = op_5199_quantized)[name = tensor<string, []>("x_605")];
+            tensor<fp32, []> const_378 = const()[name = tensor<string, []>("const_378"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [8]> x_607_pad_0 = const()[name = tensor<string, []>("x_607_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 0, 0, 1, 0])];
+            tensor<string, []> x_607_mode_0 = const()[name = tensor<string, []>("x_607_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 8, 2, 144]> x_607 = pad(constant_val = const_378, mode = x_607_mode_0, pad = x_607_pad_0, x = x_605)[name = tensor<string, []>("x_607")];
+            tensor<int32, [4]> var_5207 = const()[name = tensor<string, []>("op_5207"), val = tensor<int32, [4]>([1, 8, -1, 2])];
+            tensor<fp32, [1, 8, 144, 2]> x_609 = reshape(shape = var_5207, x = x_607)[name = tensor<string, []>("x_609")];
+            tensor<int32, [4]> var_5211_begin_0 = const()[name = tensor<string, []>("op_5211_begin_0"), val = tensor<int32, [4]>([0, 0, 1, 0])];
+            tensor<int32, [4]> var_5211_end_0 = const()[name = tensor<string, []>("op_5211_end_0"), val = tensor<int32, [4]>([1, 8, 144, 2])];
+            tensor<bool, [4]> var_5211_end_mask_0 = const()[name = tensor<string, []>("op_5211_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
+            tensor<fp32, [1, 8, 143, 2]> var_5211 = slice_by_index(begin = var_5211_begin_0, end = var_5211_end_0, end_mask = var_5211_end_mask_0, x = x_609)[name = tensor<string, []>("op_5211")];
+            tensor<int32, [4]> var_5212 = const()[name = tensor<string, []>("op_5212"), val = tensor<int32, [4]>([1, 8, 2, 143])];
+            tensor<fp32, [1, 8, 2, 143]> matrix_bd_93 = reshape(shape = var_5212, x = var_5211)[name = tensor<string, []>("matrix_bd_93")];
+            tensor<bool, []> matrix_ac_transpose_x_0 = const()[name = tensor<string, []>("matrix_ac_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> matrix_ac_transpose_y_0 = const()[name = tensor<string, []>("matrix_ac_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<int32, [4]> transpose_142_perm_0 = const()[name = tensor<string, []>("transpose_142_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
+            tensor<int32, [4]> transpose_143_perm_0 = const()[name = tensor<string, []>("transpose_143_perm_0"), val = tensor<int32, [4]>([0, 2, -1, -3])];
+            tensor<fp32, [1, 8, 128, 72]> transpose_143 = transpose(perm = transpose_143_perm_0, x = k_93)[name = tensor<string, []>("transpose_153")];
+            tensor<fp32, [1, 8, 2, 128]> transpose_142 = transpose(perm = transpose_142_perm_0, x = var_5195)[name = tensor<string, []>("transpose_154")];
+            tensor<fp32, [1, 8, 2, 72]> matrix_ac = matmul(transpose_x = matrix_ac_transpose_x_0, transpose_y = matrix_ac_transpose_y_0, x = transpose_142, y = transpose_143)[name = tensor<string, []>("matrix_ac")];
+            tensor<int32, [4]> matrix_bd_begin_0 = const()[name = tensor<string, []>("matrix_bd_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
+            tensor<int32, [4]> matrix_bd_end_0 = const()[name = tensor<string, []>("matrix_bd_end_0"), val = tensor<int32, [4]>([1, 8, 2, 72])];
+            tensor<bool, [4]> matrix_bd_end_mask_0 = const()[name = tensor<string, []>("matrix_bd_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
+            tensor<fp32, [1, 8, 2, 72]> matrix_bd = slice_by_index(begin = matrix_bd_begin_0, end = matrix_bd_end_0, end_mask = matrix_bd_end_mask_0, x = matrix_bd_93)[name = tensor<string, []>("matrix_bd")];
+            tensor<fp32, [1, 8, 2, 72]> var_5221 = add(x = matrix_ac, y = matrix_bd)[name = tensor<string, []>("op_5221")];
+            tensor<fp32, []> _inversed_scores_93_y_0 = const()[name = tensor<string, []>("_inversed_scores_93_y_0"), val = tensor<fp32, []>(0x1.6a09e6p-4)];
+            tensor<fp32, [1, 8, 2, 72]> _inversed_scores_93 = mul(x = var_5221, y = _inversed_scores_93_y_0)[name = tensor<string, []>("_inversed_scores_93")];
+            tensor<fp32, [1, 8, 2, 72]> scores = select(a = var_41, b = _inversed_scores_93, cond = mask_11)[name = tensor<string, []>("scores")];
+            tensor<fp32, [1, 8, 2, 72]> var_5227 = softmax(axis = var_56, x = scores)[name = tensor<string, []>("op_5227")];
+            tensor<fp32, [1, 8, 2, 72]> input_1237 = select(a = var_40, b = var_5227, cond = mask_11)[name = tensor<string, []>("input_1237")];
+            tensor<bool, []> x_611_transpose_x_0 = const()[name = tensor<string, []>("x_611_transpose_x_0"), val = tensor<bool, []>(false)];
+            tensor<bool, []> x_611_transpose_y_0 = const()[name = tensor<string, []>("x_611_transpose_y_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 8, 72, 128]> value = transpose(perm = value_perm_0, x = v)[name = tensor<string, []>("transpose_152")];
+            tensor<fp32, [1, 8, 2, 128]> x_611 = matmul(transpose_x = x_611_transpose_x_0, transpose_y = x_611_transpose_y_0, x = input_1237, y = value)[name = tensor<string, []>("x_611")];
+            tensor<int32, [4]> var_5231_perm_0 = const()[name = tensor<string, []>("op_5231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
+            tensor<int32, [3]> var_5232 = const()[name = tensor<string, []>("op_5232"), val = tensor<int32, [3]>([1, -1, 1024])];
+            tensor<fp32, [1, 2, 8, 128]> var_5231 = transpose(perm = var_5231_perm_0, x = x_611)[name = tensor<string, []>("transpose_151")];
+            tensor<fp32, [1, 2, 1024]> input_1239 = reshape(shape = var_5232, x = var_5231)[name = tensor<string, []>("input_1239")];
+            tensor<fp32, [1, 2, 1024]> input_1241 = linear(bias = linear_2_bias_0, weight = module_layers_23_self_attn_linear_out_weight_quantized, x = input_1239)[name = tensor<string, []>("linear_214")];
+            tensor<fp32, [1, 2, 1024]> input_1243 = add(x = input_1233, y = input_1241)[name = tensor<string, []>("input_1243")];
+            tensor<int32, [1]> x_615_axes_0 = const()[name = tensor<string, []>("x_615_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> x_615 = layer_norm(axes = x_615_axes_0, beta = module_layers_23_norm_conv_bias, epsilon = var_38, gamma = module_layers_23_norm_conv_weight, x = input_1243)[name = tensor<string, []>("x_615")];
+            tensor<int32, [3]> input_1245_perm_0 = const()[name = tensor<string, []>("input_1245_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<string, []> input_1247_pad_type_0 = const()[name = tensor<string, []>("input_1247_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> input_1247_strides_0 = const()[name = tensor<string, []>("input_1247_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> input_1247_pad_0 = const()[name = tensor<string, []>("input_1247_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> input_1247_dilations_0 = const()[name = tensor<string, []>("input_1247_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> input_1247_groups_0 = const()[name = tensor<string, []>("input_1247_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> input_1245 = transpose(perm = input_1245_perm_0, x = x_615)[name = tensor<string, []>("transpose_150")];
+            tensor<fp32, [1, 2048, 2]> input_1247 = conv(dilations = input_1247_dilations_0, groups = input_1247_groups_0, pad = input_1247_pad_0, pad_type = input_1247_pad_type_0, strides = input_1247_strides_0, weight = module_layers_23_conv_pointwise_conv1_weight_quantized, x = input_1245)[name = tensor<string, []>("input_1247")];
+            tensor<int32, []> x_617_split_num_splits_0 = const()[name = tensor<string, []>("x_617_split_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> x_617_split_axis_0 = const()[name = tensor<string, []>("x_617_split_axis_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_617_split_0, tensor<fp32, [1, 1024, 2]> x_617_split_1 = split(axis = x_617_split_axis_0, num_splits = x_617_split_num_splits_0, x = input_1247)[name = tensor<string, []>("x_617_split")];
+            tensor<fp32, [1, 1024, 2]> x_617_split_1_sigmoid = sigmoid(x = x_617_split_1)[name = tensor<string, []>("x_617_split_1_sigmoid")];
+            tensor<fp32, [1, 1024, 2]> x_617 = mul(x = x_617_split_0, y = x_617_split_1_sigmoid)[name = tensor<string, []>("x_617")];
+            tensor<fp32, [1, 1024, 2]> input_1249 = select(a = var_40, b = x_617, cond = var_565)[name = tensor<string, []>("input_1249")];
+            tensor<bool, []> new_x_interleave_0 = const()[name = tensor<string, []>("new_x_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 1024, 10]> new_x = concat(axis = var_56, interleave = new_x_interleave_0, values = (cache, input_1249))[name = tensor<string, []>("new_x")];
+            tensor<int32, [3]> cache_last_time_cur_begin_0 = const()[name = tensor<string, []>("cache_last_time_cur_begin_0"), val = tensor<int32, [3]>([0, 0, 2])];
+            tensor<int32, [3]> cache_last_time_cur_end_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_0"), val = tensor<int32, [3]>([1, 1024, 10])];
+            tensor<bool, [3]> cache_last_time_cur_end_mask_0 = const()[name = tensor<string, []>("cache_last_time_cur_end_mask_0"), val = tensor<bool, [3]>([true, true, true])];
+            tensor<fp32, [1, 1024, 8]> cache_last_time_cur = slice_by_index(begin = cache_last_time_cur_begin_0, end = cache_last_time_cur_end_0, end_mask = cache_last_time_cur_end_mask_0, x = new_x)[name = tensor<string, []>("cache_last_time_cur")];
+            tensor<string, []> x_619_pad_type_0 = const()[name = tensor<string, []>("x_619_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, []> x_619_groups_0 = const()[name = tensor<string, []>("x_619_groups_0"), val = tensor<int32, []>(1024)];
+            tensor<int32, [1]> x_619_strides_0 = const()[name = tensor<string, []>("x_619_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_619_pad_0 = const()[name = tensor<string, []>("x_619_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_619_dilations_0 = const()[name = tensor<string, []>("x_619_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1024, 2]> x_619 = conv(dilations = x_619_dilations_0, groups = x_619_groups_0, pad = x_619_pad_0, pad_type = x_619_pad_type_0, strides = x_619_strides_0, weight = module_layers_23_conv_depthwise_conv_weight_quantized, x = new_x)[name = tensor<string, []>("x_619")];
+            tensor<int32, [3]> input_1251_perm_0 = const()[name = tensor<string, []>("input_1251_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> x_621_axes_0 = const()[name = tensor<string, []>("x_621_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1251 = transpose(perm = input_1251_perm_0, x = x_619)[name = tensor<string, []>("transpose_149")];
+            tensor<fp32, [1, 2, 1024]> x_621 = layer_norm(axes = x_621_axes_0, beta = module_layers_23_conv_batch_norm_bias, epsilon = var_38, gamma = module_layers_23_conv_batch_norm_weight, x = input_1251)[name = tensor<string, []>("x_621")];
+            tensor<int32, [3]> input_1253_perm_0 = const()[name = tensor<string, []>("input_1253_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1024, 2]> input_1253 = transpose(perm = input_1253_perm_0, x = x_621)[name = tensor<string, []>("transpose_148")];
+            tensor<fp32, [1, 1024, 2]> input_1255 = silu(x = input_1253)[name = tensor<string, []>("input_1255")];
+            tensor<string, []> x_623_pad_type_0 = const()[name = tensor<string, []>("x_623_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> x_623_strides_0 = const()[name = tensor<string, []>("x_623_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> x_623_pad_0 = const()[name = tensor<string, []>("x_623_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> x_623_dilations_0 = const()[name = tensor<string, []>("x_623_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> x_623_groups_0 = const()[name = tensor<string, []>("x_623_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 1024, 2]> x_623 = conv(dilations = x_623_dilations_0, groups = x_623_groups_0, pad = x_623_pad_0, pad_type = x_623_pad_type_0, strides = x_623_strides_0, weight = module_layers_23_conv_pointwise_conv2_weight_quantized, x = input_1255)[name = tensor<string, []>("x_623")];
+            tensor<int32, [3]> input_1257_perm_0 = const()[name = tensor<string, []>("input_1257_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 2, 1024]> input_1257 = transpose(perm = input_1257_perm_0, x = x_623)[name = tensor<string, []>("transpose_147")];
+            tensor<fp32, [1, 2, 1024]> input_1259 = add(x = input_1243, y = input_1257)[name = tensor<string, []>("input_1259")];
+            tensor<int32, [1]> input_1261_axes_0 = const()[name = tensor<string, []>("input_1261_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> input_1261 = layer_norm(axes = input_1261_axes_0, beta = module_layers_23_norm_feed_forward2_bias, epsilon = var_38, gamma = module_layers_23_norm_feed_forward2_weight, x = input_1259)[name = tensor<string, []>("input_1261")];
+            tensor<fp32, [1, 2, 4096]> input_1263 = linear(bias = linear_1_bias_0, weight = module_layers_23_feed_forward2_linear1_weight_quantized, x = input_1261)[name = tensor<string, []>("linear_215")];
+            tensor<fp32, [1, 2, 4096]> input_1265 = silu(x = input_1263)[name = tensor<string, []>("input_1265")];
+            tensor<fp32, [1, 2, 1024]> input_1269 = linear(bias = linear_2_bias_0, weight = module_layers_23_feed_forward2_linear2_weight_quantized, x = input_1265)[name = tensor<string, []>("linear_216")];
+            tensor<fp32, []> var_5311 = const()[name = tensor<string, []>("op_5311"), val = tensor<fp32, []>(0x1p-1)];
+            tensor<fp32, [1, 2, 1024]> var_5312 = mul(x = input_1269, y = var_5311)[name = tensor<string, []>("op_5312")];
+            tensor<fp32, [1, 2, 1024]> input = add(x = input_1259, y = var_5312)[name = tensor<string, []>("input")];
+            tensor<int32, [1]> audio_signal_axes_0 = const()[name = tensor<string, []>("audio_signal_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [1, 2, 1024]> audio_signal = layer_norm(axes = audio_signal_axes_0, beta = module_layers_23_norm_out_bias, epsilon = var_38, gamma = module_layers_23_norm_out_weight, x = input)[name = tensor<string, []>("audio_signal")];
+            tensor<int32, [3]> obj_1_perm_0 = const()[name = tensor<string, []>("obj_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, []> obj_5_axis_0 = const()[name = tensor<string, []>("obj_5_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [24, 1, 70, 1024]> obj_5 = stack(axis = obj_5_axis_0, values = (var_479, var_683, var_887, var_1091, var_1295, var_1499, var_1703, var_1907, var_2111, var_2315, var_2519, var_2723, var_2927, var_3131, var_3335, var_3539, var_3743, var_3947, var_4151, var_4355, var_4559, var_4763, var_4967, cache_last_channel_cur))[name = tensor<string, []>("obj_5")];
+            tensor<int32, []> obj_7_axis_0 = const()[name = tensor<string, []>("obj_7_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [24, 1, 1024, 8]> obj_7 = stack(axis = obj_7_axis_0, values = (var_578, var_782, var_986, var_1190, var_1394, var_1598, var_1802, var_2006, var_2210, var_2414, var_2618, var_2822, var_3026, var_3230, var_3434, var_3638, var_3842, var_4046, var_4250, var_4454, var_4658, var_4862, var_5066, cache_last_time_cur))[name = tensor<string, []>("obj_7")];
+            tensor<int32, [1]> var_5328 = add(x = cache_len, y = max_audio_length_1)[name = tensor<string, []>("op_5328")];
+            tensor<fp32, []> const_384 = const()[name = tensor<string, []>("const_384"), val = tensor<fp32, []>(-0x1.fffffep+127)];
+            tensor<string, []> var_5328_promoted_dtype_0 = const()[name = tensor<string, []>("op_5328_promoted_dtype_0"), val = tensor<string, []>("fp32")];
+            tensor<fp32, []> var_45_promoted = const()[name = tensor<string, []>("op_45_promoted"), val = tensor<fp32, []>(0x1.18p+6)];
+            tensor<fp32, [1]> var_5328_promoted = cast(dtype = var_5328_promoted_dtype_0, x = var_5328)[name = tensor<string, []>("cast_2")];
+            tensor<fp32, [1]> clip_1 = clip(alpha = const_384, beta = var_45_promoted, x = var_5328_promoted)[name = tensor<string, []>("clip_1")];
+            tensor<int32, [4]> var_5338_perm_0 = const()[name = tensor<string, []>("op_5338_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<int32, [4]> var_5341_perm_0 = const()[name = tensor<string, []>("op_5341_perm_0"), val = tensor<int32, [4]>([1, 0, 2, 3])];
+            tensor<string, []> var_5346_dtype_0 = const()[name = tensor<string, []>("op_5346_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<string, []> var_5351_dtype_0 = const()[name = tensor<string, []>("op_5351_dtype_0"), val = tensor<string, []>("int32")];
+            tensor<fp32, [1, 1024, 2]> encoded = transpose(perm = obj_1_perm_0, x = audio_signal)[name = tensor<string, []>("transpose_144")];
+            tensor<fp32, [1, 24, 70, 1024]> cache_channel_out = transpose(perm = var_5338_perm_0, x = obj_5)[name = tensor<string, []>("transpose_145")];
+            tensor<fp32, [1, 24, 1024, 8]> cache_time_out = transpose(perm = var_5341_perm_0, x = obj_7)[name = tensor<string, []>("transpose_146")];
+            tensor<int32, [1]> encoded_length = cast(dtype = var_5346_dtype_0, x = clip_0)[name = tensor<string, []>("cast_0")];
+            tensor<int32, [1]> cache_len_out = cast(dtype = var_5351_dtype_0, x = clip_1)[name = tensor<string, []>("cast_1")];
+        } -> (encoded, encoded_length, cache_channel_out, cache_time_out, cache_len_out);
+}
\ No newline at end of file