Replace chunk variants with int8 quantized .mlmodelc files

- Convert 560ms, 160ms, 80ms from .mlpackage to compiled .mlmodelc
- Quantize encoders to int8 (3.97x compression: 2.2GB -> 564MB)
- Remove original .mlpackage files
- Total size reduction from ~6.6GB to ~1.8GB for all three variants

nemotron_coreml_160ms/{decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel → decoder.mlmodelc/analytics/coremldata.bin}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8235df1402789eef9449536ca6c7e456348372a13d7ef5cae8784dfbd655ebe9
-size 8734
+oid sha256:1357ea773a03fb0d353f1755603502d5d37249327b97ca487bb158484dc79b36
+size 243

nemotron_coreml_160ms/{joint.mlpackage/Data/com.apple.CoreML/model.mlmodel → decoder.mlmodelc/coremldata.bin}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96d68bb789827874c834847a52654230ac53c7b525cf678815ec579fd7bc3c1b
-size 3326
+oid sha256:5b76bbefd9e9e8a9ad9ec9be741bb2acd397d9629c1194b6a7e36f70b44be185
+size 492

nemotron_coreml_160ms/decoder.mlmodelc/metadata.json

@@ -0,0 +1,120 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Select" : 1,
+      "Ios17.squeeze" : 4,
+      "Ios17.gather" : 1,
+      "Ios17.lstm" : 2,
+      "Identity" : 1,
+      "Ios17.transpose" : 2,
+      "Split" : 2,
+      "Ios17.add" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Stack" : 2
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "token",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "token_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "decoder",
+    "method" : "predict"
+  }
+]

nemotron_coreml_160ms/decoder.mlmodelc/model.mil

@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            tensor<fp32, [1025, 640]> module_prediction_embed_weight = const()[name = tensor<string, []>("module_prediction_embed_weight"), val = tensor<fp32, [1025, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = token, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(1025)];
+            tensor<int32, [1, 1]> add_2 = add(x = token, y = slice_by_index_0)[name = tensor<string, []>("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = token, b = add_2, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> y_axis_1 = const()[name = tensor<string, []>("y_axis_1"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> y = gather(axis = y_axis_1, batch_dims = y_batch_dims_0, indices = select_0, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight)[name = tensor<string, []>("y")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> split_0_0, tensor<fp32, [1, 1, 640]> split_0_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in)[name = tensor<string, []>("split_0")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> split_1_0, tensor<fp32, [1, 1, 640]> split_1_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in)[name = tensor<string, []>("split_1")];
+            tensor<fp32, [2560]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624128)))];
+            tensor<fp32, [2560, 640]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2634432)))];
+            tensor<fp32, [2560, 640]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9188096)))];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze")];
+            tensor<string, []> input_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [1, 1, 640]> input_3 = transpose(perm = input_3_perm_0, x = y)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 1, 640]> input_lstm_layer_0_0, tensor<fp32, [1, 640]> input_lstm_layer_0_1, tensor<fp32, [1, 640]> input_lstm_layer_0_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze, initial_h = input_lstm_layer_0_lstm_h0_squeeze, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2, weight_ih = concat_1, x = input_3)[name = tensor<string, []>("input_lstm_layer_0")];
+            tensor<fp32, [2560]> concat_3 = const()[name = tensor<string, []>("concat_3"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15741760)))];
+            tensor<fp32, [2560, 640]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15752064)))];
+            tensor<fp32, [2560, 640]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22305728)))];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_h0_squeeze = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_1)[name = tensor<string, []>("input_lstm_h0_squeeze")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_c0_squeeze = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_1)[name = tensor<string, []>("input_lstm_c0_squeeze")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [1, 1, 640]> input_0, tensor<fp32, [1, 640]> input_1, tensor<fp32, [1, 640]> input_2 = lstm(activation = input_activation_0, bias = concat_3, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze, initial_h = input_lstm_h0_squeeze, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5, weight_ih = concat_4, x = input_lstm_layer_0_0)[name = tensor<string, []>("input")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> h_out = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_1, input_1))[name = tensor<string, []>("obj_3")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> c_out = stack(axis = obj_axis_0, values = (input_lstm_layer_0_2, input_2))[name = tensor<string, []>("obj")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<fp32, [1, 640, 1]> decoder_out = transpose(perm = transpose_0_perm_0, x = input_0)[name = tensor<string, []>("transpose_1")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = tensor<string, []>("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}

nemotron_coreml_160ms/decoder.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "18C5CE6D-76FF-4E99-9598-B0324D85FA82": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        },
-        "BB41F755-71E1-4E08-A46B-A254A8EFF24A": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        }
-    },
-    "rootModelIdentifier": "BB41F755-71E1-4E08-A46B-A254A8EFF24A"
-}

nemotron_coreml_160ms/encoder/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:45d83ed8f04162a80e63ecb8d988a33e66757086e3746995a0e59c30433abca4
-size 2350022976

nemotron_coreml_160ms/encoder/encoder.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "18210CE0-CDF0-4569-A681-F10CCF13AC4F": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        },
-        "1E80803B-2C09-4FF6-BE69-D90B01CA9B74": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        }
-    },
-    "rootModelIdentifier": "18210CE0-CDF0-4569-A681-F10CCF13AC4F"
-}

nemotron_coreml_560ms/decoder.mlpackage/Data/com.apple.CoreML/model.mlmodel → nemotron_coreml_160ms/encoder/encoder_int8.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8235df1402789eef9449536ca6c7e456348372a13d7ef5cae8784dfbd655ebe9
-size 8734
+oid sha256:c25c31b1ca93677a224c6e6e9d3aeaa14d63b0e7c9b18fa7e8d22f7e00ae18ff
+size 243

nemotron_coreml_560ms/joint.mlpackage/Data/com.apple.CoreML/model.mlmodel → nemotron_coreml_160ms/encoder/encoder_int8.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96d68bb789827874c834847a52654230ac53c7b525cf678815ec579fd7bc3c1b
-size 3326
+oid sha256:68b215265240323c5dfb95b39668f9b39b423ea340d4affc0c5d644a70971d2e
+size 669

nemotron_coreml_160ms/encoder/encoder_int8.mlmodelc/metadata.json

@@ -0,0 +1,171 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Nemotron Streaming Encoder (int8 quantized)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 2)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 2]",
+        "name" : "encoded",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 70 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 70, 1024]",
+        "name" : "cache_channel_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 1024, 8]",
+        "name" : "cache_time_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float32, Int8)",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.logicalAnd" : 3,
+      "Ios17.reshape" : 145,
+      "Ios16.softmax" : 24,
+      "Ios17.matmul" : 72,
+      "Ios17.transpose" : 224,
+      "Split" : 24,
+      "Ios17.expandDims" : 18,
+      "Select" : 72,
+      "Ios17.add" : 180,
+      "Tile" : 8,
+      "Ios17.sliceByIndex" : 147,
+      "Ios16.sigmoid" : 24,
+      "Pad" : 27,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 144,
+      "Ios16.constexprAffineDequantize" : 294,
+      "Ios17.less" : 5,
+      "Ios17.sub" : 4,
+      "Ios17.conv" : 77,
+      "Ios16.relu" : 3,
+      "Ios17.clip" : 2,
+      "Ios17.linear" : 193,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.floorDiv" : 3,
+      "Ios17.cast" : 12,
+      "Ios16.silu" : 72,
+      "Ios17.concat" : 72,
+      "Stack" : 2,
+      "Ios17.mul" : 106
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 25)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 25]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 70 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 70, 1024]",
+        "name" : "cache_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 1024, 8]",
+        "name" : "cache_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "encoder_int8",
+    "method" : "predict"
+  }
+]

nemotron_coreml_160ms/encoder/{encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel → encoder_int8.mlmodelc/weights/weight.bin}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:037e8d0375acbd8ceb92173640f1d5f0fa3a706780344bb8e1d0c1272aa82fdf
-size 638845
+oid sha256:54ceddffffaa3a7d8dd7bf227624bfd0e014a574419daf6453a0698244e8e144
+size 590870588

nemotron_coreml_160ms/joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:795588a93820f6ec4823e81d9da03b1a84e9a889c845b1e83213e4e18e3c5e67
+size 243

nemotron_coreml_160ms/joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81370bb4ea1f4e92b73c4cf7a0f8cc7e46f433d475215584afb58af27e192c62
+size 400

nemotron_coreml_160ms/joint.mlmodelc/metadata.json

@@ -0,0 +1,75 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 1025)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 1025]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.expandDims" : 2,
+      "Ios17.transpose" : 2,
+      "Ios17.linear" : 3,
+      "Ios17.add" : 1,
+      "Ios16.relu" : 1
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 1]",
+        "name" : "encoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "joint",
+    "method" : "predict"
+  }
+]

nemotron_coreml_160ms/joint.mlmodelc/model.mil

@@ -0,0 +1,25 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<fp32, [640]> module_enc_bias = const()[name = tensor<string, []>("module_enc_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [640, 1024]> module_enc_weight = const()[name = tensor<string, []>("module_enc_weight"), val = tensor<fp32, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2688)))];
+            tensor<fp32, [640]> module_pred_bias = const()[name = tensor<string, []>("module_pred_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624192)))];
+            tensor<fp32, [640, 640]> module_pred_weight = const()[name = tensor<string, []>("module_pred_weight"), val = tensor<fp32, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2626816)))];
+            tensor<fp32, [1025]> module_joint_net_2_bias = const()[name = tensor<string, []>("module_joint_net_2_bias"), val = tensor<fp32, [1025]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4265280)))];
+            tensor<fp32, [1025, 640]> module_joint_net_2_weight = const()[name = tensor<string, []>("module_joint_net_2_weight"), val = tensor<fp32, [1025, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4269504)))];
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1, 1024]> input_1 = transpose(perm = input_1_perm_0, x = encoder)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 1, 640]> enc_proj = linear(bias = module_enc_bias, weight = module_enc_weight, x = input_1)[name = tensor<string, []>("linear_0")];
+            tensor<fp32, [1, 1, 640]> input_3 = transpose(perm = input_3_perm_0, x = decoder)[name = tensor<string, []>("transpose_0")];
+            tensor<fp32, [1, 1, 640]> dec_proj = linear(bias = module_pred_bias, weight = module_pred_weight, x = input_3)[name = tensor<string, []>("linear_1")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 1, 1, 640]> var_23 = expand_dims(axes = var_23_axes_0, x = enc_proj)[name = tensor<string, []>("op_23")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = tensor<string, []>("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 1, 640]> var_25 = expand_dims(axes = var_25_axes_0, x = dec_proj)[name = tensor<string, []>("op_25")];
+            tensor<fp32, [1, 1, 1, 640]> input_5 = add(x = var_23, y = var_25)[name = tensor<string, []>("input_5")];
+            tensor<fp32, [1, 1, 1, 640]> input_7 = relu(x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp32, [1, 1, 1, 1025]> logits = linear(bias = module_joint_net_2_bias, weight = module_joint_net_2_weight, x = input_7)[name = tensor<string, []>("linear_2")];
+        } -> (logits);
+}

nemotron_coreml_160ms/joint.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "1D02C143-1820-4904-A8D1-9D058325DC09": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        },
-        "627B0BEF-7268-413A-8118-20E6785E38B1": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        }
-    },
-    "rootModelIdentifier": "1D02C143-1820-4904-A8D1-9D058325DC09"
-}

nemotron_coreml_160ms/metadata.json

@@ -24,5 +24,10 @@
   ],
   "decoder_hidden": 640,
   "decoder_layers": 2,
-  "encoder_dim": 1024
+  "encoder_dim": 1024,
+  "quantization": {
+    "encoder": "int8",
+    "baseline_size_mb": 2241.8,
+    "quantized_size_mb": 564.2
+  }
 }

nemotron_coreml_160ms/preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d69676882610f7d73b3e848f1321e1c8f992c2effb2abcdca40e260fbb263eac
+size 243

nemotron_coreml_160ms/preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f5d04b110719e59b7e84e9e6d1d5d60ae48c5e01bee99f1239956dc8a91b834
+size 430

nemotron_coreml_160ms/preprocessor.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 2,
+      "Ios17.equal" : 1,
+      "Ios17.reshape" : 2,
+      "Identity" : 1,
+      "Ios17.matmul" : 1,
+      "Select" : 3,
+      "Ios17.expandDims" : 7,
+      "Ios17.add" : 2,
+      "Ios17.sliceByIndex" : 3,
+      "Ios16.reduceSum" : 1,
+      "Shape" : 2,
+      "Ios17.gather" : 2,
+      "Ios17.logicalNot" : 1,
+      "Pad" : 1,
+      "Ios17.log" : 1,
+      "Ios17.less" : 1,
+      "Ios17.sub" : 2,
+      "Ios17.conv" : 2,
+      "Ios17.pow" : 1,
+      "Ios17.concat" : 1,
+      "Stack" : 1,
+      "Ios17.floorDiv" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.mul" : 1
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...480000",
+        "shapeRange" : "[[1, 1], [1, 480000]]",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "audio",
+        "shortDescription" : ""
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "preprocessor",
+    "method" : "predict"
+  }
+]

nemotron_coreml_160ms/preprocessor.mlmodelc/model.mil

@@ -0,0 +1,110 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, ?]> audio, tensor<int32, [1]> audio_length) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, list<tensor<int32, [2]>, ?>>>>((("DefaultShapes", {{"audio", [1, 1]}}), ("RangeDims", {{"audio", [[1, 1], [1, 480000]]}})))] {
+            tensor<int32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<int32, []>(1)];
+            tensor<int32, []> var_10 = const()[name = tensor<string, []>("op_10"), val = tensor<int32, []>(160)];
+            tensor<int32, []> var_12 = const()[name = tensor<string, []>("op_12"), val = tensor<int32, []>(0)];
+            tensor<fp32, []> var_16 = const()[name = tensor<string, []>("op_16"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, []> var_33 = const()[name = tensor<string, []>("op_33"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_34 = add(x = audio_length, y = var_33)[name = tensor<string, []>("op_34")];
+            tensor<int32, []> var_35 = const()[name = tensor<string, []>("op_35"), val = tensor<int32, []>(512)];
+            tensor<int32, [1]> var_36 = sub(x = var_34, y = var_35)[name = tensor<string, []>("op_36")];
+            tensor<int32, [1]> floor_div_0 = floor_div(x = var_36, y = var_10)[name = tensor<string, []>("floor_div_0")];
+            tensor<bool, [1]> var_39 = equal(x = audio_length, y = var_12)[name = tensor<string, []>("op_39")];
+            tensor<int32, [1]> var_40 = const()[name = tensor<string, []>("op_40"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1]> mel_length = select(a = var_40, b = floor_div_0, cond = var_39)[name = tensor<string, []>("seq_len")];
+            tensor<int32, [2]> var_42_shape = shape(x = audio)[name = tensor<string, []>("op_42_shape")];
+            tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_0 = const()[name = tensor<string, []>("select_0"), val = tensor<int32, []>(1)];
+            tensor<int32, []> gather_0_axis_1 = const()[name = tensor<string, []>("gather_0_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_0 = gather(axis = gather_0_axis_1, batch_dims = gather_0_batch_dims_0, indices = select_0, validate_indices = gather_0_validate_indices_0, x = var_42_shape)[name = tensor<string, []>("gather_0")];
+            tensor<int32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> var_43 = range_1d(end = gather_0, start = const_0, step = const_1)[name = tensor<string, []>("op_43")];
+            tensor<int32, [1]> var_44_axes_0 = const()[name = tensor<string, []>("op_44_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> var_44 = expand_dims(axes = var_44_axes_0, x = var_43)[name = tensor<string, []>("op_44")];
+            tensor<int32, [1]> var_45_axes_0 = const()[name = tensor<string, []>("op_45_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_45 = expand_dims(axes = var_45_axes_0, x = audio_length)[name = tensor<string, []>("op_45")];
+            tensor<bool, [1, ?]> timemask = less(x = var_44, y = var_45)[name = tensor<string, []>("timemask")];
+            tensor<int32, [2]> var_48_begin_0 = const()[name = tensor<string, []>("op_48_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_48_end_0 = const()[name = tensor<string, []>("op_48_end_0"), val = tensor<int32, [2]>([1, 1])];
+            tensor<bool, [2]> var_48_end_mask_0 = const()[name = tensor<string, []>("op_48_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<bool, [2]> var_48_squeeze_mask_0 = const()[name = tensor<string, []>("op_48_squeeze_mask_0"), val = tensor<bool, [2]>([false, true])];
+            tensor<fp32, [1]> var_48 = slice_by_index(begin = var_48_begin_0, end = var_48_end_0, end_mask = var_48_end_mask_0, squeeze_mask = var_48_squeeze_mask_0, x = audio)[name = tensor<string, []>("op_48")];
+            tensor<int32, [1]> var_49_axes_0 = const()[name = tensor<string, []>("op_49_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1]> var_49 = expand_dims(axes = var_49_axes_0, x = var_48)[name = tensor<string, []>("op_49")];
+            tensor<int32, [2]> var_51_begin_0 = const()[name = tensor<string, []>("op_51_begin_0"), val = tensor<int32, [2]>([0, 1])];
+            tensor<int32, [2]> var_51_end_0 = const()[name = tensor<string, []>("op_51_end_0"), val = tensor<int32, [2]>([1, 0])];
+            tensor<bool, [2]> var_51_end_mask_0 = const()[name = tensor<string, []>("op_51_end_mask_0"), val = tensor<bool, [2]>([true, true])];
+            tensor<fp32, [1, ?]> var_51 = slice_by_index(begin = var_51_begin_0, end = var_51_end_0, end_mask = var_51_end_mask_0, x = audio)[name = tensor<string, []>("op_51")];
+            tensor<int32, [2]> var_53_begin_0 = const()[name = tensor<string, []>("op_53_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_53_end_0 = const()[name = tensor<string, []>("op_53_end_0"), val = tensor<int32, [2]>([1, -1])];
+            tensor<bool, [2]> var_53_end_mask_0 = const()[name = tensor<string, []>("op_53_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp32, [1, ?]> var_53 = slice_by_index(begin = var_53_begin_0, end = var_53_end_0, end_mask = var_53_end_mask_0, x = audio)[name = tensor<string, []>("op_53")];
+            tensor<fp32, []> var_54 = const()[name = tensor<string, []>("op_54"), val = tensor<fp32, []>(0x1.f0a3d8p-1)];
+            tensor<fp32, [1, ?]> var_55 = mul(x = var_53, y = var_54)[name = tensor<string, []>("op_55")];
+            tensor<fp32, [1, ?]> var_56 = sub(x = var_51, y = var_55)[name = tensor<string, []>("op_56")];
+            tensor<bool, []> x_3_interleave_0 = const()[name = tensor<string, []>("x_3_interleave_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, ?]> x_3 = concat(axis = var_9, interleave = x_3_interleave_0, values = (var_49, var_56))[name = tensor<string, []>("x_3")];
+            tensor<bool, [1, ?]> var_59 = logical_not(x = timemask)[name = tensor<string, []>("op_59")];
+            tensor<fp32, [1, ?]> input_1 = select(a = var_16, b = x_3, cond = var_59)[name = tensor<string, []>("input_1")];
+            tensor<int32, [3]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [3]>([1, 1, -1])];
+            tensor<fp32, [1, 1, ?]> input_3 = reshape(shape = concat_1x, x = input_1)[name = tensor<string, []>("input_3")];
+            tensor<fp32, []> const_3 = const()[name = tensor<string, []>("const_3"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [6]> input_5_pad_0 = const()[name = tensor<string, []>("input_5_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 256, 256])];
+            tensor<string, []> input_5_mode_0 = const()[name = tensor<string, []>("input_5_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [1, 1, ?]> input_5 = pad(constant_val = const_3, mode = input_5_mode_0, pad = input_5_pad_0, x = input_3)[name = tensor<string, []>("input_5")];
+            tensor<int32, [2]> concat_2x = const()[name = tensor<string, []>("concat_2x"), val = tensor<int32, [2]>([1, -1])];
+            tensor<fp32, [1, ?]> input = reshape(shape = concat_2x, x = input_5)[name = tensor<string, []>("input")];
+            tensor<fp32, [257, 1, 512]> expand_dims_1 = const()[name = tensor<string, []>("expand_dims_1"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [257, 1, 512]> expand_dims_2 = const()[name = tensor<string, []>("expand_dims_2"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(526464)))];
+            tensor<int32, [1]> expand_dims_3 = const()[name = tensor<string, []>("expand_dims_3"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [1]> expand_dims_4_axes_0 = const()[name = tensor<string, []>("expand_dims_4_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, ?]> expand_dims_4 = expand_dims(axes = expand_dims_4_axes_0, x = input)[name = tensor<string, []>("expand_dims_4")];
+            tensor<string, []> conv_0_pad_type_0 = const()[name = tensor<string, []>("conv_0_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_0_pad_0 = const()[name = tensor<string, []>("conv_0_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_0_dilations_0 = const()[name = tensor<string, []>("conv_0_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_0_groups_0 = const()[name = tensor<string, []>("conv_0_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_0 = conv(dilations = conv_0_dilations_0, groups = conv_0_groups_0, pad = conv_0_pad_0, pad_type = conv_0_pad_type_0, strides = expand_dims_3, weight = expand_dims_1, x = expand_dims_4)[name = tensor<string, []>("conv_0")];
+            tensor<string, []> conv_1_pad_type_0 = const()[name = tensor<string, []>("conv_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [2]> conv_1_pad_0 = const()[name = tensor<string, []>("conv_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> conv_1_dilations_0 = const()[name = tensor<string, []>("conv_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> conv_1_groups_0 = const()[name = tensor<string, []>("conv_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [1, 257, ?]> conv_1 = conv(dilations = conv_1_dilations_0, groups = conv_1_groups_0, pad = conv_1_pad_0, pad_type = conv_1_pad_type_0, strides = expand_dims_3, weight = expand_dims_2, x = expand_dims_4)[name = tensor<string, []>("conv_1")];
+            tensor<int32, []> stack_0_axis_0 = const()[name = tensor<string, []>("stack_0_axis_0"), val = tensor<int32, []>(-1)];
+            tensor<fp32, [1, 257, ?, 2]> stack_0 = stack(axis = stack_0_axis_0, values = (conv_0, conv_1))[name = tensor<string, []>("stack_0")];
+            tensor<fp32, []> var_19_promoted = const()[name = tensor<string, []>("op_19_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [1, 257, ?, 2]> var_74 = pow(x = stack_0, y = var_19_promoted)[name = tensor<string, []>("op_74")];
+            tensor<int32, [1]> var_76_axes_0 = const()[name = tensor<string, []>("op_76_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<bool, []> var_76_keep_dims_0 = const()[name = tensor<string, []>("op_76_keep_dims_0"), val = tensor<bool, []>(false)];
+            tensor<fp32, [1, 257, ?]> var_76 = reduce_sum(axes = var_76_axes_0, keep_dims = var_76_keep_dims_0, x = var_74)[name = tensor<string, []>("op_76")];
+            tensor<fp32, [1, 257, ?]> x_11 = identity(x = var_76)[name = tensor<string, []>("x_11")];
+            tensor<fp32, [1, 128, 257]> const_4 = const()[name = tensor<string, []>("const_4"), val = tensor<fp32, [1, 128, 257]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1052864)))];
+            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, 128, ?]> x_13 = matmul(transpose_x = x_13_transpose_x_0, transpose_y = x_13_transpose_y_0, x = const_4, y = x_11)[name = tensor<string, []>("x_13")];
+            tensor<fp32, []> var_83 = const()[name = tensor<string, []>("op_83"), val = tensor<fp32, []>(0x1p-24)];
+            tensor<fp32, [1, 128, ?]> var_84 = add(x = x_13, y = var_83)[name = tensor<string, []>("op_84")];
+            tensor<fp32, []> x_epsilon_0 = const()[name = tensor<string, []>("x_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [1, 128, ?]> x = log(epsilon = x_epsilon_0, x = var_84)[name = tensor<string, []>("x")];
+            tensor<int32, [3]> var_86_shape = shape(x = x)[name = tensor<string, []>("op_86_shape")];
+            tensor<int32, []> gather_5_batch_dims_0 = const()[name = tensor<string, []>("gather_5_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> gather_5_validate_indices_0 = const()[name = tensor<string, []>("gather_5_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> select_3 = const()[name = tensor<string, []>("select_3"), val = tensor<int32, []>(2)];
+            tensor<int32, []> gather_5_axis_1 = const()[name = tensor<string, []>("gather_5_axis_1"), val = tensor<int32, []>(0)];
+            tensor<int32, []> gather_5 = gather(axis = gather_5_axis_1, batch_dims = gather_5_batch_dims_0, indices = select_3, validate_indices = gather_5_validate_indices_0, x = var_86_shape)[name = tensor<string, []>("gather_5")];
+            tensor<int32, []> const_5 = const()[name = tensor<string, []>("const_5"), val = tensor<int32, []>(0)];
+            tensor<int32, []> const_6 = const()[name = tensor<string, []>("const_6"), val = tensor<int32, []>(1)];
+            tensor<int32, [?]> mask_1 = range_1d(end = gather_5, start = const_5, step = const_6)[name = tensor<string, []>("mask_1")];
+            tensor<int32, [1]> expand_dims_0_axes_0 = const()[name = tensor<string, []>("expand_dims_0_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<int32, [1, ?]> expand_dims_0 = expand_dims(axes = expand_dims_0_axes_0, x = mask_1)[name = tensor<string, []>("expand_dims_0")];
+            tensor<int32, [1]> var_91_axes_0 = const()[name = tensor<string, []>("op_91_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [1, 1]> var_91 = expand_dims(axes = var_91_axes_0, x = mel_length)[name = tensor<string, []>("op_91")];
+            tensor<bool, [1, ?]> mask = greater_equal(x = expand_dims_0, y = var_91)[name = tensor<string, []>("mask")];
+            tensor<int32, [1]> var_93_axes_0 = const()[name = tensor<string, []>("op_93_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, [1, 1, ?]> var_93 = expand_dims(axes = var_93_axes_0, x = mask)[name = tensor<string, []>("op_93")];
+            tensor<fp32, [1, 128, ?]> mel = select(a = var_16, b = x, cond = var_93)[name = tensor<string, []>("processed_signal")];
+        } -> (mel, mel_length);
+}

nemotron_coreml_160ms/preprocessor.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "358ABE70-5E22-49A2-A249-CF11C6CCFEB7": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        },
-        "59E29668-EB5C-46CD-AD91-CFB4BBEC79F3": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        }
-    },
-    "rootModelIdentifier": "59E29668-EB5C-46CD-AD91-CFB4BBEC79F3"
-}

nemotron_coreml_560ms/decoder.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1357ea773a03fb0d353f1755603502d5d37249327b97ca487bb158484dc79b36
+size 243

nemotron_coreml_560ms/decoder.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:170217fe94bc926fc363b2afa6a0567b0da8524e7586de24a062a48a927563ac
+size 492

nemotron_coreml_560ms/decoder.mlmodelc/metadata.json

@@ -0,0 +1,120 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Select" : 1,
+      "Ios17.squeeze" : 4,
+      "Ios17.gather" : 1,
+      "Ios17.lstm" : 2,
+      "Identity" : 1,
+      "Ios17.transpose" : 2,
+      "Split" : 2,
+      "Ios17.add" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Stack" : 2
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1]",
+        "name" : "token",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "token_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "h_in",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 2 × 1 × 640)",
+        "shortDescription" : "",
+        "shape" : "[2, 1, 640]",
+        "name" : "c_in",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "decoder",
+    "method" : "predict"
+  }
+]

nemotron_coreml_560ms/decoder.mlmodelc/model.mil

@@ -0,0 +1,57 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<fp32, [2, 1, 640]> c_in, tensor<fp32, [2, 1, 640]> h_in, tensor<int32, [1, 1]> token, tensor<int32, [1]> token_length) {
+            tensor<fp32, [1025, 640]> module_prediction_embed_weight = const()[name = tensor<string, []>("module_prediction_embed_weight"), val = tensor<fp32, [1025, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<int32, []> y_batch_dims_0 = const()[name = tensor<string, []>("y_batch_dims_0"), val = tensor<int32, []>(0)];
+            tensor<bool, []> y_validate_indices_0 = const()[name = tensor<string, []>("y_validate_indices_0"), val = tensor<bool, []>(false)];
+            tensor<int32, []> greater_equal_0_y_0 = const()[name = tensor<string, []>("greater_equal_0_y_0"), val = tensor<int32, []>(0)];
+            tensor<bool, [1, 1]> greater_equal_0 = greater_equal(x = token, y = greater_equal_0_y_0)[name = tensor<string, []>("greater_equal_0")];
+            tensor<int32, []> slice_by_index_0 = const()[name = tensor<string, []>("slice_by_index_0"), val = tensor<int32, []>(1025)];
+            tensor<int32, [1, 1]> add_2 = add(x = token, y = slice_by_index_0)[name = tensor<string, []>("add_2")];
+            tensor<int32, [1, 1]> select_0 = select(a = token, b = add_2, cond = greater_equal_0)[name = tensor<string, []>("select_0")];
+            tensor<int32, []> y_axis_1 = const()[name = tensor<string, []>("y_axis_1"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> y = gather(axis = y_axis_1, batch_dims = y_batch_dims_0, indices = select_0, validate_indices = y_validate_indices_0, x = module_prediction_embed_weight)[name = tensor<string, []>("y")];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
+            tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> split_0_0, tensor<fp32, [1, 1, 640]> split_0_1 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = h_in)[name = tensor<string, []>("split_0")];
+            tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(2)];
+            tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [1, 1, 640]> split_1_0, tensor<fp32, [1, 1, 640]> split_1_1 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = c_in)[name = tensor<string, []>("split_1")];
+            tensor<fp32, [2560]> concat_0 = const()[name = tensor<string, []>("concat_0"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624128)))];
+            tensor<fp32, [2560, 640]> concat_1 = const()[name = tensor<string, []>("concat_1"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2634432)))];
+            tensor<fp32, [2560, 640]> concat_2 = const()[name = tensor<string, []>("concat_2"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(9188096)))];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_layer_0_lstm_h0_squeeze = squeeze(axes = input_lstm_layer_0_lstm_h0_squeeze_axes_0, x = split_0_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_h0_squeeze")];
+            tensor<int32, [1]> input_lstm_layer_0_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_layer_0_lstm_c0_squeeze = squeeze(axes = input_lstm_layer_0_lstm_c0_squeeze_axes_0, x = split_1_0)[name = tensor<string, []>("input_lstm_layer_0_lstm_c0_squeeze")];
+            tensor<string, []> input_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_lstm_layer_0_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [1, 1, 640]> input_3 = transpose(perm = input_3_perm_0, x = y)[name = tensor<string, []>("transpose_2")];
+            tensor<fp32, [1, 1, 640]> input_lstm_layer_0_0, tensor<fp32, [1, 640]> input_lstm_layer_0_1, tensor<fp32, [1, 640]> input_lstm_layer_0_2 = lstm(activation = input_lstm_layer_0_activation_0, bias = concat_0, cell_activation = input_lstm_layer_0_cell_activation_0, direction = input_lstm_layer_0_direction_0, initial_c = input_lstm_layer_0_lstm_c0_squeeze, initial_h = input_lstm_layer_0_lstm_h0_squeeze, output_sequence = input_lstm_layer_0_output_sequence_0, recurrent_activation = input_lstm_layer_0_recurrent_activation_0, weight_hh = concat_2, weight_ih = concat_1, x = input_3)[name = tensor<string, []>("input_lstm_layer_0")];
+            tensor<fp32, [2560]> concat_3 = const()[name = tensor<string, []>("concat_3"), val = tensor<fp32, [2560]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15741760)))];
+            tensor<fp32, [2560, 640]> concat_4 = const()[name = tensor<string, []>("concat_4"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(15752064)))];
+            tensor<fp32, [2560, 640]> concat_5 = const()[name = tensor<string, []>("concat_5"), val = tensor<fp32, [2560, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(22305728)))];
+            tensor<int32, [1]> input_lstm_h0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_h0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_h0_squeeze = squeeze(axes = input_lstm_h0_squeeze_axes_0, x = split_0_1)[name = tensor<string, []>("input_lstm_h0_squeeze")];
+            tensor<int32, [1]> input_lstm_c0_squeeze_axes_0 = const()[name = tensor<string, []>("input_lstm_c0_squeeze_axes_0"), val = tensor<int32, [1]>([0])];
+            tensor<fp32, [1, 640]> input_lstm_c0_squeeze = squeeze(axes = input_lstm_c0_squeeze_axes_0, x = split_1_1)[name = tensor<string, []>("input_lstm_c0_squeeze")];
+            tensor<string, []> input_direction_0 = const()[name = tensor<string, []>("input_direction_0"), val = tensor<string, []>("forward")];
+            tensor<bool, []> input_output_sequence_0 = const()[name = tensor<string, []>("input_output_sequence_0"), val = tensor<bool, []>(true)];
+            tensor<string, []> input_recurrent_activation_0 = const()[name = tensor<string, []>("input_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
+            tensor<string, []> input_cell_activation_0 = const()[name = tensor<string, []>("input_cell_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<string, []> input_activation_0 = const()[name = tensor<string, []>("input_activation_0"), val = tensor<string, []>("tanh")];
+            tensor<fp32, [1, 1, 640]> input_0, tensor<fp32, [1, 640]> input_1, tensor<fp32, [1, 640]> input_2 = lstm(activation = input_activation_0, bias = concat_3, cell_activation = input_cell_activation_0, direction = input_direction_0, initial_c = input_lstm_c0_squeeze, initial_h = input_lstm_h0_squeeze, output_sequence = input_output_sequence_0, recurrent_activation = input_recurrent_activation_0, weight_hh = concat_5, weight_ih = concat_4, x = input_lstm_layer_0_0)[name = tensor<string, []>("input")];
+            tensor<int32, []> obj_3_axis_0 = const()[name = tensor<string, []>("obj_3_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> h_out = stack(axis = obj_3_axis_0, values = (input_lstm_layer_0_1, input_1))[name = tensor<string, []>("obj_3")];
+            tensor<int32, []> obj_axis_0 = const()[name = tensor<string, []>("obj_axis_0"), val = tensor<int32, []>(0)];
+            tensor<fp32, [2, 1, 640]> c_out = stack(axis = obj_axis_0, values = (input_lstm_layer_0_2, input_2))[name = tensor<string, []>("obj")];
+            tensor<int32, [3]> transpose_0_perm_0 = const()[name = tensor<string, []>("transpose_0_perm_0"), val = tensor<int32, [3]>([1, 2, 0])];
+            tensor<fp32, [1, 640, 1]> decoder_out = transpose(perm = transpose_0_perm_0, x = input_0)[name = tensor<string, []>("transpose_1")];
+            tensor<int32, [1]> token_length_tmp = identity(x = token_length)[name = tensor<string, []>("token_length_tmp")];
+        } -> (decoder_out, h_out, c_out);
+}

nemotron_coreml_560ms/decoder.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "817639C9-4D0A-439F-898E-E7F344131738": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        },
-        "B84EA71A-A885-4940-BF41-D30E41D4D9EE": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        }
-    },
-    "rootModelIdentifier": "817639C9-4D0A-439F-898E-E7F344131738"
-}

nemotron_coreml_560ms/encoder/encoder.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0eb719eeb2e7f0a25103bee448f102ad41769b4c84f1500367a527fef7fbed5c
-size 639671

nemotron_coreml_560ms/encoder/encoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6f4a5b388cc5ee06460f44349c3341f51b2915b9a519f901b872b7bcb8037b33
-size 2351006016

nemotron_coreml_560ms/encoder/encoder.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "19200EE7-0C9A-49CF-B73D-7DA50A1C1625": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        },
-        "550E7DB0-0FC8-4F2C-AC65-9DFABA4F1DE6": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        }
-    },
-    "rootModelIdentifier": "19200EE7-0C9A-49CF-B73D-7DA50A1C1625"
-}

nemotron_coreml_560ms/encoder/encoder_int8.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0871de0ecbbb6db86ca91358b0ed94190d9f9d8950f11421452dd4add8fe086
+size 243

nemotron_coreml_560ms/encoder/encoder_int8.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d960e3871eb119a106bcff36eacc40dba59d243c8f803caf745610b2259a00d
+size 669

nemotron_coreml_560ms/encoder/encoder_int8.mlmodelc/metadata.json

@@ -0,0 +1,171 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "Nemotron Streaming Encoder (int8 quantized)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 7)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 7]",
+        "name" : "encoded",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "encoded_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 70 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 70, 1024]",
+        "name" : "cache_channel_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 1024, 8]",
+        "name" : "cache_time_out",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len_out",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float32, Int8)",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.logicalAnd" : 3,
+      "Ios17.reshape" : 145,
+      "Ios16.softmax" : 24,
+      "Ios17.matmul" : 72,
+      "Ios17.transpose" : 224,
+      "Split" : 24,
+      "Ios17.expandDims" : 18,
+      "Select" : 72,
+      "Ios17.add" : 180,
+      "Tile" : 8,
+      "Ios17.sliceByIndex" : 147,
+      "Ios16.sigmoid" : 24,
+      "Pad" : 27,
+      "Ios17.logicalNot" : 2,
+      "Ios17.layerNorm" : 144,
+      "Ios16.constexprAffineDequantize" : 294,
+      "Ios17.less" : 5,
+      "Ios17.sub" : 4,
+      "Ios17.conv" : 77,
+      "Ios16.relu" : 3,
+      "Ios17.clip" : 2,
+      "Ios17.linear" : 193,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.floorDiv" : 3,
+      "Ios17.cast" : 12,
+      "Ios16.silu" : 72,
+      "Ios17.concat" : 72,
+      "Stack" : 2,
+      "Ios17.mul" : 106
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 128 × 65)",
+        "shortDescription" : "",
+        "shape" : "[1, 128, 65]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 70 × 1024)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 70, 1024]",
+        "name" : "cache_channel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 24 × 1024 × 8)",
+        "shortDescription" : "",
+        "shape" : "[1, 24, 1024, 8]",
+        "name" : "cache_time",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "cache_len",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "encoder_int8",
+    "method" : "predict"
+  }
+]

nemotron_coreml_160ms/preprocessor.mlpackage/Data/com.apple.CoreML/model.mlmodel → nemotron_coreml_560ms/encoder/encoder_int8.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f0d5f1d74ad734692015afbf06bc9f614bb1163215caf5fc21daaa1462b59dc5
-size 12961
+oid sha256:04ababb22ec2d940f3ff4eafc79651549a129f2493d4eb43119faa6575e2489d
+size 591117860

nemotron_coreml_560ms/joint.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:795588a93820f6ec4823e81d9da03b1a84e9a889c845b1e83213e4e18e3c5e67
+size 243

nemotron_coreml_560ms/joint.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d02fd9c4cbb21aa6364d4630b752f5d2926dc77da6b693e6e8f19903027f38f4
+size 400

nemotron_coreml_560ms/joint.mlmodelc/metadata.json

@@ -0,0 +1,75 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1 × 1025)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1, 1025]",
+        "name" : "logits",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.expandDims" : 2,
+      "Ios17.transpose" : 2,
+      "Ios17.linear" : 3,
+      "Ios17.add" : 1,
+      "Ios16.relu" : 1
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1024 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 1024, 1]",
+        "name" : "encoder",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 640 × 1)",
+        "shortDescription" : "",
+        "shape" : "[1, 640, 1]",
+        "name" : "decoder",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "joint",
+    "method" : "predict"
+  }
+]

nemotron_coreml_560ms/joint.mlmodelc/model.mil

@@ -0,0 +1,25 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3510.2.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.9.1"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0"}})]
+{
+    func main<ios17>(tensor<fp32, [1, 640, 1]> decoder, tensor<fp32, [1, 1024, 1]> encoder) {
+            tensor<fp32, [640]> module_enc_bias = const()[name = tensor<string, []>("module_enc_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [640, 1024]> module_enc_weight = const()[name = tensor<string, []>("module_enc_weight"), val = tensor<fp32, [640, 1024]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2688)))];
+            tensor<fp32, [640]> module_pred_bias = const()[name = tensor<string, []>("module_pred_bias"), val = tensor<fp32, [640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2624192)))];
+            tensor<fp32, [640, 640]> module_pred_weight = const()[name = tensor<string, []>("module_pred_weight"), val = tensor<fp32, [640, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2626816)))];
+            tensor<fp32, [1025]> module_joint_net_2_bias = const()[name = tensor<string, []>("module_joint_net_2_bias"), val = tensor<fp32, [1025]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4265280)))];
+            tensor<fp32, [1025, 640]> module_joint_net_2_weight = const()[name = tensor<string, []>("module_joint_net_2_weight"), val = tensor<fp32, [1025, 640]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4269504)))];
+            tensor<int32, [3]> input_1_perm_0 = const()[name = tensor<string, []>("input_1_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [3]> input_3_perm_0 = const()[name = tensor<string, []>("input_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<fp32, [1, 1, 1024]> input_1 = transpose(perm = input_1_perm_0, x = encoder)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [1, 1, 640]> enc_proj = linear(bias = module_enc_bias, weight = module_enc_weight, x = input_1)[name = tensor<string, []>("linear_0")];
+            tensor<fp32, [1, 1, 640]> input_3 = transpose(perm = input_3_perm_0, x = decoder)[name = tensor<string, []>("transpose_0")];
+            tensor<fp32, [1, 1, 640]> dec_proj = linear(bias = module_pred_bias, weight = module_pred_weight, x = input_3)[name = tensor<string, []>("linear_1")];
+            tensor<int32, [1]> var_23_axes_0 = const()[name = tensor<string, []>("op_23_axes_0"), val = tensor<int32, [1]>([2])];
+            tensor<fp32, [1, 1, 1, 640]> var_23 = expand_dims(axes = var_23_axes_0, x = enc_proj)[name = tensor<string, []>("op_23")];
+            tensor<int32, [1]> var_25_axes_0 = const()[name = tensor<string, []>("op_25_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [1, 1, 1, 640]> var_25 = expand_dims(axes = var_25_axes_0, x = dec_proj)[name = tensor<string, []>("op_25")];
+            tensor<fp32, [1, 1, 1, 640]> input_5 = add(x = var_23, y = var_25)[name = tensor<string, []>("input_5")];
+            tensor<fp32, [1, 1, 1, 640]> input_7 = relu(x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp32, [1, 1, 1, 1025]> logits = linear(bias = module_joint_net_2_bias, weight = module_joint_net_2_weight, x = input_7)[name = tensor<string, []>("linear_2")];
+        } -> (logits);
+}

nemotron_coreml_560ms/joint.mlpackage/Manifest.json

@@ -1,18 +0,0 @@
-{
-    "fileFormatVersion": "1.0.0",
-    "itemInfoEntries": {
-        "41F607D9-A4C0-44BF-9B0C-68A035B461DA": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Weights",
-            "name": "weights",
-            "path": "com.apple.CoreML/weights"
-        },
-        "CDF0E0C6-BE68-4A6A-AE37-EF979D58F14B": {
-            "author": "com.apple.CoreML",
-            "description": "CoreML Model Specification",
-            "name": "model.mlmodel",
-            "path": "com.apple.CoreML/model.mlmodel"
-        }
-    },
-    "rootModelIdentifier": "CDF0E0C6-BE68-4A6A-AE37-EF979D58F14B"
-}

nemotron_coreml_560ms/metadata.json

@@ -24,5 +24,10 @@
   ],
   "decoder_hidden": 640,
   "decoder_layers": 2,
-  "encoder_dim": 1024
+  "encoder_dim": 1024,
+  "quantization": {
+    "encoder": "int8",
+    "baseline_size_mb": 2242.7,
+    "quantized_size_mb": 564.4
+  }
 }

nemotron_coreml_560ms/preprocessor.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d69676882610f7d73b3e848f1321e1c8f992c2effb2abcdca40e260fbb263eac
+size 243

nemotron_coreml_560ms/preprocessor.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f4b9e3acfcb0eb8c3c0270c42b603341a9e7674d2c747b6fdf6bd89a0b5f3cb
+size 430

nemotron_coreml_560ms/preprocessor.mlmodelc/metadata.json

@@ -0,0 +1,106 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "storagePrecision" : "Float32",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "mel",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "mel_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "modelParameters" : [
+
+    ],
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Range1d" : 2,
+      "Ios17.equal" : 1,
+      "Ios17.reshape" : 2,
+      "Identity" : 1,
+      "Ios17.matmul" : 1,
+      "Select" : 3,
+      "Ios17.expandDims" : 7,
+      "Ios17.add" : 2,
+      "Ios17.sliceByIndex" : 3,
+      "Ios16.reduceSum" : 1,
+      "Shape" : 2,
+      "Ios17.gather" : 2,
+      "Ios17.logicalNot" : 1,
+      "Pad" : 1,
+      "Ios17.log" : 1,
+      "Ios17.less" : 1,
+      "Ios17.sub" : 2,
+      "Ios17.conv" : 2,
+      "Ios17.pow" : 1,
+      "Ios17.concat" : 1,
+      "Stack" : 1,
+      "Ios17.floorDiv" : 1,
+      "Ios17.greaterEqual" : 1,
+      "Ios17.mul" : 1
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-01-15",
+      "com.github.apple.coremltools.source" : "torch==2.9.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "inputSchema" : [
+      {
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1...480000",
+        "shapeRange" : "[[1, 1], [1, 480000]]",
+        "formattedType" : "MultiArray (Float32 1 × 1)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1]",
+        "name" : "audio",
+        "shortDescription" : ""
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1)",
+        "shortDescription" : "",
+        "shape" : "[1]",
+        "name" : "audio_length",
+        "type" : "MultiArray"
+      }
+    ],
+    "generatedClassName" : "preprocessor",
+    "method" : "predict"
+  }
+]