Add files using upload-large-folder tool

README.md

@@ -6,6 +6,7 @@ tags:
 - ref2va
 - int8
 - convrot
+- input-major
 - turbo
 - h3ddle
 - pulpcut
@@ -36,24 +37,25 @@ package (Qwen3-VL-32B INT8 text encoder, video/audio VAEs, tokenizer) from
 FL2VA transformer alongside it if you also want prompt-only and keyframe
 generation.
 
-## H3ddle FC2 performance sidecar
+## H3ddle full input-major checkpoint
 
-`minimax_h3_ref2va_pruned_turbo_int8_convrot_fc2_input_major.safetensors`
-is an optional H3ddle performance sidecar derived from the transformer in
-this repository. It contains only the 50 INT8 `mlp.fc2.weight` matrices, with
-their storage transposed from `[output, input]` to `[input, output]`. Values,
-quantization scales, reference conditioning, and model behavior are unchanged.
+`minimax_h3_ref2va_pruned_turbo_int8_convrot_input_major.safetensors` is the
+H3ddle-optimized form of the primary transformer. All 200 quantized core
+projections across 50 blocks are stored `[input, output]` instead of
+`[output, input]`. No tensor is dequantized or numerically changed; scales,
+ConvRot metadata, reference conditioning, and non-core tensors are preserved.
 
-H3ddle installs it beside the transformer and selects it automatically after
-validating its format version, source file size, exact source-header
-fingerprint, and all 50 tensor schemas. It cannot silently be used with a
-different checkpoint. The original transformer remains available as the
-fallback and for runtimes that do not understand the sidecar.
+This layout is selected by a versioned marker and validated against every
+projection shape before H3ddle runs it. It is **not** compatible with runtimes
+that assume the original output-major Comfy layout. The original transformer
+and FC2-only sidecar remain in this repository for those runtimes and for older
+H3ddle releases; current managed installs download only the full input-major
+checkpoint.
 
-A real 512-class, 50-block Ref2VA parity run produced identical baseline and
-sidecar hashes: video `85a5ccfc5a4d8075`, audio `731e24ae9dc2e7ec`.
-The matched cold pair measured 43.753 seconds without the sidecar and 40.736
-seconds with it; the broader FL2VA A/B/B/A benchmark measured a 7.15% gain.
+On a 32 GiB M1 Pro, a matched 512x896 reference-conditioned, eight-pass
+generation reduced complete transformer time from 541.2 to 515.3 seconds:
+**4.8%**. Total generation fell from 631.4 to 595.9 seconds, and the generated
+reference output remained good.
 
 ## Why this merge was made and republished
 
@@ -139,6 +141,7 @@ of the MiniMax H3 Community License apply unchanged.
 | File | Bytes | SHA-256 |
 |---|---|---|
 | `minimax_h3_ref2va_pruned_turbo_int8_convrot.safetensors` | 20,970,379,854 | `e64cef63bc2785bcd72e6103c52aa78c6cd2c4f9870a7ce79675083fd65cf2e7` |
+| `minimax_h3_ref2va_pruned_turbo_int8_convrot_input_major.safetensors` | 20,970,380,012 | `5ca6696fe1cd9a8f254594ac67ee541f151b2377735dea3557364bd868270463` |
 | `minimax_h3_ref2va_pruned_turbo_int8_convrot_fc2_input_major.safetensors` | 3,853,522,260 | `0ad6a5673abdf842c39d4d8de7c34c971a420b64bd5f79eb6f4331c5bfb5cd97` |
 
 ## Reproducibility references
@@ -148,7 +151,9 @@ The conversion is a single dependency-free Python script,
 in the H3ddle repository, including the strength-0 self-check used to validate
 the pipeline against the official file.
 
-The FC2 sidecar is reproducible with
+The full optimized layout is reproducible with
+[`Scripts/repack-h3-input-major.py`](https://github.com/AlexanderIstomin/h3ddle/blob/main/Scripts/repack-h3-input-major.py).
+The legacy FC2 sidecar is reproducible with
 [`Scripts/optimize-h3-fc2-sidecar.py`](https://github.com/AlexanderIstomin/h3ddle/blob/main/Scripts/optimize-h3-fc2-sidecar.py).
 
 ## Contact

minimax_h3_ref2va_pruned_turbo_int8_convrot_input_major.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ca6696fe1cd9a8f254594ac67ee541f151b2377735dea3557364bd868270463
+size 20970380012