# Modification record

## Source

- Repository: `nvidia/parakeet-tdt_ctc-0.6b-ja`
- Revision: `44edb27eea9317daf89333e75eb830db4b1cc298`
- Checkpoint SHA-256: `a2ed2aab9c82cce4e9b699dca669184d7525c7b5e1eb8c3f7b789421bc3637e8`

## Export

1. The pinned NeMo checkpoint was loaded in FP32.
2. The TDT branch was exported to an encoder and fused decoder/joint graph.
3. The CTC branch was exported as one monolithic graph.
4. The standalone TDT encoder export was rejected after numerical validation
   against the PyTorch encoder.
5. A faithful shared encoder was extracted from the CTC graph at
   `/Transpose_2_output_0` with the encoded length from `/Cast_output_0`.
6. The extracted encoder was validated against the PyTorch encoder.
7. The CTC projection, transpose, and log-softmax were extracted into a small
   `ctc-head-model.onnx` graph.

Export environment:

- NeMo: `3.1.0+b331a3488`
- PyTorch: `2.12.0+cpu`
- ONNX: `1.22.0`

## Quantization

Only constant-weight `MatMul` operations in the shared encoder were quantized:

```text
onnxruntime.quantization.quantize_dynamic(
    op_types_to_quantize=["MatMul"],
    weight_type=QInt8,
    per_channel=True,
    use_external_data_format=True,
)
```

- Quantization ONNX Runtime: `1.27.0`
- Runtime validation ONNX Runtime: `1.24.4`
- Execution provider: CPUExecutionProvider
- TDT decoder/joint: exported FP32, not quantized
- CTC head: exported FP32, not quantized
- Vocabulary: unchanged from the pinned export

The distribution omits the redundant monolithic CTC graph. Both TDT and CTC
reuse the same quantized encoder.

No fine-tuning was performed.