---
license: cc-by-4.0
language:
- en
- de
- es
- fr
pipeline_tag: automatic-speech-recognition
tags:
- automatic-speech-recognition
- onnx
- directml
- nemo
- canary
base_model:
- nvidia/canary-1b-flash
---

# NVIDIA canary-1b-flash — DirectML-safe ONNX

ONNX export of [nvidia/canary-1b-flash](https://huggingface.co/nvidia/canary-1b-flash) whose **encoder is re-exported
in the TorchScript idiom** (`torch.onnx.export(dynamo=False)`), so the model runs on the ONNX Runtime
**DirectML EP** (and every other EP).

## Why

istupakov's `canary-1b-flash-onnx` encoder is a torch-**dynamo** export. On the DirectML EP it is a two-sided
trap (isolated by graph bisection): dynamic shapes crash the Reshape/attention kernels
(`MLOperatorAuthorImpl.cpp:2597`, then a D3D12 device-removal `887A0020`), and forcing static shapes
fails session creation in `InferAndVerifyOutputSizes` (`:2853`) — both unfixed ORT-DML defects around
the dynamo `view` idiom (upstream onnxruntime #26826 / #26944; the DML EP is in maintenance mode).
NVIDIA's Parakeet FastConformer, exported via TorchScript, runs fine on DML — so this repo re-exports
the **same encoder** the same way.

## What changed

- `encoder-model.onnx` (+ `encoder-model.int8.onnx`): re-exported from the `nvidia/canary-1b-flash` NeMo
  checkpoint via `torch.onnx.export(dynamo=False, opset=17)`, same I/O contract as istupakov
  (`audio_signal[B,128,T]`, `length[B]` → `encoder_embeddings[B,S,1024]`, `encoder_mask[B,S]`).
  Numerically identical to istupakov's encoder on CPU (max|Δ| ≈ 4e-6 — export-tracer float noise).
- `decoder-model.onnx` / `decoder-model.int8.onnx` / `config.json` / `vocab.txt`: **unchanged** from
  [istupakov/canary-1b-flash-onnx](https://huggingface.co/istupakov/canary-1b-flash-onnx) — the DML crash was
  encoder-only; the AED decoder is byte-for-byte the same.

Produced by [WinSTT](https://github.com/dahshury/WinSTT)'s `canary_encoder_export.py`.