Qwen3 Asr 1.7b Onnx FP16

--- license: apache-2.0 library_name: sherpa-onnx pipeline_tag: automatic-speech-recognition tags: - sherpa-onnx - onnx - fp16 - qwen3-asr - speech-recognition --- # Qwen3-ASR-1.7B — sherpa-onnx fp16 export An **fp16 ONNX** export of [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B), packaged for [sherpa-onnx](https://github.com/k2-fsa/sherpa-onnx) (`OfflineRecognizer.from_qwen3_asr`). Intended for **GPU** inference (CUDA execution provider): ~4.5 GB on disk / ~4–5 GB VRAM, roughly half the fp32 export, with full-precision accuracy (fp16 is numerically lossless for this model — validated equal to fp32 to the decimal). > **Why fp16, not int8?** Dynamic INT8 quantization severely degrades this 1.7B model > (e.g. FLEURS Vietnamese WER 5.9 → 18.3). fp16 preserves accuracy; int8 is not offered here. ## Files ``` conv_frontend.onnx # audio frontend (fp32) encoder.onnx (+ .data) # fp16 decoder.onnx (+ .data) # fp16 tokenizer/ # copied unchanged from the base checkpoint ``` ## Usage (sherpa-onnx, Python) ```python import sherpa_onnx rec = sherpa_onnx.OfflineRecognizer.from_qwen3_asr( conv_frontend="conv_frontend.onnx", encoder="encoder.onnx", decoder="decoder.onnx", tokenizer="tokenizer", provider="cuda", # fp16 targets GPU num_threads=2, sample_rate=16000, max_total_len=4096, # defaults (512/128) silently truncate long turns max_new_tokens=1024, ) s = rec.create_stream() s.accept_waveform(16000, audio_f32) # 16 kHz mono float32 rec.decode_stream(s) print(s.result.text) ``` ## How this export was produced 1. Official `Qwen/Qwen3-ASR-1.7B` safetensors → ONNX (fp32) via the pipeline at [Wasser1462/Qwen3-ASR-onnx](https://github.com/Wasser1462/Qwen3-ASR-onnx) (conv_frontend, encoder, decoder graphs). 2. fp32 → fp16 with **`onnxruntime.transformers.optimizer.optimize_model(model_type="qwen3")`**, which fuses the RMSNorm/attention primitives into fused ops with correct fp16 kernels, then `OnnxModel.convert_float_to_float16(keep_io_types=True)`. (Fusing before converting is required — the standalone float16 converters produce an unloadable graph on the raw RMSNorm primitives.) No fine-tuning or weight modification beyond the fp32→fp16 cast. License: Apache-2.0 (inherited from the base model). See `LICENSE` and `NOTICE`.