tokihisu-qwen3-asr-0.6b-webgpu

Qwen3-ASR-0.6B WebGPU FP16 for TOKIHISU

This repository contains a browser-oriented ONNX conversion of Qwen/Qwen3-ASR-0.6B. The conversion source is the official native Transformers checkpoint Qwen/Qwen3-ASR-0.6B-hf at revision 7f1569a48a89f3e3f4dc3a5c9d28bddd903bc76c.

Files

• onnx/audio_encoder_fp16.onnx + external data: vectorized windowed audio encoder.
• onnx/decoder_with_past_fp16.onnx + external data: merged prefill/decode graph with KV cache.
• processor/: tokenizer, feature extractor and chat-template files.
• manifest.json: exact graph contract, hashes, revisions and conversion settings.

Runtime contract

• Audio features: 128-bin log-Mel input, padded to a multiple of 800 frames. Trim projected audio embeddings using audio_token_mask.
• Decoder: 28 layers, 8 KV heads, head dimension 128.
• Seed every KV input with one zero position. Keep cache position zero masked with -65504 for the whole utterance. This avoids zero-sized WebGPU tensors.
• Keep present.* outputs on GPU with ONNX Runtime Web preferredOutputLocation: 'gpu-buffer' or preallocated GPU tensors.
• The decoder expects an additive causal attention mask and audio embeddings aligned to the prompt's audio-token positions.

ONNX Runtime Web

Use onnxruntime-web/webgpu, explicitly select the webgpu execution provider, and pass the external-data file listed in manifest.json through the session externalData option. Use a current Chrome or Edge build whose WebGPU adapter exposes shader-f16.

License

The conversion code and generated metadata are MIT licensed. Converted model weights remain under the upstream Apache-2.0 license. Preserve the upstream attribution and license when redistributing this repository.