jiangzhuo9357/Qwen3-ASR-1.7B-ONNX

Qwen3-ASR-1.7B — ONNX for the browser (onnxruntime-web / WebGPU), layout v2

ONNX export of Qwen/Qwen3-ASR-1.7B (Apache-2.0) packaged for in-browser inference with onnxruntime-web, built for the Sokuji local-inference lane. Same layout v2 as jiangzhuo9357/Qwen3-ASR-0.6B-ONNX (2026-09-03): no embedding table inside any graph, one shared weights file per precision, int8 embedding table, and a prompt_config.json that carries every constant a client needs — a client that runs the 0.6B runs this model unchanged by reading the dims from that file. Tooling and measurements: benchmark/qwen3-asr-webgpu/ in the Sokuji repo (results/1.7b-notes.md for the 0.6B comparison).

Pick a variant

Shared by both: embed_tokens.int8.bin (311 MB) + embed_scales.f32.bin (0.6 MB), prompt_config.json, mel_filters.json, tokenizer.json, tokenizer_config.json, vocab.json, added_tokens.json, config.json — about 0.33 GB, so the total download is ≈ 2.7 GB (q4) or ≈ 2.0 GB (q4f16). prompt_config.json → variants lists the same file roles machine-readably. Decoders are MatMulNBits int4, RTN, block 32, accuracy level 4; q4f16 additionally runs activations and I/O in fp16 with RMSNorm / softmax / rotary kept in fp32. Without WebGPU the wasm execution provider is far too slow for live use. Note that Chrome's Vulkan adapters on Linux (NVIDIA) and on the GB10 expose no shader-f16, so those get q4; Windows (D3D12) and macOS (Metal) get q4f16.

Graph contracts

• —encoder.*: mel [1, 128, T] fp32 → audio_features [1, A, 2048] fp32 (fp32 I/O for both encoders).
• —decoder_init.*: input_embeds [1, S, 2048], position_ids [1, S] int64 → logits [1, 1, 151936] (last position only), present_keys / present_values [28, 1, 8, S, 128].
• —decoder_step.*: input_embeds [1, 1, 2048], position_ids [1, 1], past_keys, past_values → logits [1, 1, 151936], present_* [28, 1, 8, S+1, 128].
• —The KV cache has the same shape as the 0.6B's (28 layers, 8 kv heads, head dim 128): 224 KB per token in fp32, 112 KB in fp16. Only the hidden size (2048 vs 1024) and the encoder (24 layers, d 1024) differ.
• —q4f16 graphs use fp16 for every float tensor above; q4 uses fp32. Both decoders of a variant reference the same decoder_weights.*.data; hand onnxruntime-web the buffer once (externalData: [{ path, data }]) for each session.
• —Keep the KV cache on the GPU between steps (preferredOutputLocation: { present_keys: 'gpu-buffer', present_values: 'gpu-buffer' }) and dispose the superseded tensors.

Building the prompt (see prompt_config.json)

Identical to the 0.6B card; prompt_config.json carries the numbers.

1. 1.Log-mel, Whisper-compatible: 16 kHz, nfft 400, hop 160 (periodic Hann, centred, reflect padding), power spectrum, the 128-bin Slaney filterbank in `melfilters.json, log10 with a 1e-10 floor, clamp to max − 8, (x + 4) / 4`, last frame dropped.
2. 2.Audio tokens for T mel frames: A = conv(conv(conv(T mod 100))) + 13 · floor(T / 100) with conv(t) = floor((t + 1) / 2).
3. 3.Prompt ids: prompt.prefix_ids + A × prompt.audio_pad_id + prompt.suffix_ids.
4. 4.Embeddings: row id of the int8 table times embed_scales[id] → fp32 (or fp16 for q4f16); overwrite the A rows at the pad positions with the encoder output. That tensor is input_embeds for decoder_init.
5. 5.Greedy decode with decoder_step until <|endoftext|> (151643) or <|im_end|> (151645); max_new_tokens 256.
6. 6.Output begins with language <Name><asr_text>; drop everything up to and including asr_text_id (151704). When the language is known, append language_prefix_ids[<iso>] to the prompt (same ids as for the 0.6B — same tokenizer).

Validation

• —v2 FP32 graphs are token-for-token identical to the v1 FP32 export on English, Japanese and Chinese test clips.
• —int4 block 32 / q4f16: English and Chinese clips identical to FP32; the Japanese clip differs in one word (a hard phrase both FP32 and int4 get partly wrong).
• —Encoder fusion removes 26.5 % of encoder nodes and 30 % of decoder nodes; int8 embedding max dequantisation error 1.2e-3.

Measured in the browser (whole pipeline, warm, 13 clips: 8 ja, 4 zh, 1 en; medians)

Quality against the 0.6B on the same clips: Japanese CER 0.147 → 0.070 (the 1.7B fixes the proper-noun / kana slips of the 0.6B); English and Chinese were already near-perfect. The step loop is dispatch-bound on every GPU, which is why a decoder with ~2.8× the weights costs only 13–43 % more per token; prefill and the encoder scale closer to the parameter ratio.

Provenance

Exported with andrewleech/qwen3-asr-onnx (qwen-asr 0.0.6, transformers 4.57.6, torch 2.14), with the layout-v2 prefill wrapper (host-provided embeddings, last-position logits), the pipeline's optimize_graphs.py (RMSNorm → SimplifiedLayerNormalization, encoder BiasGelu / SkipLayerNormalization) and share_weights.py, onnxruntime's MatMulNBitsQuantizer (RTN, 4-bit, block 32, accuracy level 4) and float16 converter. Scripts: benchmark/qwen3-asr-webgpu/export_v2/ (size-agnostic since 2026-09-03).