dkhokhlov/whisper-tiny-hqq-4bit

HQQ 4-bit Whisper-Tiny

Includes fp16 + fp32 ONNX exports — the same HQQ model runs on CPU ONNX Runtime with no HQQ runtime. The fp16 export is a fp16-only graph (lower RAM, slower on CPU-ORT); the fp32 export is the recommended CPU compute and matches the published benchmark. HQQ ships no ONNX exporter; this repo adds one (see ONNX export).

Model card source for dkhokhlov/whisper-tiny-hqq-4bit.

Related models

• —`dkhokhlov/whisper-base-hqq-4bit` — HQQ 4-bit, whisper-base (CPU eval)
• —`dkhokhlov/whisper-small-hqq-4bit` — HQQ 4-bit, whisper-small (A10 GPU eval)
• —Source model: `openai/whisper-tiny` (fp32)
• —Benchmark + code: `dkhokhlov/whisper-cascade`

Summary

`openai/whisper-tiny` quantized with HQQ 4-bit grouped quantization for CPU inference. Resident weight RAM (fp16 compute) is 57.53 MB, 23.8% smaller than the unquantized fp16 model (75.52 MB). fp16 compute is WER-neutral; the published WER benchmark uses fp32 compute for cross-model comparability. The key setting is mixed precision: the whole encoder stack and fc1 are kept at 8-bit, the remaining decoder linears are 4-bit (see the repo README for the full config and the config-sweep ablation).

English (fleurs en_us, n=100) WER is 0.1367 vs 0.1381 fp32 (-1.0%), within n=100 noise. HQQ is within 5% relative of fp32 on every tested config (5 fleurs + 4 talkbank).

Results

English (fleurs en_us, n=100, fp32 compute):

HQQ is within 5% relative of fp32 on every tested config. The full multilingual and telephone WER tables, the cross-reference against whisper-base/whisper-small, and the size-by-component breakdown are in the repo README.

Load and use

The model auto-detects the spoken language and transcribes (multilingual Whisper behavior). Pass language to force a language when it is known.

import hqq_asr
pipe = hqq_asr.build_pipeline("dkhokhlov/whisper-tiny-hqq-4bit", quant="hqq")
text = pipe({"array": audio, "sampling_rate": 16000})["text"]                       # auto-detect
text = pipe({"array": audio, "sampling_rate": 16000},
             generate_kwargs={"language": "spanish", "task": "transcribe"})["text"]   # force

Command line (this repository):

make asr MODEL_ASR=dkhokhlov/whisper-tiny-hqq-4bit QUANT=hqq AUDIO=clip.wav

ONNX export (CPU ONNX Runtime)

This repo ships two ONNX exports of the same HQQ model, differing only in compute dtype:

• —fp16 (default): `encoder_model.onnx` + `decoder_model_merged.onnx` — a fp16-only graph (zero fp32 ops). It uses eager attention so the attention scale stays a fp16 Mul (SDPA would decompose it to Sqrt→Div in fp32). On CPU ONNX Runtime it runs ~4.4× slower than the fp32 export (ORT-CPU upcasts fp16→fp32 internally — fp16 is not a primary CPU compute format, so the slowdown is expected) but loads ~29% less RAM.
• —fp32: `encoder_model-fp32.onnx` + `decoder_model_merged-fp32.onnx` — the recommended CPU compute and the benchmark compute. Faster on CPU ORT; matches the published fp32 WER benchmark.

Both keep the packed uint8 W_q and the per-group scale/zero as ONNX initializers and emit the unpack + dequant as standard ONNX ops (opset 18), so each graph carries the exact HQQ weights, not a re-dequantized dense copy. Whisper is an encoder-decoder model, so each export is two ONNX graphs; the autoregressive generation loop (argmax, KV-cache, EOS stop) runs in Python in ORTModelForSpeechSeq2Seq, calling the encoder once and the decoder once per token:

The merged decoder carries the no-past (first step) and with-past (cached steps) branches behind one control-flow switch, so one session handles the whole generation; the separate un-merged decoder files optimum emits are not shipped.

Both exports reproduce the HQQ WER (0.1367). The fp32 export exact-matches the HQQ manifest (0 mismatches); the fp16 export matches the WER and differs from the fp32 manifest by one case-only token (a fp16-vs-fp32 rounding effect, WER-neutral).

Load via ONNX Runtime (the no-suffix files are the fp16 default):

import hqq_asr
pipe = hqq_asr.build_pipeline("dkhokhlov/whisper-tiny-hqq-4bit", quant="onnx")
text = pipe({"array": audio, "sampling_rate": 16000})["text"]

Reproduce the fp16 export and the gate (set HQQ_COMPUTE_DTYPE=fp32 for the fp32 export):

make onnx HQQ_REPO=dkhokhlov/whisper-tiny-hqq-4bit ONNX_OUT=build/whisper-tiny-hqq-onnx-fp16
make hqq-reference HQQ_REPO=dkhokhlov/whisper-tiny-hqq-4bit EVAL_OUT=build/hqq_reference_tiny_fp16.json
make eval-onnx ONNX_OUT=build/whisper-tiny-hqq-onnx-fp16 \
     HQQ_REFERENCE_MANIFEST=build/hqq_reference_tiny_fp16.json EVAL_OUT=build/eval_onnx_tiny_fp16.json

The export spec and the two validation gates are in docs/onnx.md in the repo.

Reproduce

# 1. Quantize locally (writes whisper-tiny-hqq-4bit/).
python quantize.py

# 2. Measure baseline WER (fp32).
EVAL_LIMIT=100 MODEL_ASR=openai/whisper-tiny EVAL_CONFIG=en_us \
  EVAL_OUT=eval_baseline.json python eval_wer.py

# 3. Measure HQQ WER.
EVAL_LIMIT=100 QUANT=hqq MODEL_ASR=./whisper-tiny-hqq-4bit EVAL_CONFIG=en_us \
  EVAL_OUT=eval_hqq.json python eval_wer.py

# 4. Telephone benchmark (talkbank segment split).
EVAL_DATASET=diabolocom/talkbank_4_stt EVAL_CONFIG=en EVAL_SPLIT=segment EVAL_LIMIT=100 \
  MODEL_ASR=openai/whisper-tiny EVAL_OUT=talkbank_en_fp32.json python eval_wer.py

# 5. Publish (needs a Hugging Face write token).
PUSH=1 HQQ_REPO=dkhokhlov/whisper-tiny-hqq-4bit python quantize.py

License

MIT. Derived from `openai/whisper-tiny` (Apache-2.0) and HQQ. The quantized weights inherit the openai/whisper license terms.

Citation

See the repo README for the BibTeX entry.

Full details

Quantization config, config-sweep ablation, safetensors format, the full WER tables (multilingual fleurs, talkbank telephone, cross-reference), and the resident-RAM-by-component breakdown are in the repo README. Per-config WER evidence JSONs are committed under eval_multilingual/ and eval_telephone/ in `dkhokhlov/whisper-cascade`.