Skip to content
Other

Qwen3-Omni-30B-A3B thinker DFlash

драфт для спекулятивного декодирования

2 min read

Это DFlash-драфт для thinker-части Qwen3-Omni-30B-A3B-Instruct — 48-слойной MoE-модели, которая отвечает на текстовые, аудио, видео и image-промпты. Автор дообучил z-lab/Qwen3-Coder-30B-A3B-DFlash на ответах самой thinker-модели: около 170 тыс. текстовых и аудиопромптов, плюс train-сплит GSM8K. Драфт использует эмбеддинги и output head целевой модели, поэтому отдельной головы у него нет.

Ключевые особенности

  • Спекулятивное декодирование: block size 8, до 7 предложений за раунд, некаузально внутри блока.
  • 8 слоёв Qwen3: hidden 2048, intermediate 6144, 32 головы / 4 KV-головы, head_dim 128.
  • Опирается на выходы слоёв 1, 12, 23, 34 и 45 целевой 48-слойной модели; контекст — 65 536 позиций.
  • Дообучение на 170 740 промптах и 169 681 ответе thinker-модели; аудио — транскрипция, суммаризация, перевод и другое.
  • Запускается через sglang-omni с DFLASH; точность на проверках не меняется.

Пример

python -m sglang_omni.cli serve --model-path Qwen/Qwen3-Omni-30B-A3B-Instruct --text-only --port 8000 \
  --thinker.engine.speculative_algorithm DFLASH \
  --thinker.engine.speculative_draft_model_path wilsonzheng0327/Qwen3-Omni-30B-A3B-Instruct-thinker-dflash
ПодробнееНа H100, greedy, sglang-omni: LibriSpeech test-clean c1 — 12,9 клипа/с против 6,9 без драфта (1,86x); MT-Bench c16 — 1619 tok/s против 1052 (1,54x); GSM8K c1 — 449–450 tok/s (1,85x).
Нужен bf16 и память под драфт; для видео-промптов требуется --thinker.engine.mem_fraction_static 0.90, иначе ~14k-токенные промпты отклоняются.
Ограничения: открытый чат и письмо дают около 1,1x при c1, короткие ответы и видео почти не ускоряются; в обучении не было изображений и видео, речь на выходе с драфтом не поддерживается. Замеры только на H100.
Discuss Telegram 0 24

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.9 GB · ready files1.5 GB
  • FP8 / INT8recommendedweights 0.4 GB1.0 GB
  • 4-bitweights 0.2 GB0.8 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # wilsonzheng0327/Qwen3-Omni-30B-A3B-Instruct-thinker-dflash
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download wilsonzheng0327/Qwen3-Omni-30B-A3B-Instruct-thinker-dflash --local-dir ~/models/Qwen3-Omni-30B-A3B-Instruct-thinker-dflash
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of Qwen3-Coder-30B-A3B-DFlash (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 2All on Hugging Face
Fine-tunes 1All on Hugging Face

About the model

wilsonzheng0327/Qwen3-Omni-30B-A3B-Instruct-thinker-dflash
Author
wilsonzheng0327
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

240

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments