Skip to content
Other

Qwen3-8B-5layer-MoE-16E2A-v1

компактный DSpark-draft с MoE-FFN

1 min readupdated

Это draft-модель для спекулятивного декодирования Qwen3-8B от yushengsu, собранная по рецепту SpecForge PR #1031. В ней 5 decoder-слоёв, GQA 32 heads / 8 KV heads и MoE-FFN: 16 экспертов по 768, top-2. Для serving нужен MoE serving package из PR #1016.

Ключевые особенности

  • DSpark-draft для Qwen3-8B: всего 5 decoder-слоёв.
  • MoE-FFN: 16 экспертов × 768 на слой, top-2; dense FFN на 12 288 разделён на 16 частей.
  • DSpark projector поверх target-слоёв 1/9/17/25/33, block size 16.
  • Обучение на PerfectBlend regen: 1 420 903 промпта, 6 эпох, 64 586 шагов, 0.1 CE + 0.9 L1.
  • Экспорты каждые 1 000 шагов и в концах эпох ep1–ep6; нормализованы под SGLang 0.5.18.

Пример

Сценарий: в SGLang выбранный checkpoint подключается как draft к Qwen3-8B. Например, на step-15000 в логе acc 0.7665 и walk accept length 6.091 — по карточке это лучший walk accept length среди приведённых шагов.

ПодробнееЛучший лог: step-15000 — acc 0.7665, ce_loss 0.9047, walk accept length 6.091.
Обучение: одна нода 8× H200, 2 SGLang capture server + 6 DDP trainer; MoE load-balance loss 0.005, expert bias 0.001/шаг.
Это не самостоятельная LLM, а 5-слойный draft; стандартных бенчмарков в карточке нет.
Discuss Telegram 1

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 2.1 GB · ready files2.8 GB
  • FP8 / INT8recommendedweights 1.0 GB1.7 GB
  • 4-bitweights 0.6 GB1.1 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.

    # yushengsu/Qwen3-8B-5layer-MoE-16E2A-v1
    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 yushengsu/Qwen3-8B-5layer-MoE-16E2A-v1 --local-dir ~/models/Qwen3-8B-5layer-MoE-16E2A-v1
    # 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

  1. Qwen3-8B-Basefamily root
  2. Qwen3-8B
  3. Qwen3-8B-5layer-MoE-16E2A-v1

New posts about versions of Qwen3-8B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face

About the model

yushengsu/Qwen3-8B-5layer-MoE-16E2A-v1
Author
yushengsu
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

01

Ask AI about this model

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

Comments

Computer vision

GALA-0.7B

GALA-0.7B