Skip to content
Other

HSpec-Qwen27B-4m3a-epoch1

драфтер, который ускоряет Qwen3.8-27B

2 min readupdated

Это промежуточный чекпоинт H-Spec-драфтера для Qwen/Qwen3.8-27B: небольшой модели на 3.52B параметров, которая предлагает токены, а основная 27B-модель их проверяет. Так работает speculative decoding — генерация становится быстрее без смены основной модели. Архитектура — mamba_attn_hybrid; на инференсе драфтер читает KV-кэш верификатора прямо в процессе, без отдельного extraction-сервера. Снимок сделан в середине 1-й эпохи (шаг 29 470, около 38% первой эпохи из двух). Пригодится тем, кто уже обслуживает Qwen3.8-27B и хочет попробовать ускорение через спекулятивное декодирование.

Ключевые особенности

  • Драфтер на 3.52B параметров для Qwen3.8-27B, обученный по рецепту H-Spec.
  • Архитектура mamba_attn_hybrid и частичный RoPE: theta 1e7, partial_rotary_factor 0.25, head_dim 256.
  • Обслуживается на depth 8; заявлены loss 0.458, accept rate 0.642 и accept length 4.07.
  • Требуется форк vLLM: stock-версия не содержит метод mamba_attn_hybrid и путь partial-RoPE.

Что можно сделать

  • Поднять сервер vLLM с Qwen3.8-27B и этим драфтером, указав --spec-method mamba_attn_hybrid и --spec-tokens 8.
  • Запустить оценку throughput на HumanEval через scripts/evaluate/evaluate.py из speculators.
  • Использовать launcher scripts/launch_vllm.py, чтобы собрать аргументы и сохранить provenance эксперимента.

Пример

Сначала нужны форки vLLM и speculators из карточки; затем можно скачать чекпоинт и запустить сервер:

huggingface-cli download MeganEFlynn/HSpec-Qwen27B-4m3a-epoch1 --local-dir hspec-ckpt
vllm serve Qwen/Qwen3.8-27B \
  --spec-model hspec-ckpt \
  --spec-method mamba_attn_hybrid \
  --spec-tokens 8 \
  --max-model-len 8196 \
  --attention-backend FLASHINFER
ПодробнееМетрики снимка: loss 0.458, accept rate 0.642, accept length 4.07.
Валидация при depth 8 измеряет acceptance до позиции 7: 0.822 / 0.728 / 0.659 / 0.604 / 0.559 / 0.520 / 0.486.
Нужны точные ветки/коммиты форка vLLM и speculators; optimizer state в репозиторий не включён.
Это промежуточный чекпоинт, а не финальная модель.
Discuss Telegram 0 48

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 6.6 GB · ready files7.7 GB
  • FP8 / INT8recommendedweights 3.3 GB4.1 GB
  • 4-bitweights 1.8 GB2.5 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 5060 Ti 8GB

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.

    # MeganEFlynn/HSpec-Qwen27B-4m3a-epoch1
    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 MeganEFlynn/HSpec-Qwen27B-4m3a-epoch1 --local-dir ~/models/HSpec-Qwen27B-4m3a-epoch1
    # 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 Qwen3.8-27B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 2All on Hugging Face

About the model

MeganEFlynn/HSpec-Qwen27B-4m3a-epoch1
Author
MeganEFlynn
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

480

Ask AI about this model

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

Comments

Computer vision

Artsy RLHF

Artsy RLHF