Skip to content
Other

HSpec-Qwen27B

RoPE-исправленный драфтер для Qwen3.8-27B

1 min read

Спекулятивный драфтер H-Spec с архитектурой mamba_attn_hybrid для целевой модели Qwen/Qwen3.8-27B. Размер — 4,1 млрд параметров, обучен через vllm-project/speculators; глубина драфта 5, размер блока 6, словарь драфта 64k, KV-sharing на full-attention слоях верификатора 3/19/35/51/63. Чекпойнт заменяет предыдущую pre-RoPE-fix версию: Q/K-поворот драфтера теперь соответствует partial rotary целевой модели.

Ключевые особенности

  • Спекулятивное декодирование с глубиной драфта 5 и размером блока 6.
  • Обучен на 49,5k сэмплов OpenPerfectBlend (Qwen3.8-27B regenerated), 1 эпоха, 7×H100.
  • Исправлен RoPE: Q/K-поворот соответствует partial rotary целевой модели, включая neox pairing j↔j+32.
  • Hidden states и shared KV извлекались онлайн из целевой модели.
  • Запуск через vLLM с --spec-method mamba_attn_hybrid и --spec-tokens 5.

Пример

vllm serve Qwen/Qwen3.8-27B \
  --spec-model MeganEFlynn/HSpec-Qwen27B \
  --spec-method mamba_attn_hybrid \
  --spec-tokens 5 \
  --max-model-len 8196 --attention-backend FLASHINFER
ПодробнееВалидация: val loss 0.735, val accept rate 0.511, val EAL 2.998.
HumanEval acceptance length — 2.5590 (101 865 драфтов; позиции 0–4: 0.623 / 0.390 / 0.254 / 0.173 / 0.119).
Для запуска нужна Qwen3.8-27B и vLLM с поддержкой mamba_attn_hybrid; VRAM в карточке не указана.
Discuss Telegram 0 56

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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
    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 --local-dir ~/models/HSpec-Qwen27B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to HSpec-Qwen27B
  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 HSpec-Qwen27B (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

MeganEFlynn/HSpec-Qwen27B
Author
MeganEFlynn
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

560

Ask AI about this model

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

Comments

Computer vision

UAV-PUR-8B

UAV-PUR-8B