Skip to content
Computer vision

Qwen3.8-27B FCF ck250

файнтюн для ускорения кода

2 min readupdated

Это Qwen/Qwen3.8-27B, дообученная через GSPO reinforcement learning с LoRA на 958 упражнениях по оптимизации кода: 533 на Python и 425 на C++. Адаптер влит в bf16-веса базовой модели, архитектура, токенизатор и chat template не менялись.

Ключевые особенности

  • База — Qwen3.8-27B на 27.78B параметров; модель заявлена к запуску как есть в vLLM.
  • Обучение — GSPO RL на задачах code-to-code оптимизации, description-to-code и оптимизации на уровне репозитория.
  • LoRA — rank-stabilized, rank 16, α 4 или 2, без dropout, на всех 64 текстовых слоях.
  • Vision tower и MTP head не адаптировались — изменения только в текстовой части.
  • Адаптер merged в bf16-веса со stochastic rounding (seed 850); checkpoint 250.

Пример

Сценарий из карточки: подайте функцию на Python или C++ и попросите ускорить её без потери корректности. Модель тренировали на code-to-code оптимизации, description-to-code и оптимизации на уровне репозитория — например, на задачах ParEval, PolyBench и EffiBench-X.

ПодробнееБенчмарки: PolyBench Python — speedup 7.33× против 5.88× у базы, корректность 92.7% против 86.0%; EffiBench-X Python Pass@1 — 89.8% против 87.7%; SWE-fficiency Lite — speedup 9.02× против 8.50×.
Требования к VRAM в карточке не указаны; это 27.78B-модель с bf16-весами, заявлен запуск в vLLM.
Ограничения: адаптированы только текстовые слои, vision tower и MTP head не менялись; замеры при reasoning effort low.
Детали обучения: GSPO RL, LoRA, checkpoint 250, 8 completions на промпт (64 за шаг), stochastic rounding seed 850.
Discuss Telegram 0 33

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 52 GB · ready files57 GB
  • FP8 / INT8recommendedweights 26 GB29 GB
  • 4-bitweights 15 GB17 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.

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.

    # fastcoder-factory/qwen3.8-27b-FCF-ck250
    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 fastcoder-factory/qwen3.8-27b-FCF-ck250 --local-dir ~/models/qwen3.8-27b-FCF-ck250
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to qwen3.8-27b-FCF-ck250
  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

fastcoder-factory/qwen3.8-27b-FCF-ck250
Author
fastcoder-factory
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

330

Ask AI about this model

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

Comments