Skip to content
Other

Qwen3.8-27B Sparse F2

разреженный студент для локального чата

2 min read

Это per-token sparse student модели Qwen3.8-27B (HauhauCS Uncensored-Aggressive, Q6_K_P GGUF) от автора verysweatydog. Студент сохраняет все 64 слоя, но каждый токен читает только часть весов: низкоранговые проекции attention/GDN, 1-битный sketch-генератор кандидатов для FFN и shortlist в LM head. Есть готовый GGUF Q4_K_M на 20,9 ГБ (3,82 BPW) и патч llama.cpp.

Ключевые особенности

  • Готовый к запуску qwen3.8-27b-sparse-F2-Q4_K_M.gguf на 20,9 ГБ; нужен патченный llama.cpp.
  • Разреженность настраивается GGUF-ключами без переэкспорта: ffn_candidates, gdn_gate_k, ffn_active, head_shortlist.
  • Декод батчами до 8 токенов идёт по sparse-путям; prefill — по dense-путям с теми же весами.
  • Обучение — дистилляция по логитам учителя на ~12M токенов, 1500 шагов; данные очищены от GSM8K test.
  • Модель uncensored, без дополнительного safety tuning; автор ещё не запускал её на реальном iGPU.

Пример

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
git checkout f477f79
git apply /path/to/llama.cpp-sparse-qwen35.patch
cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j --target llama-cli
./build/bin/llama-cli -m qwen3.8-27b-sparse-F2-Q4_K_M.gguf -ngl 99 -cnv
ПодробнееDev KL 0.2268 и dev top-1 85.6%; math KL 0.1648 и top-1 88.4% против учителя.
С ffn_candidates=8192 и gdn_gate_k=2048: dev KL 0.1974, math KL 0.1461, но +16% байт на токен.
GSM8K: 115/128 (89.8%) — столько же, сколько у учителя.
20,9 ГБ GGUF + KV-кэш помещается в 32 ГБ; 16 слоёв full attention, 48 DeltaNet с фиксированным state.
Prefill стоит примерно как у dense-модели; экономия — на генерации.
Discuss Telegram 0

Will it run?

Runs in GGUF Q4_K_M

The most precise of the ready quantised files fits.

  • GGUF Q4_K_Mweights 19 GB · ready files22 GB

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: GGUF Q4_K_M — from RTX 4090

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.

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

verysweatydog/qwen3.8-27b-moe
Author
verysweatydog
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

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

Comments