Skip to content
Other

Jasper-K-27B

RL-файнтюн Qwen для оптимизации GPU-ядер

1 min readupdated

Jasper-K-27B — 27B-модель от Ragentile, полученная RL-постобучением Qwen3.8-27B с reasoning effort medium. Она генерирует корректные реализации CUDA-ядер и пытается ускорять их через многотуровое взаимодействие со средой выполнения.

Ключевые особенности

  • Multi-rollout, multi-turn GRPO: модель получает фидбек о компиляции, корректности и производительности и дорабатывает код.
  • Награда за улучшение подталкивает не останавливаться на первом корректном ядре, а искать более быстрые варианты.
  • Обучение на 180 задачах KernelBenchFixed L1–L2; ещё 20 задач — для мониторинга и выбора чекпоинтов.
  • 50 задач LeetCuda от команды RL из University of Alberta используются только для теста.
  • В комплекте веса и harness для multi-turn/multi-rollout инференса; бэкенд оценки нужно подключать свой.

Пример запуска

export CUDAMASTER_GPUS=0,1,2,3,4,5,6,7
export JASPER_K_EVALUATOR_ENDPOINT=http://127.0.0.1:9000
python3 harness.py --problems /absolute/path/to/problems --out ./run_out --backend cuda --samples 16 --timeout 7200
ПодробнееБенчмарки из блога: KernelBench 2.289× → 3.471× (+51.64%), LeetCuda 2.161× → 2.580× (+19.39%) в METR Best GM Speedup @16×4.
Дефолтное развёртывание — 2 реплики × TP4, всего 8 устройств; контекст сервера 122 880 токенов, до 4 туров на траекторию.
Нужны Python 3.11+, vLLM/Transformers и собственный evaluator через адаптер: внутренний CudaEval не поставляется.
Без evaluator endpoint каждая траектория делает только один тур. Лицензия — Apache 2.0.
Discuss Telegram 5

Will it run?

Runs in FP8 / INT8

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

  • BF16weights 50 GB · ready files56 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 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

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.

    # ragentile/Jasper-K-27B
    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 ragentile/Jasper-K-27B --local-dir ~/models/Jasper-K-27B
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Jasper-K-27B
  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

ragentile/Jasper-K-27B
Author
ragentile
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

05

Ask AI about this model

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

Comments