Skip to content
Other

Gemma 4 E2B

multi-seed QLoRA: пайплайн для медицинского SFT

2 min read

Репозиторий-эксперимент от adityag-india: обвязка для multi-seed QLoRA на Gemma 4 E2B в 4-bit NF4 на медицинских чат-данных в формате user -> assistant JSONL. Это не готовый чекпоинт, а пайплайн со стадиями preflight → model gate → smoke → все сиды → агрегация. Собран и проверен на RunPod с RTX A6000 48 GB, PyTorch 2.8 / cu128.

Ключевые особенности

  • Гейты падают до траты GPU-времени: проверка torch/CUDA, реальный kernel-проб (matmul, SDPA, bitsandbytes NF4), схема и sha256 файлов данных, chat-template.
  • Повторный запуск резюмируется: завершённые сиды пропускаются, прерванный сид продолжается с последней законченной эпохи.
  • По умолчанию 5 сидов по 10 эпох; агрегация считает mean, std, min, max и 95% CI и отказывается смешивать конфиги.
  • Оценка в отдельном процессе: test loss / perplexity / token accuracy, ROUGE-1/2/L на фиксированных 200 примерах и строка сравнения с базовой моделью.
  • Кастомный train loop вместо Trainer: loss токен-точный по всему окну накопления, отдельный guard от FP32 upcast.
  • В preflight порог ≥20 GiB GPU, только LoRA-параметры обучаемы, base NLL sanity <6 nats/token.

Пример

bash setup_runpod.sh
export HF_TOKEN=hf_xxx
cp /path/to/*train*.jsonl data/
tmux new -s exp
python run_experiment.py --stage all 2>&1 | tee -a outputs/run.log
ПодробнееГиперпараметры: LoRA r=32, alpha=64, dropout 0.05 на q,k,v,o,gate,up,down; lr 1e-4, cosine, max_length 1536, эффективный batch 8.
Обучение шло на 48 GB A6000, но те же настройки воспроизводятся на 22 GiB L4 при micro-batch 1×8.
Сид управляет инициализацией LoRA, dropout и порядком данных; бит-репродуцируемости нет из-за bf16/4-bit.
Готовых метрик и бенчмарков в карточке нет — публикуется пайплайн, а не результаты.
Discuss Telegram 0

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.

    # adityag-india/gemma4-e2b-medical-multiseed
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    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 adityag-india/gemma4-e2b-medical-multiseed --local-dir ~/models/gemma4-e2b-medical-multiseed
    # 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 gemma4-e2b-medical-multiseed (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

adityag-india/gemma4-e2b-medical-multiseed
Author
adityag-india
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