Skip to content
Other

FlexDM-LoRA-v2.1

Gemma-4-31B для решений по вариантам

1 min read

LoRA-адаптер от 1000luca для google/gemma-4-31B-it под структурированные решения: есть состояние (документ, диалог, список устройств и т.п.) и один или несколько multiple-choice / yes-no вопросов о нём. Модель не генерирует ответ, а выдаёт вероятность каждого варианта из логитов меток на Answer: (. Адаптер рассчитан на запуск в vLLM; served-версия имеет rank 128 и strength 0.7.

Ключевые особенности

  • Обучение на 180 101 вопросе; в v2.1 добавлено около 65 тыс. вопросов для форматов, слабых в v2.0.
  • Два прогона обучения усреднены; served rank 128, strength 0.7.
  • Ответ — распределение вероятностей по вариантам, без генерации текста.
  • Поддержка 2–255 вариантов и до 131 071 токена на строку; более длинные запросы отклоняются.
  • Для vLLM на 10 full-attention слоях k_proj LoRA продублирован как v_proj, чтобы V получал обученное изменение.

Пример

python -m decision_index run --engine jevlab.vllm_lora_reader:VllmLoraReader --option lora=/path/to/this/repo --option layout=header --option read_twice=true --option twice_max_tokens=3000 --option temperature_choice=10.9757 --option temperature_choice_log_slope=-1.7016 --compact --out runs/flexdm-lora
ПодробнееDecision Index 0.3 public: 68.99 (v2.0 — 66.49).
Для запуска в vLLM нужно около 75 ГБ свободной VRAM на одной GPU; база квантуется FP8 per-block.
Ограничения: 131 071 токен на строку и 2–255 вариантов; длинные строки отклоняются без усечения.
Лицензия адаптера — CC BY-SA 4.0; базовая Gemma — Apache 2.0.
Discuss Telegram 0 23

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 62 GB · ready files69 GB
  • FP8 / INT8recommendedweights 33 GB37 GB
  • 4-bitweights 20 GB23 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.

This is a LoRA: counted together with its base model gemma-4-31B-it.

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.

    # 1000luca/FlexDM-LoRA-v2.1
    # ===== 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 google/gemma-4-31B-it --local-dir ~/models/gemma-4-31B-it
    hf download 1000luca/FlexDM-LoRA-v2.1 --local-dir ~/models/FlexDM-LoRA-v2.1
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to gemma-4-31B-it
  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

Requires the base model

google/gemma-4-31B-it

This model is a LoRA / adapter. Download the base model first, then load this one on top.

  1. gemma-4-31Bfamily root
  2. gemma-4-31B-it
  3. FlexDM-LoRA-v2.1

New posts about versions of gemma-4-31B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
Merges 3All on Hugging Face

About the model

1000luca/FlexDM-LoRA-v2.1
Author
1000luca
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

230

Ask AI about this model

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

Comments