Skip to content
Other

Lemma-2.2M-init2

микро-LLM для арифметики на токенах

1 min readupdated

Lemma-2.2M-init2 — экспериментальная микро-LLM на 2,2 млн параметров от oscar128372. Она создана для проверки числовых рассуждений на уровне токенов и гибридного разреженно-рекуррентного внимания при жёстких ограничениях по параметрам и вычислениям. Претрейн с нуля шёл на одной NVIDIA T4.

Ключевые особенности

  • 2,2 млн параметров: 3 слоя UnifiedTransformerBlock (d_model=192, 4 головы, tied embeddings, logit softcapping 30.0).
  • Токенизатор на 2048 BPE-токенов с regex-изоляцией: каждая цифра и операторы + - * / = ( ) — отдельные токены.
  • Residual Pyramid Attention: локальное окно W=8 с Multi-Octave RoPE, retrieval через SRVQ и Triton-ядро, глобальные регистры G=8.
  • MSCSM и FFN: gated depthwise causal conv с тремя dilated-ветвями; Harmonic Differential StarCarrier FFN.
  • Обучение с нуля: 1× NVIDIA T4, ~6 часов претрейна (30 000 шагов, ~1,97 млрд токенов) и ~1 час посттренинга.

Пример

Чтобы загрузить веса, нужен класс UnifiedLM из modeling.py, CUDA-совместимая GPU и установленный triton: модель использует кастомные модули и Triton-ядро build_causal_index_fast_kernel.

ПодробнееБенчмарки zero-shot: HellaSwag 27.91%, ARC-Easy 34.34%, PIQA 54.95%, ArithMark-3.0 39.50% (acc_norm).
Intelligence Index Open SLM Leaderboard: +8.19.
Контекст — 512 токенов; это research preview, не для продакшена и фактической надёжности.
Для загрузки нужны CUDA-совместимая GPU и triton.
Discuss Telegram 1 3

Will it run?

Runs in F16

Fits without quantisation.

  • F16weights 0.0 GB · ready files0.5 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 · F16 — 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.

    # oscar128372/Lemma-2.2M-init2
    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 oscar128372/Lemma-2.2M-init2 --local-dir ~/models/Lemma-2.2M-init2
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Lemma-2.2M-init2
  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 Lemma-2.2M-init2 (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

oscar128372/Lemma-2.2M-init2
Author
oscar128372
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

31

Ask AI about this model

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

Comments

Other

d3-edge

d3-edge