Skip to content
Other

Fitted R-lenses

линзы для скрытых активаций LLM

1 min read

Набор fitted R-lenses от автора tangang8: линзы отображают скрытые активации ранних слоёв модели в residual stream её финального transformer-слоя. После output normalization и unembedding эти активации превращаются в token predictions. В комплекте — варианты для Qwen3-4B/8B/14B, Llama-3.1-8B, Gemma-2-2B и Gemma-3-12B-it.

Ключевые особенности

  • R-линзы переводят активации fitted decoder block в residual space целевого слоя.
  • Сохраняют forward computation модели, меняя только backward rules при подгонке.
  • Поддержка Qwen3-4B/8B/14B, Llama-3.1-8B, Gemma-2-2B и Gemma-3-12B-it.
  • Каждый каталог содержит lens_fp32.pt и config.json с базовой моделью, ревизией чекпоинта и слоями.
  • Подгонка на WikiText-103 с лимитом 128 токенов; в сохранённых линзах от 426 до 676 пассажей.

Пример

from huggingface_hub import hf_hub_download
from jlens import JacobianLens

path = hf_hub_download(
    repo_id="tangang8/jspace-lenses",
    filename="qwen3-4b/rlens/lens_fp32.pt",
)
lens = JacobianLens.load(path)
ПодробнееФайлы содержат FP32-матрицы линз; веса базовой языковой модели загружаются отдельно.
Номера слоёв zero-based и относятся к выходам decoder block до финальной нормализации.
В карточке нет бенчмарков и требований к VRAM; для применения нужен соответствующий чекпоинт модели.
Discuss Telegram 0

Will it run?

Runs in FP32

Fits without quantisation.

  • FP32weights 3.8 GB · ready files4.7 GB
  • BF16weights 1.9 GB2.6 GB
  • FP8 / INT8recommendedweights 1.0 GB1.5 GB
  • 4-bitweights 0.5 GB1.1 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 · FP32 — 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.

    # tangang8/jspace-lenses
    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 tangang8/jspace-lenses --local-dir ~/models/jspace-lenses
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to jspace-lenses
  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 jspace-lenses (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

tangang8/jspace-lenses
Author
tangang8
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