Skip to content
Other

Qwen3.8-27B ICA Lens

полная карта 64 слоёв для интерпретации LLM

2 min readupdated

Предвычисленный ICA Lens для Qwen/Qwen3.8-27B и FP8-варианта: 327 680 пар «слой–компонент» на всех 64 слоях при d_model = 5120. Это не отдельная LLM, а инструмент для интерпретации residual stream и словарного чтения; архитектура базы — 48 Gated DeltaNet linear-attention слоёв + 16 full-attention. Автор — lmxxf.

Ключевые особенности

  • Строгая двуязычная подгонка 1:1 EN/ZH: 200 704 токена WikiText-103 и китайской Wikipedia, проверка на 51 200 отложенных токенах.
  • Кросс-языковая полисемия: для apple и 苹果 независимо выбирается один и тот же axis ID в 42 из 64 слоёв.
  • Совместное чтение ICA-Lens × J-lens меняет 81,5% Top-5 токенов на 6 633 осях и собирает чистые синонимы вроде for / для / für.
  • 83,76% осей имеют excess kurtosis κ > 1; 44,08% — κ > 10.
  • Загрузка по слоям: по 200 MiB на слой, всего 12,50 GiB; layer_XX.safetensors подтягиваются по требованию.

Пример

from icalens import ICALens

lens = ICALens.from_pretrained('lmxxf/Qwen3.8-27B-ica-lens')
print(lens.available_layers)  # 0..63
h = ...  # [seq_len, 5120]
scores = lens.transform(h, layer=41)
ПодробнееНа 51 200 отложенных токенах: 83,76% осей с excess kurtosis κ > 1, 44,08% — с κ > 10.
Один и тот же axis ID для apple/苹果 выбран в 42 из 64 слоёв.
Совместное чтение ICA-Lens × J-lens меняет 81,5% Top-5 токенов на 6 633 осях.
Каждый слой — 200 MiB, всего 12,50 GiB; загрузка по требованию.
Ограничения: эвристическое чтение, 50 итераций FastICA не гарантируют сходимости; высокий kurtosis — не семантическая метка.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.2 GB · ready files0.7 GB
  • FP8 / INT8recommendedweights 0.1 GB0.6 GB
  • 4-bitweights 0.1 GB0.6 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 · BF16 — 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.

    # lmxxf/Qwen3.8-27B-ica-lens
    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 lmxxf/Qwen3.8-27B-ica-lens --local-dir ~/models/Qwen3.8-27B-ica-lens
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Qwen3.8-27B-ica-lens
  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

  1. Qwen3.8-27Bfamily root
  2. Qwen3.8-27B-FP8
  3. Qwen3.8-27B-ica-lens

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

lmxxf/Qwen3.8-27B-ica-lens
Author
lmxxf
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