Skip to content
Other

Cross-Layer Transcoder

рентген Qwen3-8B-Base

1 min readupdated

ByteDance-Seed опубликовала Cross-Layer Transcoder (CLT) для Qwen3-8B-Base. Это интерпретационный инструмент: 36 слоёв × 32 768 латентов, 94,2 млрд параметров в fp32, общий размер 377 ГБ. Латент каждого слоя читает residual stream перед MLP и записывает в выходы MLP этого и всех следующих слоёв.

Ключевые особенности

  • При инференсе совместно выбираются 36·k = 360 латентов (k = 10) по score a_l[i] · W_dec_norm[l, i].
  • Полный decode требует файлы W_dec_*.safetensors на 357 ГБ; всего checkpoint занимает 377 ГБ.
  • Демо запускается через torchrun на 8 GPU; для encode-only хватает одной 80 ГБ GPU, если не скачивать декодеры.
  • Нормализация уже вшита в веса: CLT принимает raw resid_mid и восстанавливает raw mlp_out.
  • Обучен на Dolma 3, Dolci-Think и Dolci-Instruct (5:2:1, ~3,4 млрд токенов); веса и код — Apache 2.0.

Пример

huggingface-cli download ByteDance-Seed/Cross-Layer-TopK-CLT --local-dir clt
torchrun --nproc_per_node 8 hf_encode_demo.py
ПодробнееБенчмарков в карточке нет: это не LLM, а CLT для анализа Qwen3-8B-Base.
Полный decode требует fp32 и 377 ГБ весов; encode-only — одна 80 ГБ GPU без файлов W_dec_*.safetensors.
CLT обучен на fp32-активациях, базовую модель тоже надо запускать в fp32.
Токены с нормой resid_mid или mlp_out ≥20× среднего по слою отбрасываются, как при обучении.
Discuss Telegram 5

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 18 GB · ready files20 GB
  • FP8 / INT8recommendedweights 9.0 GB10 GB
  • 4-bitweights 5.1 GB6.1 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 2080 Ti · BF16 — from RTX 4090

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.

    # ByteDance-Seed/Cross-Layer-TopK-CLT
    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 ByteDance-Seed/Cross-Layer-TopK-CLT --local-dir ~/models/Cross-Layer-TopK-CLT
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to Cross-Layer-TopK-CLT
  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 Qwen3-8B-Base (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 4All on Hugging Face

About the model

ByteDance-Seed/Cross-Layer-TopK-CLT
Author
ByteDance-Seed
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

05

Ask AI about this model

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

Comments

Language models

Pluto-2.1

Pluto-2.1
Computer vision

PANORAMA-2B

PANORAMA-2B