Skip to content
Other

SkinTokens GGUF BF16×3

точность почти F32 на Vulkan без NaN

1 min readupdated

BF16-weight GGUF-чекпоинты от chibifire для проекта V-Sekai-fire/skin-tokens-ggml. Получены из опубликованных F32 GGUF-весов LocalAI-io/SkinTokens-GGUF и предназначены для ригинга 3D-персонажей через skin-tokens-ggml.

Ключевые особенности

  • Решает проблему NaN у F16-весов на Vulkan coopmat2: upstream ggml-vulkan не имеет пайплайна F16-weight/F32-activation matmul и молча делает lossy F32→F16.
  • Каждый матричный тензор разбит на два BF16-тензора — name (hi) и name.lo (lo) — по схеме XLA ALG_DOT_BF16_BF16_F32_X3.
  • hi + lo восстанавливает исходное F32-значение с относительной ошибкой около 2e-4, а суммирование cross-термов даёт почти F32-точность на нативном bf16×bf16 Vulkan.
  • Norm-, embedding- и sensitive-тензоры оставлены в F32 — как в «sensitive»-наборе проектного convert_to_gguf.py.
  • Нужна совместимая версия skin-tokens-ggml: изменение с mul_mat_x3 и обнаружением .lo-тензоров вошло в main через PR #33 (commit 0fccf15).

Пример

hf download chibifire/skintokens-ggml-bf16x3 --local-dir models/SkinTokens-GGUF/BF16X3
skintokens-cli rig models/SkinTokens-GGUF/BF16X3 character.glb character-rigged.glb --device vulkan
ПодробнееНа reference humanoid mesh: 24 предсказанных joint — точное совпадение с F32, 19 различных костей, распределение весов в пределах 0.01% от F32.
Плейн BF16 без hi/lo даёт 90 joint вместо 24, поэтому нужен именно BF16×3.
Требования к VRAM в карточке не указаны. Лицензия upstream и производных — MIT.
Discuss Telegram 0 37

Will it run?

Runs in GGUF

The most precise of the ready quantised files fits.

  • GGUFweights 2.2 GB · ready files2.9 GB

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: GGUF — 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.

    # chibifire/skintokens-ggml-bf16x3
    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 chibifire/skintokens-ggml-bf16x3 --local-dir ~/models/skintokens-ggml-bf16x3
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to skintokens-ggml-bf16x3
  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 skintokens-ggml-bf16x3 (quants, finetunes, LoRAs) will show up under the bell.

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

About the model

chibifire/skintokens-ggml-bf16x3
Author
chibifire
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

370

Ask AI about this model

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

Comments