Skip to content
Other

NegWAM

безопасное планирование для end-to-end автопилота

2 min readupdated

NegWAM — это safety post-training framework для unified world-action моделей в end-to-end автономном вождении. Он учится на near-expert failure trajectories — небезопасных альтернативах, которые остаются геометрически близкими к экспертным траекториям. В репозитории — веса NegWAM Transformer (6.24 млрд параметров), конфиг и результаты оценки NAVSIM v2.

Ключевые особенности

  • Repulsive anchors уводят предсказания от опасных альтернатив, а scene-conditioned rider tokens учат совместимость сцены и траектории.
  • Вспомогательная rider-ветка нужна только при обучении и исключена из inference path.
  • На входе — камеры, ego state/history и навигационные команды; на выходе — 8 ego poses (x, y, heading) на 4 секунды с шагом 0.5 с через NAVSIM agent.
  • Кастомная Diffusers-реализация WanTransformer3DNavSimModel; веса в Safetensors BF16, файл ~12.49 ГБ, лицензия Apache 2.0.
  • Для запуска нужны кастомная реализация и eval-инструменты из GitHub-репозитория; config.json и веса должны лежать в одной директории.

Пример

Сценарий из карточки: модель получает наблюдения с камер, историю ego-состояния и навигационную команду, а затем планирует 8 ego poses (x, y, heading) на 4 секунды вперёд с шагом 0.5 с через NAVSIM agent. Это планировщик траектории, а не генератор видео или чат-модель.

ПодробнееБенчмарки NAVSIM v2: EPDMS 40.4 на navhard (two-stage) и 93.3 на navtest; шкала 0-100.
Веса: ~12.49 ГБ в BF16; для оценки нужна кастомная реализация и инструменты из GitHub.
Для navhard в отчёте используется combined summary score, а не среднее по всем строкам CSV.
Лицензия Apache 2.0; auxiliary rider branch не входит в inference path.
Discuss Telegram 0 18

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 12 GB · ready files13 GB
  • FP8 / INT8recommendedweights 5.8 GB6.9 GB
  • 4-bitweights 3.3 GB4.1 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 5060 Ti 8GB · BF16 — from RTX 5080

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.

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

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

About the model

Zizizi-hao/NegWAM
Author
Zizizi-hao
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

180

Ask AI about this model

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

Comments