Skip to content
Robotics

Kinematic Residual VLA Checkpoints

чекпоинты для VLA-роботов

2 min readupdated

Набор исследовательских чекпоинтов от adi2440 для оценки и продолжения fine-tuning моделей DynamicVLA/DOM, DOM с velocity conditioning, KRV100, SmolVLA, VLASH и π0.5. Включает baseline- и KR-варианты, а также разные истории обучения. Полный снапшот — около 313 ГБ до дедупликации, предусмотрена выборочная загрузка.

Ключевые особенности

  • Чекпоинты для KRV100, DOMV, DOM, SmolVLA, VLASH и π0.5 — с baseline для сопоставимых сравнений.
  • Несколько training scope: full, large-audited-197198, 20k-fresh, 20k-continuation, mixed-full-then-20k и full-historical.
  • Native- и adapter-чекпоинты: для DOM/KRV нужны Transformers 5.2.0 и LeRobot 0.3.3, для адаптеров — Transformers 4.53.3 и LeRobot 0.4.1.
  • Загрузка через tools/checkpoint_tools.py по алиасам; можно скачать только нужный чекпоинт, а не весь набор.
  • Production-снапшоты содержат optimizer, scheduler, sample cursor и RNG state для продолжения fine-tuning.
  • Это кастомные research-экспорты: передавать adapter checkpoint.pt напрямую в stock LeRobot from_pretrained() не поддерживается.

Пример

from tools.checkpoint_tools import load_model

policy, metadata = load_model(".", "krv100-full-best", device="cuda")
ПодробнееПолный снапшот — около 313 ГБ до дедупликации; загрузка выборочная.
KRV100 full: 203 151 тренировочный эпизод; DOMV large-audited-197198: 197 198 эпизодов (97.066% опубликованных).
π0.5 + KR: 6 полных эпох и часть 7-й; последний VLASH baseline тоже неполный.
Closed-loop оценка требует Isaac Sim 4.5.0 и Isaac Lab 2.2.1. Лицензии: S-Lab non-commercial для DynamicVLA, Gemma terms для π0.5/PaliGemma, Apache для SmolVLA/LeRobot.
Discuss Telegram 0

Will it run?

Runs in FP8 / INT8

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 38 GB · ready files42 GB
  • FP8 / INT8recommendedweights 19 GB21 GB
  • 4-bitweights 11 GB12 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 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.

    # adi2440/kinematic-residual-vla-checkpoints
    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 adi2440/kinematic-residual-vla-checkpoints --local-dir ~/models/kinematic-residual-vla-checkpoints
    # The model is in ~/models. How to run it is in the model card (link above).
  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. SmolLM2-360Mfamily root
  2. dynamic-vla-DOM
  3. kinematic-residual-vla-checkpoints

New posts about versions of SmolLM2-360M (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
LoRAs and adapters 8All on Hugging Face
Merges 1All on Hugging Face

About the model

adi2440/kinematic-residual-vla-checkpoints
Author
adi2440
Category
Robotics
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

Other

BIGR

BIGR