Skip to content
Back to skills

Model Inference

ASecurity

"Run or plan Motus CUDA inference for real-world images and

  • 247 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 8, 2026
devopsapi

Works with

  • api

Security analysis

A100/100

Pro scans all 4 files and shows the line behind each finding

Scanned September 8, 2026

npx -y skills add VectorSpaceLab/AREX-Skill --skill model-inference --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Model Inference?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Model Inference
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vectorspacelab-model-inference-0520a6d3/badge)](https://www.skillsdirectory.com/skills/vectorspacelab-model-inference-0520a6d3)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: model-inference
description: "Run or plan Motus CUDA inference for real-world images and
  RoboTwin policy evaluation, including checkpoint loading, T5 conditioning,
  output contracts, and VRAM troubleshooting."
disable-model-invocation: true
metadata:
  disco-role: operating
license: Apache 2.0
---

# Motus model inference

Use this route for Motus image-to-video/action prediction, pre-encoded T5
conditioning, checkpoint compatibility, or RoboTwin policy deployment.

- Read [api-reference.md](references/api-reference.md) for model/checkpoint
  contracts and tensor shapes.
- Read [workflows.md](references/workflows.md) for real-world and RoboTwin
  command recipes and asset preflight.
- Read [troubleshooting.md](references/troubleshooting.md) for VRAM, CUDA,
  flash-attn, checkpoint, image, and external-runtime failures.
- Use [data-preparation](../data-preparation/SKILL.md) for dataset/image layout
  preparation and [training](../training/SKILL.md) for training configuration.

## Operating procedure

1. Select real-world no-environment inference or RoboTwin evaluation. Confirm
   the checkpoint family, embodiment YAML, WAN/VAE assets, VLM checkpoint, and
   input image format before loading a model.
2. For real-world inference, provide one three-view concatenated RGB image,
   state/config matching the embodiment, and a language instruction. Prefer a
   pre-encoded WAN T5 tensor when VRAM is limited.
3. Run the parser/help preflight, then verify CUDA, GPU memory, file paths, and
   action/image dimensions. Actual model construction and denoising are
   CUDA-only and require large downloaded weights; do not substitute a CPU
   import for an inference result.
4. Inspect both returned artifacts: predicted future frames are `[B,C,F,H,W]`
   in `[0,1]`, and actions are `[B, action_chunk_size, action_dim]`.
5. For RoboTwin, configure the external simulator/runtime and policy paths
   separately. The policy route may execute robot/simulation side effects and
   is not a safe generic smoke test.

Launches, downloads, and simulator runs require explicit user approval.

Files in this skill

  • SKILL.md2.1 KB
  • references/api-reference.md2.1 KB
  • references/troubleshooting.md2.2 KB
  • references/workflows.md1.7 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.