cosmos3-inference

Guide users through running Cosmos3 inference — offline batch generation, online serving with Ray and Gradio, parallelism options, input formats, sampling parameters, and prompt upsampling. Use when the user asks "how do I run inference", "how do I generate a video", "how do I serve the model", "what parameters should I use", or any question about running the model to produce outputs.

Zur Installation springen Dateien prüfen

Repository: NVIDIA/cosmos-framework
Letzte Quellaktivität: 1. Juni 2026 um 03:50
Erkannte Sprache von SKILL.md: Englisch
Sterne: 588
Forks: 188

Installationsoptionen

Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.

Mit einem KI-Assistenten installieren

Diesen Prompt kopieren und in den verwendeten KI-Assistenten einfügen.

Follow https://skillsmp.com/skill-install/prompt.md to install only the Agent Skill at https://github.com/NVIDIA/cosmos-framework/tree/main/.agents/skills/cosmos3-inference.

Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.

npx skills add https://github.com/NVIDIA/cosmos-framework --skill cosmos3-inference

Der Befehl bleibt in einer Zeile. Scrollen Sie horizontal, um ihn vor dem Kopieren vollständig zu prüfen.

Sie bevorzugen eine lokale Kopie? Laden Sie die Dateien herunter, die SkillsMP derzeit vorliegen.

Quelldateien prüfen

Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.

SKILL.md

Quellanweisungen · Schreibgeschützte Vorschau

name: cosmos3-inference
description: Guide users through running Cosmos3 inference — offline batch generation, online serving with Ray and Gradio, parallelism options, input formats, sampling parameters, and prompt upsampling. Use when the user asks "how do I run inference", "how do I generate a video", "how do I serve the model", "what parameters should I use", or any question about running the model to produce outputs.

# Cosmos3 Inference ## When to use this skill - Use when a user wants to generate images or videos with Cosmos3 - Use when a user asks about inference parameters, input formats, or parallelism - Use when a user wants to set up online serving (Ray Serve, Gradio) - Use when a user asks about prompt engineering or upsampling - For environment or import errors, hand off to **cosmos3-env-troubleshoot** ## Path convention All paths below are relative to the cosmos3 package root (`../../../` from this skill file). All `uv run` / `python` commands should also be run from there. ## Where to find answers | User question | Go to | | ----------------------------------------------------------------------- | --------------------------------------------------------------------------------------- | | How do I run inference? (single-GPU, multi-GPU) | `README.md` § Inference | | Which model should I use? (Nano vs Super, memory, shift) | `README.md` § Models | | Which modality? (t2i, t2v, i2v, examples) | `README.md` § Modalities | | What parallelism preset? (latency vs throughput) | `README.md` § Inference | | What input fields are available? (prompt, vision_path, num_frames, ...) | `docs/inference.md` § Sample Arguments | | What are the default parameter values? | `cosmos_framework/inference/defaults/<model_mode>/sample_args.json` (per-modality JSON) | | How do I use custom defaults? | `docs/inference.md` § Custom Defaults | | How do I override a parameter? (precedence) | `docs/faq.md` § How do I override a default parameter? | | What is the shift parameter? | `docs/faq.md` § What is the `shift` parameter? | | How many frames can I generate? (resolution caps) | `docs/faq.md` § How many frames can I generate? | | How do I start Ray Serve / Gradio / submit requests? | `docs/faq.md` § How do I run online inference with Ray? | | How do I upsample short prompts? | `docs/faq.md` § Prompt upsampling | | How do I use the low-level API? (examples/) | `examples/inference.py` (model API) / `examples/inference_pipeline.py` (pipeline API) | | All CLI flags | `uv run --all-extras --group=cu130 python -m cosmos_framework.scripts.inference --help` | ## Things not obvious from the docs - **Path resolution**: relative paths in input JSON files are resolved relative to the **JSON file's directory**, not the working directory. - **Seed**: always pass `--seed` for reproducible results. Without it, a random seed is used each time. - **Resume**: interrupted runs can be resumed by re-running the same command — existing outputs are skipped automatically. - **`--keep-going`**: continues processing remaining samples after a per-sample failure (e.g. guardrail rejection). Used in online serving by default. - **Unique names**: every sample in a run must have a unique `name` field, or the script will error. ## Related skills | Skill | When to use | | -------------------------------------- | ---------------------------------------------- | | `../cosmos3-setup/SKILL.md` | Installation and environment setup | | `../cosmos3-codebase-nav/SKILL.md` | Finding files, parameters, and configs in code | | `../cosmos3-env-troubleshoot/SKILL.md` | Debugging environment and runtime errors |

Auf GitHub ansehen

Zurück zur Suche Zurück zur Installation

cosmos3-qwen3-vl-robovqa-sft

Fine-tune Qwen3-VL-8B-Instruct, a vision-language model from outside the Cosmos family, with the Cosmos Framework posttrain SFT path on robot-video question answering (RoboVQA), on one GPU, and score it before and after with the same evaluator. Use when the user asks to "fine-tune Qwen3-VL with Cosmos Framework", "train a non-Cosmos model or a VLM in the framework", "run SFT on robot videos / RoboVQA", "reproduce the Qwen3-VL RoboVQA result", or wants a worked post-training example with a measured before and after (environments, data preparation, preflight, training, Hugging Face export, evaluation). Not for Cosmos3 checkpoints (use cosmos3-post-training) and not for LoRA (the SFT path trains every parameter).

2026-10-31

cosmos3-generate-media

Generate images, videos, and audio-synchronized videos with the Cosmos3 audiovisual generator models (Edge / Nano / Super / Super-4Step) through any of the seven supported serving paths: diffusers in-process, cosmos_framework torchrun, vLLM-Omni, SGLang, TensorRT-LLM, distilled TensorRT-LLM, or NIM. Use when the user asks "generate a video from a prompt", "run text2image / text2video / image2video with Cosmos3", "serve a Cosmos3 generator", or asks which Cosmos3 generation backend fits their GPU. Self-contained: sample prompts, reference images and negative prompts ship in this skill's assets/.

2026-10-30

cosmos3-skill-authoring

Author new Cosmos3 agent skills and their eval configs, and get them through the skill eval. Use when adding a skill under .agents/skills/, migrating a cookbook notebook or README into a skill, writing or updating an eval.config, choosing eval cases and assertions, fixing a skill after a failed or self-corrected eval cell, reading an eval report on the dashboard, or preparing a skill PR. Covers the required structure, the writing rules that decide first-attempt pass/fail, the eval.config schema (multi-case evals), how the evaluator judges, and how to run the eval in GitHub CI or locally. Do not use it to run the Cosmos3 workloads themselves (the per-workload skills do that) or to change the evaluator.

2026-10-30

Alle aus diesem Repository ansehen

cosmos3-inference

Quellinformationen

Installationsoptionen

Mit einem KI-Assistenten installieren

Quelldateien prüfen

Mehr aus diesem Repository

Mit einem KI-Assistenten installieren

Mehr aus diesem Repository