---

name: magi-1 description: "Operate MAGI-1 autoregressive video generation inference, ComfyUI nodes, configs, and prompt enhancement assets." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

MAGI-1 repo skill

Use this skill when a task is about MAGI-1, the SandAI autoregressive video generation repository, including source-code inference, model/config selection, CUDA/runtime setup, ComfyUI custom nodes, or the bundled Dify prompt-enhancement workflow.

MAGI-1 is not a lightweight CPU library. Real generation needs CUDA, the MAGI DiT weights, T5 weights, VAE weights, special-token assets, and ffmpeg. This skill provides self-contained operating guidance and safe preflight helpers; it does not include model weights.

Route by task

User task	Read
Build or debug t2v, i2v, or v2v source-code commands	sub-skills/inference/SKILL.md
Edit or validate MAGI JSON configs, checkpoint paths, process counts, or CFG/distill/fp8 settings	sub-skills/inference/SKILL.md and sub-skills/inference/references/configuration.md
Use the Python MagiPipeline API	sub-skills/inference/references/cli-and-api.md
Install, recognize, or operate the MAGI-1 ComfyUI custom node	sub-skills/comfyui/SKILL.md
Import or adapt MAGI ComfyUI workflow JSONs	sub-skills/comfyui/references/comfyui-nodes-and-workflows.md
Prepare weights, runtime dependencies, or safe environment checks	references/installation-and-assets.md and scripts/magi_runtime_preflight.py
Understand model families and hardware expectations	references/model-and-config-overview.md
Use the prompt enhancement DSL outside MAGI generation	references/dify-prompt-enhancement.md
Diagnose install/import, asset, CUDA, ffmpeg, or cross-workflow issues	references/troubleshooting.md plus the nearest sub-skill troubleshooting reference

Common operating path

1. Choose the user-facing surface:
   • Source CLI/API inference: use the inference sub-skill.
   • ComfyUI node graph: use the comfyui sub-skill.
   • Prompt rewriting in Dify: use the Dify reference.
2. Choose a model family from references/model-and-config-overview.md. Prefer 4.5B variants for single-GPU use; 24B examples are multi-GPU configurations.
3. Download or locate local MAGI DiT, T5, and VAE weights. Edit a copied JSON config so load, t5_pretrained, and vae_pretrained point to local paths.
4. Run scripts/magi_runtime_preflight.py in the runtime Python to check CUDA, attention packages, ffmpeg, and optional source assets.
5. For source CLI/API, run sub-skills/inference/scripts/magi_config_check.py before generation and sub-skills/inference/scripts/magi_command_builder.py to print a command.
6. For ComfyUI, import one of the bundled workflow JSONs from sub-skills/comfyui/references/workflows/ and reassign every placeholder path inside the UI before queueing.
7. If anything fails, read references/troubleshooting.md first, then the workflow-specific troubleshooting page.

Strong constraints and safety notes

• Do not present a config parse, helper output, or import check as a successful video generation. A real MAGI smoke test must load weights and write a playable MP4.
• Do not tell users to run original repository examples from an absent checkout. Use this skill's bundled helpers and references to reconstruct safe commands and checks.
• Do not claim CPU-only generation support. CUDA is required by MAGI source distributed initialization and model code paths.
• Match engine_config.pp_size * engine_config.cp_size to the launched process count. The bundled inference config checker validates this preflight.
• Treat model architecture fields as checkpoint-coupled. For most tasks, copy a release-family config and edit paths, seed, output dimensions, frame count, steps, FPS, and parallel sizes only when launch topology changes.
• The ComfyUI node path is single-process/single-GPU by default because MagiProcess sets its own distributed and CUDA visibility environment variables.
• Large model downloads, full generation, and ComfyUI host operation may be expensive or environment-specific. Ask for user approval before long downloads or long generation runs.

Bundled helpers

• scripts/magi_runtime_preflight.py: checks Python version, key dependency imports, CUDA visibility, optional tiny CUDA tensor execution, ffmpeg availability, and optional MAGI source asset paths without loading model weights.
• sub-skills/inference/scripts/magi_config_check.py: validates MAGI JSON config schema, CFG/distill/fp8 rules, process-count alignment, and optional checkpoint paths.
• sub-skills/inference/scripts/magi_command_builder.py: prints a source-code inference command and recommended environment variables; it does not execute generation.
• sub-skills/comfyui/scripts/inspect_workflow_nodes.py: inspects ComfyUI workflow JSONs offline and lists MAGI nodes plus placeholder fields.

Provenance and routing metadata

• Source snapshot and evidence paths are recorded in references/repo-provenance.md.
• Managed router metadata for repo-skill import is recorded in references/repo-routing-metadata.json.