---

name: pyramid-flow description: "Use Pyramid-Flow for video generation, data precomputation, Causal Video VAE, distributed training, and reusable model-component inspection." disable-model-invocation: true metadata: disco-role: operating license: MIT

Pyramid-Flow

Use this skill when a task names Pyramid-Flow, pyramid-flow, PyramidDiTForVideoGeneration, CausalVideoVAE, the generation demos, the annotation/precompute helpers, or the training launchers in this repository.

Route by task

• Model APIs, schedulers, VAE encode/decode, and distributed helpers: read sub-skills/core-components/SKILL.md.
• Prompt/image generation, Gradio demos, multi-GPU inference, and checkpoint selection: read sub-skills/generation-inference/SKILL.md.
• JSONL annotations, dataset loading, text-feature extraction, and VAE-latent precompute: read sub-skills/data-preparation/SKILL.md.
• AR/non-AR DiT fine-tuning, Causal VAE training, and launch-time prerequisites: read sub-skills/training-workflows/SKILL.md.

Start here

• references/model-overview.md for the model families and workflow map.
• references/installation.md for import-root and dependency setup.
• references/troubleshooting.md for cross-cutting install, backend, checkpoint, and routing failures.
• references/repo-provenance.md for the source snapshot used to build this skill.
• references/repo-routing-metadata.json for managed router placement metadata.
• scripts/check_environment.py for a safe environment and bundled-helper smoke check.

Quick smoke

From a Pyramid-Flow checkout or a shell that can point to one, run:

python scripts/check_environment.py --repo PATH_TO_PYRAMID_FLOW

Use --json for machine-readable output. The helper checks the common runtime imports, CUDA visibility when torch is available, and the bundled helper script entry points without starting generation or training.

Installation and runtime baseline

• The repository snapshot has no pyproject.toml or setup.py; make the checkout root importable with PYTHONPATH or by running from the checkout root.
• Install the repo's runtime dependencies before trying any workflow that imports torch, diffusers, transformers, accelerate, sentencepiece, jsonlines, cv2, or the training/data helpers.
• Keep checkpoint downloads, model-cache creation, feature extraction, and training launches explicit. The bundled helpers only plan or preflight those workflows unless you ask them to execute.
• Treat CUDA as required for truthful generation and training. CPU is useful for bounded import, syntax, fixture, and scheduler checks only.

Safety defaults

• Do not assume the repo checkout is installed as a package; top-level imports come from the repository root.
• Do not launch generation, extraction, or training until the matching sub-skill preflight passes and the needed checkpoint or dataset artifacts are present.
• Do not route a task to a narrower sub-skill if the user is actually asking about the overall repo layout, repo routing metadata, or a cross-cutting environment issue.
• Do not rely on the original checkout remaining available after skill generation; the sub-skills and bundled references are the persistent operating knowledge.