---

name: training-workflows description: "Route Pyramid-Flow Causal VAE training, DiT fine-tuning, distributed launchers, and launch-time prerequisites." disable-model-invocation: true metadata: disco-role: operating license: MIT

Pyramid-Flow training workflows

Use this sub-skill when the task is about Pyramid-Flow training launchers or preflight checks, not inference demos or low-level model internals.

Route here for

• Autoregressive temporal-pyramid DiT video fine-tuning from train/train_pyramid_flow.py and scripts/train_pyramid_flow.sh.
• Non-AR/full-sequence DiT fine-tuning from train/train_pyramid_flow.py and scripts/train_pyramid_flow_without_ar.sh.
• Causal Video VAE stage-1 and stage-2 training from train/train_video_vae.py and scripts/train_causal_video_vae.sh.
• FSDP, sequence-parallel, sync-input, DDP, and VAE context-parallel startup issues.
• Training prerequisites: CUDA GPU count, LPIPS checkpoint, checkpoint/model paths, annotation inputs, NUM_FRAMES, VIDEO_SYNC_GROUP, CONTEXT_SIZE, and per-device batch-size invariants.

Route elsewhere

• Inference demos, Gradio apps, prompts, generated videos, and generation launchers: ../generation-inference/SKILL.md.
• Annotation authoring, JSONL schemas, text-feature extraction, and VAE-latent precompute commands: ../data-preparation/SKILL.md.
• Model API signatures, scheduler math, VAE encode/decode internals, and reusable tensor-shape details: ../core-components/SKILL.md.

Read first

1. CLI reference for verified training flags, defaults, helper signatures, and invariants.
2. Workflows for AR DiT, non-AR DiT, and two-stage Causal VAE launch planning.
3. Troubleshooting for GPU, LPIPS, FSDP, context-parallel, gradient-checkpointing, resolution, and annotation failures.

Bundled helpers

• scripts/build_training_commands.py prints validated torchrun command shapes adapted from the three training shell launchers. It never starts a training run.
• scripts/check_training_prereqs.py validates command-builder invariants, CUDA/distributed visibility, optional source syntax, and optional annotation samples before launch. It never launches torchrun.

For any expensive run, build or check the command first, fix all preflight failures, then copy the printed command into the actual training environment.