VeOmni

Create a pull request for the current branch. Handles uncommitted changes, generates a PR title matching the `[{modules}] {type}: {description}` format enforced by CI, and fills in the PR description template. Trigger: 'create pr', 'open pr', 'submit pr', 'make pr'.

Use this skill when adding or refreshing a patchgen-generated modeling file for a VeOmni model under its generated directory — GPU-only or GPU+NPU, dense or MoE, text-only / VLM / Omni-thinker+talker. Covers: creating GPU and NPU patchgen configs, using patchgen decorators (replace_class/override_method/replace_function/modify_init/add_post_import_block/drop_import_names), reusing sibling-model patches via name_map, handling MoE weight-loading (CheckpointTensorConverter + fused gate_up_proj layout), multimodal/VLM forward with Ulysses SP, excluding speech/vocoder subtrees in Omni models (talker/token2wav/DiT/BigVGAN), wiring __init__.py for the patchgen-generated classes, running codegen, and adding test cases. Trigger: 'port a model to patchgen', 'add patchgen for a model', 'transformers v5 migration', 'add NPU patchgen'. Do NOT edit files under generated/ manually — always regenerate via patchgen.

Use this skill when adding support for a new model to VeOmni. Covers the full lifecycle: analyzing the HuggingFace model, creating model patches, defining parallel plans, writing configs, integrating with the trainer, and testing. Trigger: 'add model', 'support new model', 'integrate a model', 'new model support'.

Use this skill when adding a new optimized kernel or operator to veomni/ops/. Covers the full lifecycle: understanding VeOmni's ops architecture (KERNEL_REGISTRY + OpSlot dispatch, with a thin function-pointer shim for a few legacy global ops), implementing the kernel, registering it, adding tests, and documenting it. Trigger: 'add op', 'new kernel', 'add attention variant', 'new fused op', 'add triton kernel', 'optimize operator'.

Use this skill for performance profiling and optimization. Two modes: (1) Analyze existing profile files (Chrome traces, memory snapshots) — write scripts to parse and summarize metrics per user requirements. (2) Generate profiles during development — configure ProfileConfig, run training, collect traces, analyze bottlenecks, and suggest optimizations. Trigger: 'profile', 'performance', 'slow', 'MFU', 'throughput', 'bottleneck', 'memory usage', 'trace', 'optimize training speed'.

Use this skill when updating dependencies managed by uv: bumping a package version, upgrading the uv tool itself, updating torch/CUDA stack, switching transformers version, or regenerating the lockfile. Trigger: 'update dependency', 'bump version', 'upgrade uv', 'update torch', 'update lockfile', 'uv sync fails'.

Use this skill for ANY bug, error, crash, wrong output, loss divergence, gradient explosion, test failure, CUDA error, distributed training hang, checkpoint load failure, or unexpected behavior. Covers both quick fixes (clear root cause) and complex debugging (unclear cause). Trigger: 'fix bug', 'fix error', 'broken', 'crash', 'doesn't work', 'fails with', 'loss NaN', 'training hangs', 'FSDP error', 'OOM'.

VeOmni-specific checklist for feature development and refactoring. Covers impact analysis across modalities, trainer hierarchy, data pipeline, and distributed code. Use before implementing any non-trivial change. For model-specific or ops-specific work, use veomni-new-model or veomni-new-op instead. Trigger: 'add feature', 'implement', 'refactor', 'reorganize', 'new capability'.

Use this skill before committing ANY code change — this is a mandatory gate in the commit flow. Also trigger proactively when: you've made changes across multiple files and want to check consistency, you're unsure if a fix is safe, a change touches shared infrastructure (BaseTrainer, distributed, model loading, data pipeline), or a change is larger than a few lines. The review launches a subagent that checks implementation quality, multi-file consistency, and known constraint violations, then rates the change as safe/needs-attention/risky.