veomni-develop
SkillAI & modelsThis skill gives your AI a safety checklist for adding features and refactoring code in the VeOmni training framework. Once added, it can analyze the impact of a planned change across modalities, the trainer hierarchy, the data pipeline, and distributed code before anything is implemented. Use it before any non-trivial change, and switch to veomni-new-model or veomni-new-op when the work is model-specific or ops-specific.
Available today. Use it from your connected AI after setup.
No other account needed.
Add the skill, then ask your AI to run the checklist before you start any feature or refactor in the VeOmni codebase. For model-specific or ops-specific work, point it to veomni-new-model or veomni-new-op instead.
Then ask your AI: use the veomni-develop skill
What your AI can do with it
- Run a safety checklist before non-trivial changes to VeOmni
- Analyze the impact of changes across modalities
- Check how refactors affect the trainer hierarchy
- Assess the impact of changes on the data pipeline
- Review changes to distributed code
- Tell you when veomni-new-model or veomni-new-op is the better fit
What this skill tells your AI
The instructions your AI receives, as published by bytedance-seed/veomni in .agents/skills/veomni-develop/SKILL.md and read by ahel’s review.
Impact Analysis
Before implementing, check which areas your change affects:
| Area | What to check | Why it matters |
|---|---|---|
veomni/trainer/ | All trainer subclasses (TextTrainer, VLMTrainer, DitTrainer, RL trainers) | Changing BaseTrainer method signatures breaks all subclasses |
veomni/data/data_collator.py | All modalities (text, VLM, DiT) | Collators are tightly coupled to model-specific preprocessing |
veomni/distributed/ | FSDP2 + ExtraParallel/MoE/SP paths | Shared distributed code is used by many downstream modalities |
veomni/models/auto.py, loader.py | Model registry, import-time side effects | MODELING_REGISTRY is populated at import time; moving registrations breaks loading |
configs/ | YAML config keys | Renaming config keys breaks existing training configs silently |
veomni/models/transformers/*/ | __init__.py registration entry points | All models ship a patchgen-generated v5 path under generated/; never import or call legacy modeling_<m>.py or apply_veomni_<m>_patch() (these no longer exist) |
Refactoring Safety Rules
When restructuring code (same behavior, better structure):
- Baseline first: run
pytest tests/before any change, record results. - One change per commit: ONE structural change → update ALL callers → verify tests match baseline → commit.
- Never batch multiple refactoring steps into one commit.
- Check baseline again at the end — results must be identical.
Common Traps
veomni.models.autoregistration depends on import-time side effects — moving registrations into functions or delaying them breaks model loading.- Renaming config keys silently breaks existing YAML configs in
configs/— grep all YAML files first. veomni.distributedmodules feed into ExtraParallel/MoE/SP — touching shared code may affect every modality, so run the cross-cutting parallel tests.- Data collators in
veomni/data/data_collator.pyare coupled toDEFAULT_DATA_COLLATE_INFO— adding new tensor keys requires updating the collate info table. MainCollatorhas strict SP ordering (pad → slice → FA kwargs → slice position_ids) — reordering breaks SP correctness.position_ids == 0marks segment boundaries for FA varlen — any transform that produces position_ids must preserve this convention.
Documentation
Before committing, check if the change requires documentation updates:
- New/changed API → update or create docs in
docs/. - New/changed config fields → update config examples in
configs/and relevant docs. - Architecture change → update
.agents/knowledge/architecture.md. - New constraint discovered → add to
.agents/knowledge/constraints.md.
When to Use Other Skills
- New model →
/veomni-new-model - New op/kernel →
/veomni-new-op - Bug fix or debugging →
/veomni-debug - Dependency update →
/veomni-uv-update
Signals
- GitHub stars
- 2k
- Forks
- 272
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
veomni-develop- Source
- github.com/bytedance-seed/veomni