veomni-migrate-transformers-v5

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.

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

81%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

An exceptionally actionable, well-sequenced protocol with strong validation feedback loops, but it is over-long: the Common Pitfalls section repeats points already made inline and could be consolidated or moved to a reference file to improve token efficiency.

Suggestions

Deduplicate the Common Pitfalls section: many entries (MoE layout dispatch, name_map dropping MoE-only fields, dummy_forward dtype lookup, copy-pasting sibling converter docstrings) restate guidance already given in the phase bodies — keep one canonical location and cross-reference.

Move the lengthy pitfalls catalog and the multimodal-metadata checklist into a bundled reference file under ./references/ and link to it, reducing SKILL.md toward an overview while preserving the detail on demand.

Trim trigger-phrase redundancy in the description and condense repeated patch-primitive caveats (e.g. the name_map dataclass caveat appears in both the table and the pitfalls) to tighten token budget.

Dimension	Reasoning	Score
Conciseness	The body is information-dense and avoids beginner-concept padding, but at ~1030 lines it is heavily padded — the 'Common Pitfalls' section re-explains issues already covered inline (e.g. MoE layout dispatch, name_map dataclass field dropping, dummy_forward dtype) with noticeable repetition.	3 / 5
Actionability	Provides concrete, mostly copy-paste-ready guidance: a PatchConfig skeleton, a patch-primitives table mapping effect to decorator, regen/pytest/ruff commands with exact flags, and shape-dispatch snippets — all executable.	5 / 5
Workflow Clarity	A clear 9-phase sequence with explicit validation checkpoints and feedback loops (make quality, patchgen --check / --fix, pytest -k acceptance criteria, mandatory /veomni-review pre-commit gate), matching the top anchor.	5 / 5
Progressive Disclosure	References are clearly signaled at the top ('References (read first, load on demand)' with bulleted doc paths) and phases are well-organized, but no bundle files exist and a large amount of detail (the pitfalls catalog, metadata checklist) is inlined rather than split into separate files.	4 / 5
	Total	17 / 20 Passed

Description

92%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A highly specific, third-person description that clearly states both capabilities and trigger conditions with concrete example phrases. Its only weakness is mild verbosity and jargon-dense trigger terms that may miss less-expert phrasings.

Dimension	Reasoning	Score
Specificity	Enumerates multiple concrete actions — 'creating GPU and NPU patchgen configs', 'using patchgen decorators (replace_class/override_method/replace_function/modify_init/add_post_import_block/drop_import_names)', 'handling MoE weight-loading (CheckpointTensorConverter + fused gate_up_proj layout)' — giving comprehensive coverage rather than vague abstraction.	5 / 5
Completeness	Explicitly answers both 'what' (a long 'Covers:' clause listing capabilities) and 'when' (a dedicated 'Trigger:' clause with concrete phrases), matching the top anchor.	5 / 5
Trigger Term Quality	Includes several natural trigger phrases ("'port a model to patchgen'", "'transformers v5 migration'", "'add NPU patchgen'") but they are somewhat jargon-heavy and lack common synonyms or variations a less expert user might say.	4 / 5
Distinctiveness Conflict Risk	Targets a very specific niche — VeOmni transformers v5 patchgen-generated modeling — with model-specific triggers, making overlap with unrelated skills minimal.	5 / 5
	Total	19 / 20 Passed

Validation

87%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
skill_md_line_count	SKILL.md is long (1030 lines); consider splitting into references/ and linking	Warning
referenced_paths_exist	Referenced path issues: 1 missing, 1 deeper-than-1-level	Warning

	Total	14 / 16 Passed

Repository: ByteDance-Seed/VeOmni
Path: .agents/skills/veomni-migrate-transformers-v5/SKILL.md
Commit: c465a63

Reviewed: about 19 hours ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.