CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

cu-lengths-attention-flow

Bilingual guide for understanding how cu_lengths controls attention behavior across ViT and LLM stages, and how patch_positions scope differs between the two

Quality

74%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Passed

No findings from the security scan

Fix and improve this skill with Tessl

tessl review fix ./.opencode/skills/cu-lengths-attention-flow/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

67%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.

A dense, well-organized knowledge skill whose content is almost entirely non-obvious, repo-specific information with concrete code anchors — its substance is excellent. The dominant weaknesses are the systematic English/Chinese duplication that roughly doubles token cost and the absence of any explicit 'when to use' framing or debugging checkpoints in the body.

Suggestions

Eliminate the bilingual duplication (or move one language to a reference file): every paragraph appears twice, roughly doubling the skill's token footprint without adding information for the model reading it.

Move the detailed code excerpts (SFT attention-mask path, sequence-parallelism padding) into a references/ file, keeping SKILL.md as a leaner overview with one-level-deep pointers.

Add a short 'Debugging checklist' that sequences how to apply the knowledge (e.g. check cu_lengths shape → determine packed vs non-packed → verify cu_seqlens construction), giving the reasoning skill explicit validation checkpoints.

DimensionReasoningScore

Conciseness

Every fact is repo-specific and non-obvious (file roles, branch conditions, Flash Attention parameters) with no filler explaining concepts Claude already knows, but nearly every section is duplicated verbatim in English and Chinese (e.g. 'The entire sequence is ONE attention domain...' / '整个序列是一个 attention 域...'), roughly doubling token cost, and the core 'patch_positions does not carry into the LLM' insight is restated four times. Fits 'mostly efficient but could be tightened'; not 2 because each individual explanation is dense with no generic padding, not 4 because the systematic bilingual duplication exceeds minor trimming.

3 / 5

Actionability

Provides real code excerpts with exact file paths and precise branch conditions ('if cu_lengths.shape == torch.Size([1, 1])', 'PackedSeqParams(qkv_format="thd", cu_seqlens_q=cu_lengths[0], ...)') plus a quick-reference table mapping each field to where it is set and what it controls. Not 5 because several snippets are abridged rather than copy-paste ready (the SP-padding block references undefined 'pad_size'/'new_end_kv', the SFT snippet uses an unbound 'zero' tensor); not 3 because the guidance is concrete and executable in intent, not pseudocode.

4 / 5

Workflow Clarity

As a reasoning/knowledge skill it sequences the material as the actual data flow — cu_lengths generation (Sources A/B), forward-function branching, LLM attention behavior, SP padding, then pitfalls — introduced by an overview diagram. Not 5 because there are no explicit checkpoints guiding how to apply the knowledge when debugging (e.g. 'to diagnose unexpected cross-sample attention, check X first'), and the SFT path section appears after the main narrative without a transition tying it into the flow; not 3 because the sequence is complete and coherent with the diagram anchoring it.

4 / 5

Progressive Disclosure

No bundle files exist, so everything lives in a single ~270-line SKILL.md, but it is well-sectioned with an overview diagram, tables, a quick reference, and a pitfalls list that make navigation easy. Fits 'good structure; most content appropriately placed; minor organization gaps'; not 5 because detailed code blocks (SFT path, SP padding) are inlined where reference files would keep the overview lean, and not 3 because headers, tables, and the diagram provide genuinely clear navigation rather than buried content.

4 / 5

Total

15

/

20

Passed

Description

70%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, distinctive description that names exact mechanism symbols rather than generic capabilities. Its main weakness is the complete absence of 'when to use' trigger guidance, and it omits 'packed/packing' — one of the skill's central concepts — from its trigger vocabulary.

Suggestions

Add an explicit trigger clause, e.g. 'Use when debugging packed vs non-packed attention, cross-sample isolation, or attention boundaries in LLaVA-OneVision2 training.'

Include 'packed'/'packing', 'packed sequences', and 'cu_seqlens' as trigger terms since much of the skill concerns packed vs non-packed behavior.

Consider naming the repo/framework context (LLaVA-OneVision2 / Megatron) in the description to sharpen routing when multiple attention-related skills exist.

DimensionReasoningScore

Specificity

Names two concrete mechanisms with their exact symbols — 'how cu_lengths controls attention behavior across ViT and LLM stages' and 'how patch_positions scope differs' — which goes beyond naming the domain to specific actions. Not 5 because coverage is a single reasoning task rather than multiple comprehensive capabilities; not 3 because two distinct mechanisms are explicitly described, not just 1-2 minimal actions.

4 / 5

Completeness

The 'what' is clear ('Bilingual guide for understanding how cu_lengths controls attention behavior...') but there is no 'Use when...' clause or equivalent explicit trigger guidance, which caps completeness at 3 per the judging guidelines. Not 4 because the 'when' is entirely absent rather than just implicit; not 2 because the 'what' half is concrete and specific.

3 / 5

Trigger Term Quality

Includes the natural terms a user in this codebase would actually say — 'cu_lengths', 'patch_positions', 'attention', 'ViT', 'LLM'. Not 5 because 'packed'/'packing' (a central concept of the skill) and related variations like 'Flash Attention' or 'sequence packing' are missing; not 3 since the present keywords are strong and domain-natural rather than generic.

4 / 5

Distinctiveness Conflict Risk

'cu_lengths' and 'patch_positions' attention scoping across ViT and LLM stages is a narrow technical niche with distinct trigger symbols that virtually no other skill would claim. Matches the anchor for a clear niche with minimal conflict risk; no neighboring anchor fits better.

5 / 5

Total

16

/

20

Passed

Validation

93%

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

Validation — 15 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

Total

15

/

16

Passed

Repository

EvolvingLMMs-Lab/LLaVA-OneVision-2

Path

.opencode/skills/cu-lengths-attention-flow/SKILL.md

Reviewed

Table of Contents

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.