nemo-mbridge-perf-cpu-offloading

Validate and use CPU offloading in Megatron Bridge, including layer-level activation offloading and fractional optimizer state offloading with HybridDeviceOptimizer.

Quality

70%

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

Fix and improve this skill with Tessl

tessl review fix ./skills/nemo-mbridge-perf-cpu-offloading/SKILL.md

Quality

Content

82%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.

The body is a tight, highly actionable reference: executable commands, decision tables, parameter reference, and a strong failure-diagnosis feedback loop. The main improvement is moving the inline upstream source snippets into a reference file for better progressive disclosure.

Suggestions

Move the multi-file upstream Code Anchors snippets into a dedicated reference file (e.g. references/code_anchors.md) and link to it from SKILL.md to reduce inline token load.

Add an explicit numbered 'workflow' checklist (choose mechanism -> enable -> run minimal command -> verify against success criteria -> consult Failure Diagnosis on error) to make the implicit section sequence a true validated workflow.

If the quoted transformer_config.py constraints are stable, consider replacing the raw source snippets with a concise summary table plus a single file:line pointer per constraint.

Dimension	Reasoning	Score
Conciseness	Lean and free of concept-explanation padding, with compact decision and parameter tables; the inline Code Anchors section quotes ~70 lines of upstream source that could be trimmed or moved to a reference file, keeping it just below the 5 anchor.	4 / 5
Actionability	Provides copy-paste config blocks, CLI overrides, a fully runnable 'uv run python scripts/training/run_recipe.py ...' command, and pytest verification commands covering the common cases.	5 / 5
Workflow Clarity	A coherent section-ordered flow (Quick Decision -> Enablement -> Run -> Verification) with explicit success criteria and a Failure Diagnosis table that supplies symptom->cause->confirm->fix feedback loops, though it is not a rigid numbered checklist.	4 / 5
Progressive Disclosure	Clearly signaled one-level-deep references (@docs/..., @skills/.../card.yaml) and clean section organization, but no bundle files exist and the Code Anchors content is all inlined rather than split into a separate reference.	4 / 5
	Total	17 / 20 Passed

Description

58%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.

The description is specific and well-targeted to a distinct niche, but the description field alone omits explicit 'when to use' trigger guidance (delegated to when_to_use) and lacks the natural user phrasings that would push trigger quality higher.

Suggestions

Fold a brief 'Use when ...' trigger clause with natural user phrasings ('move optimizer to CPU', 'reduce GPU memory', 'OOM after changing offload config') directly into the description field so it is self-contained.

Add common synonyms/file-flag variants users actually say (e.g. 'cpu_offloading', 'optimizer_cpu_offload', 'optimizer_offload_fraction') to the description rather than only to when_to_use.

Consider mentioning the diagnosis/investigation use case (OOM or crash after a config change) in the description to broaden natural trigger coverage.

Dimension	Reasoning	Score
Specificity	Names the domain and several concrete actions ('Validate and use CPU offloading', 'layer-level activation offloading', 'fractional optimizer state offloading with HybridDeviceOptimizer'), but leans technical and has minor coverage gaps relative to the comprehensive 5 anchor.	4 / 5
Completeness	Clearly states what the skill does but the description field itself lacks a 'Use when...' clause or equivalent trigger guidance; per the rubric, a missing trigger clause caps completeness at 3.	3 / 5
Trigger Term Quality	Contains relevant keywords ('CPU offloading', 'activation offloading', 'optimizer state offloading') but the natural user-facing synonyms and config flags ('move optimizer to CPU', 'optimizer_offload_fraction') live in the separate when_to_use field rather than the description.	3 / 5
Distinctiveness Conflict Risk	Targets a specific framework and named class (Megatron Bridge, HybridDeviceOptimizer) giving a clear niche, with only minor overlap risk against sibling nemo-mbridge-perf skills.	4 / 5
	Total	14 / 20 Passed

Validation

93%

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 — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
referenced_paths_exist	Referenced path issues: 1 missing, 1 deeper-than-1-level	Warning

	Total	15 / 16 Passed

Repository: NVIDIA/skills
Path: skills/nemo-mbridge-perf-cpu-offloading/SKILL.md
Commit: 994b870

Reviewed: ~15 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.