CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

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

81%

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

SKILL.md
Quality
Evals
Security

Quality

Content

88%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 an exceptionally actionable reference: decision tables, copy-paste config, a runnable command, explicit verification criteria, and a failure-diagnosis matrix with feedback loops. Its main weakness is structure — all detail lives inline in a long SKILL.md with no bundle reference files to split depth from overview.

Suggestions

Move the Code Anchors section (quoted Mcore/Bridge source excerpts) into a references/ file (e.g., references/code-anchors.md) and keep only file:line pointers in SKILL.md to lighten the always-loaded body.

Consolidate the Enablement code blocks and the Minimal Runnable Command into a single end-to-end example to remove the minor duplication between them.

Fix the pytest verification command — it passes two -k flags ('-k "cpu_offload"' twice), where the second silently overrides the first; use a single -k expression or separate pytest invocations per file.

DimensionReasoningScore

Conciseness

Dense and tabular with no padding or explanation of concepts Claude already knows, but the ~230-line body inlines full source-code anchor excerpts and a complete parameter table, with minor overlap between the Enablement and Minimal Runnable Command sections. Fits 'efficient; minor instances that could be trimmed' rather than 'every token earns its place'.

4 / 5

Actionability

Fully executable, copy-paste-ready guidance throughout: concrete Python config snippets, CLI overrides, a runnable recipe command ('uv run python scripts/training/run_recipe.py ...'), and specific pytest invocations covering the common cases.

5 / 5

Workflow Clarity

A clear sequence flows from the Quick Decision table through Enablement to a Minimal Runnable Command, followed by an explicit Verification section with success criteria, and a Failure Diagnosis table providing symptom→cause→confirm→fix feedback loops for error recovery.

5 / 5

Progressive Disclosure

Well-sectioned body with a clearly signaled References section, but the deep detail (config parameter tables, ~70 lines of quoted code anchors) is inlined rather than split into bundle reference files, which would better match the one-level-deep reference ideal.

4 / 5

Total

18

/

20

Passed

Description

66%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 clearly and concretely states what the skill does with domain-specific terminology, but omits explicit 'when to use' trigger guidance within the description field itself and misses a few natural keyword variants. Trigger guidance exists only in the separate when_to_use field.

Suggestions

Fold an explicit 'Use when...' clause into the description field (e.g., 'Use when enabling CPU offload to reduce GPU memory or debugging OOM after offloading config changes') so the description stands alone.

Add natural keyword variants users would say, such as 'CPU offload', 'free up GPU memory', or '.yaml offloading config', alongside the existing technical terms.

Mention the commit-investigation use case (debugging OOM/crashes after CPU offloading config changes) in the description itself, since that is a distinct triggering scenario.

DimensionReasoningScore

Specificity

Names several concrete capabilities — 'Validate and use CPU offloading', 'layer-level activation offloading', 'fractional optimizer state offloading with HybridDeviceOptimizer' — with only minor gaps in coverage (e.g., what validation entails).

4 / 5

Completeness

The 'what' is clear and concrete, but the description field itself contains no 'Use when...' clause or equivalent trigger guidance (that lives in the separate when_to_use field), which caps completeness at 3.

3 / 5

Trigger Term Quality

Good keyword coverage ('CPU offloading', 'activation offloading', 'optimizer state offloading', 'HybridDeviceOptimizer', 'Megatron Bridge'), but natural variants like 'CPU offload' or 'reduce GPU memory' are missing, so it falls short of the comprehensive-synonym anchor.

4 / 5

Distinctiveness Conflict Risk

Clear niche anchored by 'Megatron Bridge' and 'HybridDeviceOptimizer', with only minor overlap risk against generic CPU-offload/ZeRO-offload topics in the broader training space.

4 / 5

Total

15

/

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

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

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.