pytorch-fsdp

Expert guidance for Fully Sharded Data Parallel training with PyTorch FSDP - parameter sharding, mixed precision, CPU offloading, FSDP2

Quality

58%

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 ./optional-skills/mlops/pytorch-fsdp/SKILL.md

Quality

Content

35%

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 auto-generated documentation dump that is verbose, only partially actionable, and structurally weak — it inlines reference-grade content and points to reference files that do not exist. It needs heavy distillation into guidance plus genuine reference files.

Suggestions

Replace the scraped Patterns wall with a concise quick-start: a minimal FSDP setup snippet (init_process_group, FullyShardedDataParallel wrapping, checkpoint saving) plus 3-4 copy-paste patterns, moving the rest into references/.

Create the referenced files (getting_started/tutorials, api, guides) or remove those pointers, and ensure each reference path in the body maps to a real file in references/.

Add an explicit workflow with validation checkpoints (e.g., verify world_size, check sharding with sum_of_full_params, validate checkpoint load) for the destructive/batch operations FSDP involves.

Dimension	Reasoning	Score
Conciseness	The body is a monolithic wall of scraped documentation (Patterns 1-8), including 'Created On / Last Updated On' metadata and full inline API reference prose that explains concepts Claude already knows. Not a 2 because the bulk is unedited documentation dump with no token-efficiency discipline.	1 / 3
Actionability	Some executable code exists (e.g., the init_device_mesh example in Pattern 4), but most patterns pair long prose with a code fence containing only a bare identifier ('Join', 'torch.distributed', 'new_group()', 'NCCL', 'True'). Not a 3 because those code fences are not copy-paste-ready guidance, and not a 1 because real executable examples do appear.	2 / 3
Workflow Clarity	'Working with This Skill' lists loose steps (start with getting_started, use category files) but has no validation checkpoints or feedback loops, and references files that do not exist (getting_started, tutorials, api, guides). Not a 3 because of missing validation/verification, and not a 1 because some sequenced guidance is present.	2 / 3
Progressive Disclosure	Headers provide structure and the body correctly points to the real one-level reference (other.md), but the bulk of detailed documentation is dumped inline in SKILL.md rather than split into reference files, and several referenced files (getting_started, tutorials, api, guides) are absent. Not a 3 because of the inline documentation wall; not a 1 because there is a genuine reference file and navigation intent.	2 / 3
	Total	7 / 12 Passed

Description

82%

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 uses strong, natural trigger terms for a well-defined niche, but it omits an explicit 'Use when...' clause, capping completeness. Adding a trigger guidance clause would round it out.

Suggestions

Add an explicit 'Use when...' clause (e.g., 'Use when setting up Fully Sharded Data Parallel training, sharding model parameters across GPUs, or debugging FSDP/FSDP2 issues') to lift completeness to 3.

Lead with verb-driven actions ('Shard parameters, configure mixed precision, offload to CPU') to make capabilities read as actions rather than a noun list.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete capabilities — 'parameter sharding, mixed precision, CPU offloading, FSDP2' — naming specific FSDP techniques rather than vague abstractions. Not a 2 because it enumerates several distinct, concrete sub-capabilities beyond just naming the domain.	3 / 3
Completeness	Answers 'what' (expert guidance for FSDP training with the listed features) but has no explicit 'Use when...' trigger clause. Capped at 2 per the guideline that a missing explicit trigger clause limits completeness; it is not a 1 because the 'what' is clearly stated.	2 / 3
Trigger Term Quality	Good coverage of natural terms a user would say: 'Fully Sharded Data Parallel', 'PyTorch FSDP', 'mixed precision', 'CPU offloading', 'FSDP2'. Not a 2 because it includes both the full name and common abbreviations/variations users actually use.	3 / 3
Distinctiveness Conflict Risk	Targets a clear, specialized niche (PyTorch FSDP / Fully Sharded Data Parallel) with distinct triggers unlikely to overlap with unrelated skills. Not a 2 because the specificity of FSDP makes mis-triggering unlikely.	3 / 3
	Total	11 / 12 Passed

Validation

81%

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 — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning
metadata_field	'metadata' should map string keys to string values	Warning
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	13 / 16 Passed

Repository: NousResearch/hermes-agent
Path: optional-skills/mlops/pytorch-fsdp/SKILL.md
Commit: 4c9628e

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.