deepspeed

Expert guidance for distributed training with DeepSpeed - ZeRO optimization stages, pipeline parallelism, FP16/BF16/FP8, 1-bit Adam, sparse attention

Quality

40%

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 ./backend/cli/skills/ml-training/deepspeed/SKILL.md

Quality

Content

22%

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

The body is essentially a scraped-documentation dump: massive unformatted tutorial text inlined into SKILL.md, with only a thin shell of real structure around it. Executable code exists but is poorly organized, and progressive disclosure is undermined by useless reference descriptions and inline content duplication.

Suggestions

Replace the inlined 'Common Patterns' tutorial dumps with a concise overview (a few lines + one minimal code example per topic) and move the full tutorial text into the appropriate reference files, pointing to them with one-level-deep links.

Add a real step-by-step workflow for common DeepSpeed tasks (e.g. enable ZeRO, launch with `deepspeed --hostfile`, validate with `ds_report`) with explicit validation/feedback checkpoints.

Rewrite the Reference Files list with meaningful descriptions of what each file contains and which task it supports, and add the missing index.md entry.

Dimension	Reasoning	Score
Conciseness	The 'Common Patterns' section pastes entire scraped tutorials (DeepNVMe, MoE, LRRT) as single unformatted paragraphs, explains concepts Claude already knows (libaio, Python tab completion), repeats the 1GB-tensor example four times, and includes 'Updated: November 5, 2025 / Previous Next' boilerplate — a verbose, padded wall of text.	1 / 3
Actionability	It does contain executable code and commands (aio_handle/gds_handle creation, sync_pwrite/async_pwrite, deepspeed_config JSON blocks, ds_nvme_tune, `deepspeed --hostfile`), but much is embedded in REPL `>>>` sessions and buried inside documentation prose rather than presented as clean, copy-paste-ready snippets.	2 / 3
Workflow Clarity	There is no sequenced workflow for actually using DeepSpeed; 'Working with This Skill' offers only vague directional hints ('Start with... tutorials reference files', 'Use the appropriate category reference file') and there are no validation checkpoints for risky training/compilation operations.	1 / 3
Progressive Disclosure	A reference file list and section structure exist, but the per-file descriptions are tautological ('08.md - 08 documentation', 'mii.md - Mii documentation', 'other.md - Other documentation'), index.md is omitted from the listing, and full tutorial content is dumped inline in the body instead of being progressively disclosed into the reference files.	2 / 3
	Total	6 / 12 Passed

Description

57%

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 to DeepSpeed and lists relevant features, but it lacks concrete action verbs and an explicit 'Use when' trigger clause. It is distinctive but incomplete on the when-to-use side.

Suggestions

Rewrite feature labels as concrete actions, e.g. 'Configure ZeRO optimization stages, set up pipeline parallelism, and enable FP16/BF16/FP8 mixed-precision training'.

Add an explicit trigger clause such as 'Use when the user is training or debugging large models with DeepSpeed, configuring multi-GPU/multi-node training, or tuning ZeRO/optimizer settings'.

Broaden trigger coverage with natural phrasings users actually say, like 'large model training', 'multi-GPU training', and 'distributed PyTorch training'.

Dimension	Reasoning	Score
Specificity	Names the domain ('distributed training with DeepSpeed') and several features ('ZeRO optimization stages, pipeline parallelism, FP16/BF16/FP8, 1-bit Adam, sparse attention'), but these are feature labels rather than concrete verbs, so it does not reach the 'lists multiple specific concrete actions' bar.	2 / 3
Completeness	It clearly states what the skill covers but provides no 'Use when...' clause or equivalent explicit trigger guidance, so per the rubric guideline completeness is capped at 2.	2 / 3
Trigger Term Quality	It includes natural terms a DeepSpeed user would say ('DeepSpeed', 'distributed training', 'pipeline parallelism') but leans heavily on jargon ('ZeRO', '1-bit Adam', 'FP8') and omits common variations like 'large model training' or 'multi-GPU training'.	2 / 3
Distinctiveness Conflict Risk	It targets a named, specific tool (DeepSpeed) with distinctive technical triggers (ZeRO, 1-bit Adam), giving it a clear niche unlikely to fire for unrelated skills.	3 / 3
	Total	9 / 12 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
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/deepspeed/SKILL.md
Commit: 52845c3

Reviewed: about 10 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.