huggingface-accelerate

Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.

Quality

68%

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/accelerate/SKILL.md

Quality

Content

71%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 well-structured, highly actionable overview with exemplary progressive disclosure into three real reference files. Its weaknesses are redundancy — the multi-GPU workflow restates the quick-start example and repeats prepare() boilerplate across every workflow — and the absence of explicit validation checkpoints in the launch/checkpoint workflows.

Suggestions

Delete or drastically shorten Workflow 1's 'Original script'/'With Accelerate' pair, which fully duplicates the quick-start conversion; point back to it instead.

Show the prepare() and backward() calls only once and note 'same pattern as quick start' in subsequent workflows to cut ~60 lines of repeated boilerplate.

Add a verification step after distributed launch and checkpointing, e.g. checking torch.distributed.get_world_size() or comparing state_dict hashes across ranks before/after accelerator.save_state().

Dimension	Reasoning	Score
Conciseness	The body is code-dense with no conceptual padding Claude already knows, but Workflow 1 ('From single GPU to multi-GPU') duplicates the entire quick-start conversion example (~40 lines), and each workflow repeats the same prepare()/backward() boilerplate with filler comments like '# Everything else is automatic!'. This fits 'mostly efficient but could be tightened' better than the 'minor instances' of level 4.	3 / 5
Actionability	Concrete, mostly copy-paste ready code and exact CLI commands ('accelerate launch --multi_gpu --num_processes 8 train.py', full DeepSpeed JSON config). Minor gaps keep it below 5: the quick-start snippet is diff-style (+/-) rather than directly runnable, and the workflow scripts reference an undefined 'dataset' and incomplete forward passes.	4 / 5
Workflow Clarity	The core sequence (convert script → interactive config → launch) is clear and well-ordered, with a dedicated 'Common issues' troubleshooting section serving as error-recovery guidance. It misses 5 because there are no explicit validation/verification checkpoints (e.g., confirming all ranks are alive or a smoke-test command) after risky steps like distributed checkpointing or multi-node launches.	4 / 5
Progressive Disclosure	The body is an overview with clearly signaled, one-level-deep references — all three linked files (references/megatron-integration.md, references/custom-plugins.md, references/performance.md) exist and contain the promised material — and each link is annotated with what it covers. Advanced detail is appropriately split out and navigation is easy.	5 / 5
	Total	16 / 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.

A specific, capability-dense description with strong natural trigger keywords for the distributed-training domain. Its main defect is the complete absence of explicit 'when to use' guidance, which caps completeness, plus mild buzz ('Simplest', 'HuggingFace ecosystem standard').

Suggestions

Append an explicit trigger clause, e.g. 'Use when the user wants to distribute or scale PyTorch training across multiple GPUs/nodes, or mentions DDP, DeepSpeed, FSDP, or accelerate launch.'

Replace the unverifiable over-claim 'HuggingFace ecosystem standard' (and 'Simplest') with concrete facts, e.g. 'Used by HuggingFace Transformers, TRL, and PEFT.'

Add common phrasing variations users say, such as 'multi-GPU', 'multi-node', or 'data parallel', to strengthen trigger matching.

Dimension	Reasoning	Score
Specificity	Lists several specific capabilities — 'Automatic device placement, mixed precision (FP16/BF16/FP8)', 'Interactive config, single launch command', 'Unified API for DeepSpeed/FSDP/Megatron/DDP' — matching the 'several specific actions; minor gaps' anchor. It falls short of 5 because 'HuggingFace ecosystem standard' is an over-claim/buzz phrase rather than a concrete capability, and it is above 3 because it goes well beyond naming the domain with 1-2 actions.	4 / 5
Completeness	The 'what' is clear and concrete, but there is no 'Use when...' clause or equivalent explicit trigger guidance anywhere ('HuggingFace ecosystem standard' does not state when to use the skill). Per the judging guidelines, a missing 'Use when...' clause caps completeness at 3.	3 / 5
Trigger Term Quality	Good natural keyword coverage: 'distributed training', 'PyTorch script', 'DeepSpeed', 'FSDP', 'Megatron', 'DDP', 'mixed precision' — phrases users would actually say. Not 5 because common variations like 'multi-GPU', 'multi-node', 'data parallel', or 'training script' phrasings are missing; not 3 because the included terms are natural rather than jargon-only.	4 / 5
Distinctiveness Conflict Risk	Clear niche — unified distributed training across named backends — with distinct triggers, so it is mostly distinguishable from sibling skills. Minor overlap risk remains with dedicated DeepSpeed/FSDP/Megatron or HuggingFace-Trainer skills since those backend names appear here too, keeping it below 5.	4 / 5
	Total	15 / 20 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

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

	Total	14 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/accelerate/SKILL.md
Commit: 4082a2e

Reviewed: 1 day 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.