CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

pytorch-fsdp2

Adds PyTorch FSDP2 (fully_shard) to training scripts with correct init, sharding, mixed precision/offload config, and distributed checkpointing. Use when models exceed single-GPU memory or when you need DTensor-based sharding with DeviceMesh. Also covers TorchTitan (guides in references/).

Quality

78%

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

Fix and improve this skill with Tessl

tessl review fix ./skills/distributed-training/pytorch-fsdp2/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

67%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, actionable overview with concrete FSDP2-specific guidance and verified one-level references for the core path. Its main weaknesses are repetition across the contract/checklist/debug/common-issues sections and the nested torchtitan reference structure.

Suggestions

Consolidate the repeated rules (bottom-up fully_shard, model(input) vs forward(), optimizer after sharding) into a single authoritative section and have the Debug checklist and Common issues sections link to it rather than restating it.

Add inline validation checkpoints to the step-by-step procedure, e.g., verify parameters are DTensors before creating the optimizer and run a single train step to confirm hooks fire before adding checkpointing.

Flatten or clearly boundary the merged TorchTitan guide (references/torchtitan/) so its nested references/ subdirectory does not create a second level of disclosure from SKILL.md, or inline its key file list in the Related tools table.

DimensionReasoningScore

Conciseness

The body is mostly efficient and skill-specific, but key rules (bottom-up sharding, model(input) vs forward(), optimizer-after-sharding) are repeated three or more times across the Contract, Step 3/6, Debug checklist, and Common issues sections, so it could be meaningfully tightened. Not 2: there is no padding explaining basics Claude already knows.

3 / 5

Actionability

Concrete, correct API fragments and commands appear throughout (dist.init_process_group(backend="nccl"), torch.cuda.set_device(int(os.environ["LOCAL_RANK"])), mp_policy=MixedPrecisionPolicy(param_dtype=...), dcp.save(...), torchrun --nproc_per_node), but no complete end-to-end executable script is inline and the reference implementation is only a labeled outline, leaving minor gaps. Not 3: the fragments are real executable API calls, not pseudocode.

4 / 5

Workflow Clarity

Steps 0-7 are clearly sequenced, and the Debug checklist plus symptom-to-fix mapping in Common issues provide checkpoints and recovery guidance. Not 5: validation is retrospective (debug after failure) rather than explicit inline checkpoints such as verifying parameters are DTensors before optimizer creation or confirming sharding took effect.

4 / 5

Progressive Disclosure

Each step clearly signals its reference file with what it covers, and all referenced files exist in references/, giving good structure. Not 5: references/torchtitan/overview.md carries its own references/ subdirectory (checkpoint.md, float8.md, etc.), creating a second nesting level from SKILL.md rather than strictly one-level-deep references.

4 / 5

Total

15

/

20

Passed

Description

82%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 strong description with explicit what/when structure and domain-specific trigger terms. Its main weakness is the second-person "when you need" phrasing (penalized under the rubric's voice rule) and missing common trigger synonyms like OOM or multi-GPU.

DimensionReasoningScore

Specificity

The description lists several concrete capability areas ("init, sharding, mixed precision/offload config, and distributed checkpointing"), which matches anchor 4, but the second-person trigger phrase "when you need DTensor-based sharding" triggers the rubric's mandated 1-point penalty for non-third-person voice, reducing it to 3.

3 / 5

Completeness

It explicitly answers both what ("Adds PyTorch FSDP2 (fully_shard) to training scripts with correct init, sharding, mixed precision/offload config, and distributed checkpointing") and when ("Use when models exceed single-GPU memory or when you need DTensor-based sharding with DeviceMesh") with concrete trigger phrases, matching the anchor-5 example structure.

5 / 5

Trigger Term Quality

Good natural keyword coverage ("FSDP2", "fully_shard", "single-GPU memory", "DTensor", "DeviceMesh", "TorchTitan"), but common user phrasings like "OOM", "out of memory", "multi-GPU", or "model doesn't fit" are missing, so it falls short of the comprehensive-synonym anchor 5.

4 / 5

Distinctiveness Conflict Risk

"PyTorch FSDP2 (fully_shard)" is a clearly distinct niche with specific trigger terms; overlap risk with adjacent skills (DDP, FSDP1) is minimal since those have different APIs and the description names the FSDP2-specific entry point.

5 / 5

Total

17

/

20

Passed

Validation

81%

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

CriteriaDescriptionResult

frontmatter_unknown_keys

Unknown frontmatter key(s) found; consider removing or moving to metadata

Warning

relative_links

Relative link issues: 1 deeper-than-1-level

Warning

referenced_paths_exist

Referenced path issues: 2 deeper-than-1-level

Warning

Total

13

/

16

Passed

Repository

OpenLAIR/dr-claw

Path

skills/distributed-training/pytorch-fsdp2/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.