serving-llms-vllm

Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism.

Quality

78%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Critical

Do not install without reviewing

Fix and improve this skill with Tessl

tessl review fix ./skills/inference-serving/vllm/SKILL.md

Quality

Content

68%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 generally actionable and well-structured skill body with concrete code, clear sectioning, and clean one-level references. Its main weaknesses are placeholder code stubs and the absence of explicit validation feedback loops in the batch-inference and production-deployment workflows.

Suggestions

Add explicit validate→fix→retry checkpoints to Workflow 2 (batch inference) and Workflow 1 (production deployment), e.g. verify output schema/token counts before writing results.jsonl and re-run on failure.

Replace placeholder code stubs (e.g. '# Create test_load.py with sample requests' and Workflow 3 Step 4's comment-only block) with concrete, runnable examples.

Trim explanatory asides like 'vLLM handles batching internally / No need to manually chunk prompts' that restate what Claude already knows about the library.

Dimension	Reasoning	Score
Conciseness	Mostly efficient with copy-paste commands and code, but contains minor padding such as 'vLLM automatically batches requests for efficiency' and 'vLLM handles batching internally / No need to manually chunk prompts' that assume Claude lacks basic knowledge.	4 / 5
Actionability	Provides extensive executable code and CLI commands covering common cases, but a few blocks are stubs—'# Create test_load.py with sample requests' and Workflow 3 Step 4's '# Compare quantized vs non-quantized responses'—leaving minor gaps.	4 / 5
Workflow Clarity	Workflows are sequenced with checklists, but the batch-inference workflow (Workflow 2) and production-deployment workflow lack explicit validate→fix→retry feedback loops; per the rubric, missing validation in batch/destructive operations caps this dimension at 3.	3 / 5
Progressive Disclosure	Well-signaled one-level-deep references to four real files (server-deployment.md, optimization.md, quantization.md, troubleshooting.md) each with a descriptive label, though the three full workflows are fairly detailed inline rather than partially split out.	4 / 5
	Total	15 / 20 Passed

Description

88%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, third-person description that concretely states both capabilities and explicit use-when triggers, with good natural-language keyword coverage. The only weakness is slight overlap risk with neighboring inference-serving tools under its broad production-API trigger.

Dimension	Reasoning	Score
Specificity	Lists multiple concrete capabilities—'Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching', 'Supports OpenAI-compatible endpoints, quantization (GPTQ/AWQ/FP8), and tensor parallelism'—with comprehensive coverage of the serving domain.	5 / 5
Completeness	Explicitly answers both 'what' (serves LLMs via PagedAttention/continuous batching, OpenAI-compatible endpoints, quantization, tensor parallelism) and 'when' via a concrete 'Use when deploying production LLM APIs, optimizing inference latency/throughput, or serving models with limited GPU memory' trigger clause.	5 / 5
Trigger Term Quality	Includes natural phrases users would say such as 'deploying production LLM APIs', 'optimizing inference latency/throughput', and 'serving models with limited GPU memory', though a few common variations (e.g. 'LLM inference', 'model serving') are absent.	4 / 5
Distinctiveness Conflict Risk	Names vLLM-specific mechanisms (PagedAttention, continuous batching) giving it a clear niche, but the broad 'deploying production LLM APIs' trigger creates minor overlap risk with adjacent inference-serving skills like TensorRT-LLM or TGI.	4 / 5
	Total	18 / 20 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: OpenLAIR/dr-claw
Path: skills/inference-serving/vllm/SKILL.md
Commit: 4b35756

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.