awq-quantization

Activation-aware weight quantization for 4-bit LLM compression with 3x speedup and minimal accuracy loss. Use when deploying large models (7B-70B) on limited GPU memory, when you need faster inference than GPTQ with better accuracy preservation, or for instruction-tuned and multimodal models. MLSys 2024 Best Paper Award winner.

Quality

79%

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/awq/SKILL.md

Quality

Content

75%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.

Highly actionable and well-structured body with executable code and useful comparison tables, but progressive disclosure is weak: two dedicated reference files exist yet are not signaled or linked from the body, and overlapping content is inlined instead.

Suggestions

Replace the inline 'Common issues' section with a pointer to references/troubleshooting.md, and link references/advanced-usage.md from the relevant sections (e.g., kernel backends, custom calibration).

Add an explicit verification step to the quantize-your-own-model workflow, e.g. run a perplexity/accuracy check on the saved model before deployment.

De-duplicate the GEMM quant_config block by defining it once and referencing it from the GEMV/Marlin variants.

Dimension	Reasoning	Score
Conciseness	Dense, code-heavy body with scannable tables that mostly earns its tokens; minor redundancy such as the GEMM quant_config block repeated across sections keeps it just below fully lean.	4 / 5
Actionability	Multiple complete, copy-paste-ready code blocks cover loading, quantizing, vLLM, multi-GPU, and custom calibration, with specific configs and concrete examples for the common cases.	5 / 5
Workflow Clarity	The quantize-your-own-model flow (load → config → quantize → save) is a clear, ordered sequence with timing guidance, though it lacks an explicit verify-accuracy checkpoint after quantization.	4 / 5
Progressive Disclosure	Bundle files references/advanced-usage.md and references/troubleshooting.md exist but are never linked from the body, while the body inlines a 'Common issues' section and advanced config that clearly belong in those files—references are buried and content is mis-placed.	2 / 5
	Total	15 / 20 Passed

Description

83%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, specific description that clearly states what AWQ does and when to use it with concrete trigger conditions. The trailing 'MLSys 2024 Best Paper Award winner' is a mild over-claim/badge that adds little trigger value.

Dimension	Reasoning	Score
Specificity	Names concrete capabilities—'Activation-aware weight quantization for 4-bit LLM compression', '3x speedup', 'minimal accuracy loss'—with several specific actions, though 'compression' is the central verb and the rest are outcomes.	4 / 5
Completeness	Explicitly answers both what ('4-bit LLM compression with 3x speedup and minimal accuracy loss') and when ('Use when deploying large models (7B-70B) on limited GPU memory, when you need faster inference than GPTQ...') with concrete trigger phrases.	5 / 5
Trigger Term Quality	Strong natural terms like '4-bit LLM compression', 'limited GPU memory', 'faster inference than GPTQ', and 'multimodal models' cover what users would say, with only minor synonyms missing.	4 / 5
Distinctiveness Conflict Risk	Clear niche as a distinct quantization method explicitly contrasted with GPTQ, with minimal conflict risk; minor overlap possible with sibling quantization skills.	4 / 5
	Total	17 / 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: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/awq/SKILL.md
Commit: 3a6c3a9

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.