gptq

Post-training 4-bit quantization for LLMs with minimal accuracy loss. Use for deploying large models (70B, 405B) on consumer GPUs, when you need 4× memory reduction with <2% perplexity degradation, or for faster inference (3-4× speedup) vs FP16. Integrates with transformers and PEFT for QLoRA fine-tuning.

Quality

77%

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 ./10-optimization/gptq/SKILL.md

Quality

Content

65%

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

A highly actionable, code-rich skill body, but it runs long for a SKILL.md overview, repeats config blocks, and lacks validation checkpoints in the quantization workflow. Heavy inline detail that belongs in the reference files weakens progressive disclosure.

Suggestions

Add a verification step to the quantize-your-own-model workflow (e.g. reload the saved model and run a quick perplexity/sanity check before declaring success) to lift workflow_clarity to 3.

Collapse the three near-duplicate BaseQuantizeConfig blocks (Standard / High accuracy / Maximum accuracy) into one annotated example with a parameter trade-off table, removing redundant code.

Move the performance benchmarks, kernel-backend details, and multi-GPU/CPU-offload patterns into references/ so SKILL.md stays a concise overview pointing one level deep.

Dimension	Reasoning	Score
Conciseness	The ~440-line body is action-dense rather than pure fluff, but it includes near-duplicate BaseQuantizeConfig blocks (Standard / High accuracy / Maximum accuracy) and a 'How GPTQ works' explanation of concepts Claude already knows, so it could be tightened rather than being fully lean.	2 / 3
Actionability	Provides extensive copy-paste-ready, executable code for loading, quantizing, kernel backends, transformers/PEFT integration, and multi-GPU deployment with specific model names and parameters.	3 / 3
Workflow Clarity	Installation -> load -> quantize steps are present, but the quantize-your-own-model workflow has no validation/verification checkpoint (e.g. reload and check perplexity, confirm the saved model loads), which caps workflow clarity at 2 for a risky batch operation.	2 / 3
Progressive Disclosure	References to calibration.md, integration.md, and troubleshooting.md are real files and clearly signaled one level deep, but substantial inline content (benchmarks, kernel backends, multi-GPU/CPU-offload patterns, supported-models list) could be moved into those references rather than living in SKILL.md.	2 / 3
	Total	9 / 12 Passed

Description

90%

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 states concrete capabilities and gives explicit use-when triggers with quantified thresholds. The only issue is the second-person 'when you need' phrasing, which the rubric penalizes on the specificity dimension.

Dimension	Reasoning	Score
Specificity	Lists several concrete actions ('4-bit quantization', 'QLoRA fine-tuning', 'deploying large models'), matching the multiple-specific-actions anchor; however the second-person phrasing 'when you need 4x memory reduction' triggers the rubric's -1 specificity penalty for non-third-person voice.	2 / 3
Completeness	Clearly answers both 'what' (post-training 4-bit quantization for LLMs) and 'when' via an explicit 'Use for deploying large models ... when you need 4x memory reduction with <2% perplexity degradation' trigger clause.	3 / 3
Trigger Term Quality	Natural deploy-time keywords a user would actually say are well covered: '4-bit quantization', 'consumer GPUs', 'memory reduction', 'perplexity', 'QLoRA', 'fine-tuning', 'inference'.	3 / 3
Distinctiveness Conflict Risk	Occupies a clear niche (GPTQ group-wise 4-bit quantization + QLoRA integration) with distinct triggers unlikely to fire for unrelated skills.	3 / 3
	Total	11 / 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: Orchestra-Research/AI-Research-SKILLs
Path: 10-optimization/gptq/SKILL.md
Commit: 773a529

Reviewed: about 6 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.