debug-quantized-kernel-accuracy

A troubleshooting guide for accuracy loss in low-precision computation, such as INT8, FP8, or other quantized formats. Quantization stores numbers with fewer bits to reduce memory or computation.

In plain words
What is it for?
Use it to compare a quantized kernel with an fp32 reference, investigate shape- or batch-dependent regressions, and assess whether an accuracy drop comes from a specific step.
Why use it?
It helps locate the exact calculation that introduces more error than expected, instead of treating the whole model as the problem.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/tensormux/kernel-skills/debug-quantized-kernel-accuracy
Any agent
npx skills add tensormux/kernel-skills --skill debug-quantized-kernel-accuracy
Clone the repo
git clone --depth 1 https://github.com/tensormux/kernel-skills

Made for: Claude Code, Codex.

Per session 0 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 3,146 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00000 $0.03146
Opus 5 $0.00000 $0.01573
Sonnet 5 $0.00000 $0.00629
Haiku 4.5 $0.00000 $0.00315

Measured 2d ago against content hash 370353c9092b, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

debug-quantized-kernel-accuracy scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

skills/quantization/debug-quantized-kernel-accuracy/SKILL.md · 127 lines

How it starts

The opening of the file, as written. The whole thing — 127 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Skill: Debug Quantized Kernel Accuracy

Purpose

Guide the agent through a systematic process for diagnosing and isolating accuracy degradation in a quantized (INT8, FP8, or low-bit) kernel, from measuring the error to identifying the specific computational step responsible.

Use this when

  • A quantized kernel produces outputs that differ from the fp32 reference by more than the expected quantization error bound.
  • A model using quantized kernels shows accuracy degradation that exceeds what is expected for the chosen quantization scheme.
  • A quantization refactor introduced a regression and the specific step that broke is not obvious.
  • Debugging a quantized kernel that works correctly on some input shapes or batch sizes but fails on others.

Do not use this when

  • The error is within the expected quantization error bound (approximately 0.5 * scale per element for well-calibrated INT8) and the downstream task accuracy loss is acceptable.
  • The issue is clearly a non-accuracy bug (segfault, wrong shape, miscompilation) — fix the structural bug first.
  • The degradation is due to model-level quantization sensitivity (certain layers or operators being inherently sensitive to quantization), which requires a quantization-aware training or mixed-precision approach rather than kernel debugging.

Inputs the agent should gather first

  • The exact mathematical specification of what the quantized kernel is supposed to compute, written in terms of the original unquantized operation.
  • The quantization scheme: per-tensor, per-channel, or per-token; symmetric or asymmetric; INT8, INT4, or FP8; signed or unsigned range.
  • The scale computation method: offline calibration, per-batch dynamic quantization, or per-token dynamic quantization.
  • The accumulation dtype: INT32, FP32, FP16, or FP8.
  • The dequantization epilogue: where is scale applied, in what order, and what is the output dtype.
  • A fp32 reference output for the same inputs (required for comparison).
  • Whether the error is consistent across runs (deterministic) or varies (stochastic — possible race condition or non-deterministic reduction).

Read the full file on GitHub · 127 lines

Files

What ships with it

1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 127 lines · 0 tokens per session scan A 370353c9092b

Subscribe to this mod's changes

debug-quantized-kernel-accuracy is a skill published in the GitHub repository tensormux/kernel-skills (72 stars, last pushed 2mo ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 3,146 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

design-impl-audit

The Design-vs-Implementation Audit Copilot. A universal skill for any project, any language, any design document. Before work starts, it reminds engineers what must be delivered; after code lands, it checks what was actually delivered, what drifted, and what silently went beyond the design. Feed it a design doc plus a…

MagicKidd/Rokid-agentic-workflow · 0 tokens

skill-creator

Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.

MagicKidd/Rokid-agentic-workflow · 45 tokens

diagnose

Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression.

MagicKidd/Rokid-agentic-workflow · 66 tokens

prompt-engineering

Prompt engineering techniques and patterns. Use when writing agent commands, hooks, skills, subagent prompts, or any LLM interaction: optimizing prompts, improving output reliability, and designing production-grade prompt templates. Trigger words: prompt engineering, prompt, prompt optimization, LLM interaction.

MagicKidd/Rokid-agentic-workflow · 0 tokens

setup-matt-pocock-skills

Sets up an ## Agent skills block in AGENTS.md/CLAUDE.md and docs/agents/ so the engineering skills know this repo's issue tracker (GitHub or local markdown), triage label vocabulary, and domain doc layout. Run before first use of to-issues, to-prd, triage, diagnose, tdd, improve-codebase-architecture, or zoom-out — or…

MagicKidd/Rokid-agentic-workflow · 124 tokens

brainstorming

Brainstorming design process. Must be used before any creative work - creating features, building components, adding functionality, or modifying behavior. Transforms ideas into complete designs and specifications through collaborative dialogue. Trigger words: brainstorm, design discussion, requirement analysis…

MagicKidd/Rokid-agentic-workflow · 0 tokens