gpu perf engineering resources

The Problem

GPU performance engineering for AI inference is a fragmented field. Core knowledge lives across vendor docs, academic papers, engineering blogs, and production system source code. Engineers entering the field—or trying to build a systematic mental model—spend weeks assembling a reading list. The curated resources that do exist tend to be shallow link dumps without a pedagogical ordering or a source-quality policy.

What This Does

This repository is a structured, opinionated reading list for learning AI performance engineering, ordered from a single inference request down to GPU kernels and up to distributed systems. The README.md organizes roughly 100 resources into six tiers: GPU fundamentals, kernel optimization, programming models/profiling, inference engines, distributed inference, and current hardware. A "Start here" section gives newcomers a minimal reading path before they branch into the reference list.

The repo adds a quality-control layer beyond a static list. CONTRIBUTING.md and the .github/ISSUE_TEMPLATE/ files (add-resource.md, broken-link.md, quality-concern.md) define a source policy and a contribution workflow. A Python script, scripts/check_guide.py, enforces link and structural checks, and two GitHub Actions workflows (links.yml, quality.yml) run those checks automatically.

How It Is Wired

The repository has no application runtime; it is a documentation project with automated validation. The entry points are the GitHub Actions workflows in .github/workflows/. quality.yml runs scripts/check_guide.py, which validates the README's structure and link integrity. links.yml performs external link checking. Both workflows run on pull requests and scheduled intervals, so the "execution" is a CI pipeline that validates Markdown content—no database, network calls to external sites, or filesystem writes beyond the CI container.

File responsibilities:

• README.md — the core resource list and the only substantive content
• scripts/check_guide.py — structural and link validation logic
• .github/workflows/links.yml, quality.yml — CI triggers for the checks
• .github/ISSUE_TEMPLATE/ — contribution intake forms
• CONTRIBUTING.md — contribution policy and expectations

The blast radius is concentrated in scripts/check_guide.py and the two workflows; a broken check blocks all contributions. The rest of the repo is inert content.

How To Use It

Setup: Clone the repository. No dependencies or build step—the validation script is a single Python file.

git clone https://github.com/moses-y/gpu-perf-engineering-resources
cd gpu-perf-engineering-resources

Configuration: No environment variables or config files. The GitHub Actions workflows define their own Python environment in .github/workflows/quality.yml.

Running it: Start with the README's "Start here" section. To validate contributions locally, run:

python scripts/check_guide.py

Real-World Use

An inference platform team onboarding a new engineer uses this as a structured curriculum. The engineer reads the "Start here" path, then uses the tiered list to deep-dive into kernel optimization (Triton, CUTLASS) or production systems (vLLM, TensorRT-LLM) as their work requires. The team contributes new resources through the issue templates, and CI validates links and formatting before merge.

Code Health & Issues

Static analysis has not run on this repository, so no measured findings are available.

Structural observations:

• Med - No LICENSE file - usage and redistribution rights are unclear; the upstream repo (wafer-ai) is MIT-licensed but this fork does not carry it
• Low - No test suite - scripts/check_guide.py has no accompanying tests; a structural change to the README could break the script without CI catching it
• Low - Content-only repo - the value is entirely in the README; the validation script and workflows are thin wrappers

The Bottom Line

A well-organized, high-quality reading list with a sensible source policy and light CI validation. The pedagogical ordering is its main strength—it works as a curriculum, not just a bookmark dump. The lack of a license is the only real blocker for reuse; the fork adds nothing beyond the upstream content. Use it as a reference and onboarding tool, but check the upstream wafer-ai/gpu-perf-engineering-resources for the maintained version.