AI Compiler Study

A study workspace for the AI / GPU kernel-and-compiler landscape. It pairs a cross-linked knowledge base (wiki/) with four full upstream repos, vendored as git submodules, so every claim in the notes can be traced to real source files.

Layout

AI_Compiler_Study/
├── wiki/        ← the knowledge base. START HERE: wiki/index.md
├── modular/     Modular monorepo — the Mojo language + the MAX inference engine   (submodule)
├── triton/      OpenAI Triton — Python-embedded tile DSL                          (submodule)
├── FlagTree/    FlagTree — multi-backend Triton distribution (FlagOS)             (submodule)
└── tilelang/    TileLang — TVM-based tile DSL                                      (submodule)

Cloning

The four repos are submodules pinned to specific upstream commits. Clone recursively:

git clone --recursive git@github.com:wyanzhao/AI_Compiler_Study.git
# or, if you already cloned without --recursive:
git submodule update --init --recursive

If you only want the notes, a plain git clone gives you the wiki/ without the (large) submodule checkouts.

The one framing to keep in mind

These three subjects are not the same kind of thing:

• Triton / TileLang — kernel-authoring DSLs embedded in Python; you write one GPU kernel.
• Mojo — a general-purpose Python-family language; you write kernels, host code, whole apps.
• MAX — a graph compiler + serving engine (compare to vLLM / TensorRT-LLM / XLA, not to Triton).

See wiki/concepts/abstraction-ladder.md and wiki/comparisons/mojo-vs-triton-vs-tilelang.md.

Working in this repo

Start at wiki/index.md for the catalog and learning tracks. When you learn something new about these repos, extend the wiki following the rules in wiki/CLAUDE.md.