v09_triton_basics

What This Version Introduces

This version introduces Triton with small kernels before replacing the full attention path.

The goal is to explain Triton execution structure without mixing in the full complexity of paged attention.

Why Introduce It

Jumping directly from PyTorch paged GQA to a full Triton paged attention kernel would hide too many new ideas at once: program IDs, block sizes, masks, pointer arithmetic, compile-time constants, and launch grids.

Core Principle

A Triton kernel is written as a program that operates on a tile of data. Many programs run in parallel, and each program computes offsets from its program IDs.

The implementation keeps these ideas visible:

• Program IDs map work to tiles.
• Masks protect boundary loads and stores.
• Block sizes are explicit.
• Kernel launch shape becomes part of the design.

Files To Compare

• The new Triton kernel files.
• The PyTorch reference function it is checked against.
• Any test or smoke script that compares outputs.

Remaining Tradeoff

This version should stay intentionally small. Its job is confidence-building, not performance heroics.