For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo module

mma

TileTensor-native matrix-multiply-accumulate (MMA) operations for AMD GPU attention kernels.

Defines TiledMmaOp, a thin wrapper over the raw GPU MMA intrinsic that operates directly on register-resident TileTensor fragments, and KVMmaOp, which owns the K or V operand register tile and its shared-memory-to-register load logic for the two sequential attention GEMMs (P = Q @ K^T and O += P @ V).

Structs

• ​KVMmaOp: Owns the K or V operand register tile and its SMEM→reg load logic.
• ​TiledMmaOp: TileTensor-native MMA operation for AMD attention kernels.

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit