Mojo module

mha_structured

TileTensor-based MHA prefill kernel for gfx950.

Uses TileTensor instead of LayoutTensor for the KV buffer layer, eliminating RuntimeLayout overhead and reducing VGPR usage by 2 (254 -> 252), which gives the LLVM scheduler more freedom to produce better instruction ordering. Supports depth=64, 128, 256.

Functions

• ​barrier:
• ​block_sync_lds_direct_load:
• ​set_priority:

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