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

Mojo module

config

GFX950 attention config.

Supports both prefill (token_gen=False) and decode (token_gen=True).

Matches amd/mha.mojo config target: full_kv=True, depth_padded=False for both. Prefill: double_buffer=True. Decode: double_buffer=False, double_buffer_k_only when BN<=64, shared_kv only at depth>256 (SMEM budget).

Structs

• ​AMDStructuredConfig: Holds the tile layout and indexing helpers for GFX950 structured attention.

Functions

• ​decode_mma_shape: Return the MFMA shape the gfx950 decode kernels use for this shape.
• ​mha_decode_fold_tile_q_seq_len: Query token SLOTS the fold gives a width, padding narrow ones up.
• ​mha_decode_fold_wide_mma: Whether the MHA decode token fold runs the 32x32x64 MFMA.