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

attention

Attention struct for gfx950 MHA/MLA kernels (prefill + decode).

TileTensor-only API: no LayoutTensor in struct fields, constructor parameters, method signatures, or internal bridges. Mask application is delegated to MaskTileOp (see mask_op.mojo), which uses TileTensor.vectorize with local SIMD read/modify/write for per-element edits.

Inlines AMDStructuredConfig directly (no AttentionConfig trait indirection). Prefill kernels handle MMA inline; there is no _dma_loop/mma_qk/mma_pv helper on the struct.

Structs

• ​Attention: Holds the per-warp register, SMEM, and softmax state for a gfx950 MHA/MLA attention tile.