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 function

fa4_mma

def fa4_mma[MaskType: MHAMask, //, config: FA4Config[config.qkv_dtype, rope_dtype_=config.rope_dtype_, scale_dtype_=config.scale_dtype_], *, page_size: Int](smem: SM100AttentionSMem[config], tmem_addr: UInt32, seq_id: UInt32, score_row: UInt32, num_keys: UInt32, mask: MaskType)

Executes the FA4 MMA warp loop for SM100 Flash Attention.

Computes Q@K' into score TMEM (S0/S1) and P@V into output TMEM (O0/O1) across the KV tile sequence, coordinating with the load and softmax warps via barrier pipelines. Supports fused-KV and split-KV modes, single-Q and two-Q tile configurations, and partial-K contraction for paged sub-tiles (page_size zero disables paging; sub-BN enables partial-K).

Args:

smem (SM100AttentionSMem[config]): Shared memory allocator holding Q, K, V tiles and barriers.
tmem_addr (UInt32): Base TMEM address for this CTA's S/O accumulators.
seq_id (UInt32): Sequence index for the current query batch.
score_row (UInt32): Row offset of the query tile within the sequence.
num_keys (UInt32): Number of valid key columns to attend to.
mask (MaskType): Attention mask controlling tile iteration and masking bounds.