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 function
fa4_o_store_swizzle
def fa4_o_store_swizzle[output_type: DType, config: FA4Config[config.qkv_dtype, rope_dtype_=config.rope_dtype_, scale_dtype_=config.scale_dtype_]]() -> TensorMapSwizzle
Swizzle of the FA4 O-TMA store, i.e. its block width.
SWIZZLE_128B (64 bf16 columns per box row) unless split-K reduce-scatters O across the cluster: every partition TMA-stores only its own depth band, so a block must not straddle two partitions' bands. There the widest block is chosen that leaves the widest band no wider than the 16 B-block split would, which keeps the reduce-scatter's critical path unchanged (d128: P=2 -> 128B, P=4 -> 64B, P=6 or 16 -> NONE).
Returns: