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 module

mla_decode_sparse_kv_bf16

SM100 (B200) sparse MLA decode kernel with BF16 KV cache.

K is loaded by a BF16 + SWIZZLE_128B gather4 TMA covering the full 576-element row (tile_width=576, box_w=64). OffsetPosition[sparse=True] overrides num_keys with the sparse topk. A dedicated 4-warp gather WG (warps 8-11) decodes each tile's row indices cooperatively into the double-buffered idx_smem, then round-robins the tile's 144 gather4 issues across its 4 elected lanes under a single expect_bytes.

Supports NullMask and CausalMask. Sliding-window attention is FP8-only.

Structs

• ​MLA_SM100_Decode_Sparse_KV_BF16:

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