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
decode_head_tile
def decode_head_tile[config: MLA_SM100_Decode_Config]() -> Int
Returns this CTA's head-tile index.
The decode grid flattens (query token, head tile) into x, head tile
fastest, so the head tiles of one token are adjacent CTAs and share that
token's gathered K/V in L2. Keeping the token index off gridDim.y
lifts CUDA's 65535 cap on it.
Parameters:
- config (
MLA_SM100_Decode_Config): Decode config supplyingnum_q_headsandBM.
Returns:
Int: The head-tile index in [0, ceildiv(num_q_heads, BM)).