Module attention 

Multi-head attention (cudnnFusedAttnFwd/cudnnFusedAttnBwd) request types.

Routes through the v9 frontend OPERATION_MATMUL_DESCRIPTOR + softmax + dropout fusion path. Supports causal masking, sliding window, paged-KV (skeleton), MQA / GQA via head-count split.

Structs

AttentionParams

Attention parameters.

MultiHeadAttnBwdRequest

MHA backward request.

MultiHeadAttnFwdRequest

MHA forward request.

Enums

AttentionMask

Mask kind applied to the attention scores.

Functions

build_mha_bwd_graph

build_mha_fwd_graph