[NVIDIA] Support Cutlass MLA for Blackwell GPUs (#16032)

Signed-off-by: kaixih <kaixih@nvidia.com>
This commit is contained in:
Kaixi Hou
2025-04-27 06:29:21 -07:00
committed by GitHub
parent 756848e79e
commit ed7a29d9f8
8 changed files with 403 additions and 5 deletions
+9
View File
@@ -1525,3 +1525,12 @@ def flash_mla_with_kvcache(
num_splits,
)
return out, softmax_lse
def cutlass_mla_decode(out: torch.Tensor, q_nope: torch.Tensor,
q_pe: torch.Tensor, kv_c_and_k_pe_cache: torch.Tensor,
seq_lens: torch.Tensor, page_table: torch.Tensor,
scale: float) -> torch.Tensor:
torch.ops._C.cutlass_mla_decode(out, q_nope, q_pe, kv_c_and_k_pe_cache,
seq_lens, page_table, scale)
return out