[Kernel] support merge_attn_states CUDA kernel, 3x speedup (#16173)

Signed-off-by: DefTruth <qiustudent_r@163.com>
This commit is contained in:
DefTruth
2025-04-11 06:50:50 -06:00
committed by GitHub
parent 51baa9c333
commit e9528f6dc6
10 changed files with 519 additions and 4 deletions
+11
View File
@@ -138,6 +138,17 @@ def mla_decode_kvcache_cpu(
block_tables, seq_lens)
# merge attn states ops
def merge_attn_states(output: torch.Tensor,
prefix_output: torch.Tensor,
prefix_lse: torch.Tensor,
suffix_output: torch.Tensor,
suffix_lse: torch.Tensor,
output_lse: Optional[torch.Tensor] = None) -> None:
torch.ops._C.merge_attn_states(output, output_lse, prefix_output,
prefix_lse, suffix_output, suffix_lse)
# pos encoding ops
def rotary_embedding(
positions: torch.Tensor,