[Kernel] Fuse DeepEP-v2 decode globalize+align into one launch (#53911)

Signed-off-by: Elvir Crncevic <elvircrn@gmail.com>
This commit is contained in:
Elvir Crnčević
2026-10-13 11:10:38 -04:00
committed by GitHub
parent 669f85e686
commit b9b401b2ef
10 changed files with 686 additions and 16 deletions
+26
View File
@@ -2195,6 +2195,32 @@ def moe_align_block_size(
)
def fused_globalize_align_block_size(
topk_idx: torch.Tensor,
psum: torch.Tensor,
rank_expert_offset: int,
global_num_experts: int,
local_num_experts: int,
block_size: int,
sorted_ids: torch.Tensor,
expert_ids: torch.Tensor,
num_tokens_post_pad: torch.Tensor,
scatter_idx: torch.Tensor | None = None,
) -> None:
torch.ops._moe_C.fused_globalize_align_block_size(
topk_idx,
psum,
rank_expert_offset,
global_num_experts,
local_num_experts,
block_size,
sorted_ids,
expert_ids,
num_tokens_post_pad,
scatter_idx,
)
def batched_moe_align_block_size(
max_tokens_per_batch: int,
block_size: int,