[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
+16
View File
@@ -42,6 +42,18 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_moe_C, m) {
" Tensor? maybe_expert_map,"
" Tensor(a!)? scatter_idx=None) -> ()");
#ifdef USE_CUDA
// Fused globalize + align + sort for the DeepEP-v2 humming INDEXED decode
// path (one launch; num_recv read on device from psum[-1], cudagraph-safe).
m.def(
"fused_globalize_align_block_size(Tensor! topk_idx, Tensor psum,"
" int rank_expert_offset, int global_num_experts,"
" int local_num_experts, int block_size,"
" Tensor! sorted_ids, Tensor! expert_ids,"
" Tensor! num_tokens_post_pad,"
" Tensor(a!)? scatter_idx=None) -> ()");
#endif
// Aligning the number of tokens to be processed by each expert such
// that it is divisible by the block size, but for the batched case.
m.def(
@@ -143,6 +155,10 @@ STABLE_TORCH_LIBRARY_IMPL(_moe_C, CUDA, m) {
m.impl("topk_softplus_sqrt", TORCH_BOX(&topk_softplus_sqrt));
m.impl("moe_sum", TORCH_BOX(&moe_sum));
m.impl("moe_align_block_size", TORCH_BOX(&moe_align_block_size));
#ifdef USE_CUDA
m.impl("fused_globalize_align_block_size",
TORCH_BOX(&fused_globalize_align_block_size));
#endif
m.impl("batched_moe_align_block_size",
TORCH_BOX(&batched_moe_align_block_size));
m.impl("moe_lora_align_block_size", TORCH_BOX(&moe_lora_align_block_size));