[Perf][DSv4/DSv3.2] Add cluster-cooperative topK kernel for low-latency scenarios (#43008)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
This commit is contained in:
Roberto L. Castro
2026-06-23 16:11:00 -07:00
committed by GitHub
parent 3cc871aaf1
commit 855cd4d787
10 changed files with 1568 additions and 45 deletions
+8
View File
@@ -343,6 +343,14 @@ void persistent_topk(const torch::stable::Tensor& logits,
torch::stable::Tensor& workspace, int64_t k,
int64_t max_seq_len);
#ifdef VLLM_ENABLE_COOPERATIVE_TOPK
void cooperative_topk(const torch::stable::Tensor& logits,
const torch::stable::Tensor& lengths,
torch::stable::Tensor& output,
torch::stable::Tensor& workspace, int64_t k,
int64_t max_seq_len);
#endif
void selective_scan_fwd(
const torch::stable::Tensor& u, const torch::stable::Tensor& delta,
const torch::stable::Tensor& A, const torch::stable::Tensor& B,