skip cudagraph/DP padding in topk (#48979)

Signed-off-by: gnovack <novackgm@gmail.com>
This commit is contained in:
gnovack
2026-07-21 15:20:30 -07:00
committed by GitHub
parent 08e5067561
commit 85f638a2b8
10 changed files with 302 additions and 99 deletions
+3 -3
View File
@@ -8,19 +8,19 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_moe_C, m) {
m.def(
"topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
"token_expert_indices, Tensor gating_output, bool renormalize, Tensor? "
"bias) -> ()");
"bias, Tensor? is_padding) -> ()");
// Apply topk sigmoid to the gating outputs.
m.def(
"topk_sigmoid(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
"token_expert_indices, Tensor gating_output, bool renormalize, "
"Tensor? bias, float routed_scaling_factor) -> ()");
"Tensor? bias, float routed_scaling_factor, Tensor? is_padding) -> ()");
m.def(
"topk_softplus_sqrt(Tensor! topk_weights, Tensor! topk_indices, Tensor! "
"token_expert_indices, Tensor gating_output, bool renormalize, float "
"routed_scaling_factor, Tensor? "
"bias, Tensor? input_ids, Tensor? tid2eid) -> ()");
"bias, Tensor? input_ids, Tensor? tid2eid, Tensor? is_padding) -> ()");
// Calculate the result of moe by summing up the partial results
// from all selected experts. topk_ids/expert_map are optional and, when