[Perf][DeepSeek] Add sigmoid+bias fusion to fused_grouped_topk from TRTLLM (#28124)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
Michael Goin
2025-11-07 18:20:55 -08:00
committed by GitHub
co-authored by Robert Shaw Wentao Ye
parent 61d25dc44b
commit 0852527647
5 changed files with 149 additions and 75 deletions
+3 -2
View File
@@ -107,9 +107,10 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
// Apply grouped topk routing to select experts.
m.def(
"grouped_topk(Tensor scores, Tensor scores_with_bias, int n_group, int "
"grouped_topk(Tensor scores, int n_group, int "
"topk_group, int topk, bool renormalize, float "
"routed_scaling_factor) -> (Tensor, Tensor)");
"routed_scaling_factor, Tensor bias, int scoring_func) -> (Tensor, "
"Tensor)");
m.impl("grouped_topk", torch::kCUDA, &grouped_topk);
#endif
}