Skip unselected experts in flashinfer_trtllm (#23493)

2026-04-23 17:30:19 -07:00
parent 000a2525e1
commit d9c72bdd2b
1 changed files with 1 additions and 2 deletions
@@ -333,8 +333,7 @@ def _pack_topk_for_flashinfer_routed(
    packed_ids = topk_ids.to(torch.int32)
    packed_weights = topk_weights.to(torch.bfloat16)
    packed = (packed_ids << 16) | packed_weights.view(torch.int16).to(torch.int32)
-    # SGLang can mark padded tokens with -1 expert ids.
-    return packed.masked_fill_(packed_ids < 0, 0)
+    return packed


 def fused_experts_none_to_flashinfer_trtllm_fp8(