[fix] Add support for flashinfer MOE A2A to Qwen3 BF16 model path (#26255)

sgl-kernel/csrc/moe/moe_topk_softmax_kernels.cu

@@ -753,6 +753,14 @@ void topk_softmax(
   const int num_tokens = static_cast<int>(gating_output.size(0));
   const int topk = static_cast<int>(topk_weights.size(-1));
 
+  // No tokens on this DP rank, no need to do anything
+  if (num_tokens == 0) {
+    return;
+  }
+
+  TORCH_CHECK(num_experts > 0, "num_experts must be greater than 0");
+  TORCH_CHECK(topk > 0, "topk must be greater than 0");
+
   const bool is_pow_2 = (num_experts != 0) && ((num_experts & (num_experts - 1)) == 0);
   const bool needs_workspace = !is_pow_2 || num_experts > 512;
   const int64_t workspace_size = needs_workspace ? num_tokens * num_experts : 0;
@@ -822,4 +830,7 @@ void topk_softmax(
   } else {
     TORCH_CHECK(false, "Unsupported gating_output dtype: ", dtype);
   }
+
+  auto launch_error = cudaGetLastError();
+  TORCH_CHECK(launch_error == cudaSuccess, "topk_softmax launch error: ", cudaGetErrorString(launch_error));
 }