diff --git a/qwen3_6_scripts/moe_topk_softmax_kernels.cuh b/qwen3_6_scripts/moe_topk_softmax_kernels.cuh
index 578800f7..785b1041 100644
--- a/qwen3_6_scripts/moe_topk_softmax_kernels.cuh
+++ b/qwen3_6_scripts/moe_topk_softmax_kernels.cuh
@@ -704,9 +704,7 @@ void topk_gating_softmax_kernel_launcher(const T* gating_output,
       LAUNCH_SOFTMAX(T, 256, WARPS_PER_TB);
       break;
     default: {
-      TORCH_CHECK(softmax_workspace != nullptr)
-          << "softmax_workspace must be provided for num_experts that are "
-             "not a power of 2.";
+      TORCH_CHECK(softmax_workspace != nullptr, "softmax_workspace must be provided for num_experts that are not a power of 2.");
       static constexpr int TPB = 256;
       moe_softmax<T, TPB><<<num_tokens, TPB, 0, stream>>>(gating_output,
                                                           nullptr,
@@ -753,27 +751,18 @@ void topk_softmax(torch::Tensor& topk_weights,   // [num_tokens, topk]
   // Check data type
   TORCH_CHECK(gating_output.scalar_type() == at::ScalarType::Float ||
         gating_output.scalar_type() == at::ScalarType::Half ||
-        gating_output.scalar_type() == at::ScalarType::BFloat16)
-      << "gating_output must be float32, float16, or bfloat16";
+        gating_output.scalar_type() == at::ScalarType::BFloat16,
+        "gating_output must be float32, float16, or bfloat16");
 
   // Check dimensions
-  TORCH_CHECK(gating_output.dim() == 2)
-      << "gating_output must be 2D tensor [num_tokens, num_experts]";
-  TORCH_CHECK(topk_weights.dim() == 2)
-      << "topk_weights must be 2D tensor [num_tokens, topk]";
-  TORCH_CHECK(topk_indices.dim() == 2)
-      << "topk_indices must be 2D tensor [num_tokens, topk]";
+  TORCH_CHECK(gating_output.dim() == 2, "gating_output must be 2D tensor [num_tokens, num_experts]");
+  TORCH_CHECK(topk_weights.dim() == 2, "topk_weights must be 2D tensor [num_tokens, topk]");
+  TORCH_CHECK(topk_indices.dim() == 2, "topk_indices must be 2D tensor [num_tokens, topk]");
 
   // Check shapes
-  TORCH_CHECK(gating_output.size(0) == topk_weights.size(0))
-      << "First dimension of topk_weights must match num_tokens in "
-         "gating_output"
-      << "First dimension of topk_indices must match num_tokens in "
-         "gating_output";
+  TORCH_CHECK(gating_output.size(0) == topk_weights.size(0), "First dimension of topk_weights must match num_tokens in gating_output First dimension of topk_indices must match num_tokens in gating_output");
 
-  TORCH_CHECK(topk_weights.size(-1) == topk_indices.size(-1))
-      << "Second dimension of topk_indices must match topk in topk_weights"
-      << "topk must be less than or equal to num_experts";
+  TORCH_CHECK(topk_weights.size(-1) == topk_indices.size(-1), "Second dimension of topk_indices must match topk in topk_weights topk must be less than or equal to num_experts");
 
   const int num_experts = static_cast<int>(gating_output.size(-1));
   const int num_tokens = static_cast<int>(gating_output.size(0));
@@ -795,12 +784,9 @@ void topk_softmax(torch::Tensor& topk_weights,   // [num_tokens, topk]
   const float* bias_ptr = nullptr;
   if (correction_bias.has_value()) {
     const torch::Tensor& bias_tensor = correction_bias.value();
-    TORCH_CHECK(bias_tensor.dim() == 1)
-        << "correction_bias must be 1D tensor [num_experts]";
-    TORCH_CHECK(bias_tensor.size(0) == num_experts)
-        << "correction_bias size must match num_experts";
-    TORCH_CHECK(bias_tensor.scalar_type() == at::ScalarType::Float)
-        << "correction_bias must be float32, got " << bias_tensor.scalar_type();
+    TORCH_CHECK(bias_tensor.dim() == 1, "correction_bias must be 1D tensor [num_experts]");
+    TORCH_CHECK(bias_tensor.size(0) == num_experts, "correction_bias size must match num_experts");
+    TORCH_CHECK(bias_tensor.scalar_type() == at::ScalarType::Float, "correction_bias must be float32");
     bias_ptr = bias_tensor.data_ptr<float>();
   }