fix H100 unit test error

Author: nv-yunzheq

csrc/fused_moe/cutlass_backend/cutlass_fused_moe_instantiation.cu

@@ -54,4 +54,12 @@ template class CutlassMoeFCRunner<__nv_fp8_e4m3, __nv_fp4_e2m1, __nv_bfloat16, _
 template class CutlassMoeFCRunner<__nv_bfloat16, __nv_fp4_e2m1>;
 #endif
 #endif
+
+// Explicit instantiations for finalizeMoeRoutingKernelLauncher to ensure
+// symbols are emitted in the JIT library for common data types.
+INSTANTIATE_FINALIZE_MOE_ROUTING(half, half, half);
+INSTANTIATE_FINALIZE_MOE_ROUTING(float, float, float);
+#ifdef ENABLE_BF16
+INSTANTIATE_FINALIZE_MOE_ROUTING(__nv_bfloat16, __nv_bfloat16, __nv_bfloat16);
+#endif
 }  // namespace tensorrt_llm::kernels::cutlass_kernels

csrc/fused_moe/cutlass_backend/cutlass_fused_moe_kernels.cuh

@@ -2923,8 +2923,9 @@ void CutlassMoeFCRunner<T, WeightType, OutputType, InputType, ScaleBiasType, Ena
   finalizeMoeRoutingKernelLauncher<OutputType, UnfusedGemmOutputType>(
       static_cast<UnfusedGemmOutputType const*>(gemm_output), final_output, fc2_expert_biases,
       unpermuted_final_scales, unpermuted_row_to_permuted_row, permuted_row_to_unpermuted_row,
-      token_selected_experts, expert_first_token_offset, num_rows, hidden_size, k,
-      num_experts_per_node, parallelism_config, enable_alltoall, enable_pdl, stream);
+      token_selected_experts, expert_first_token_offset, num_rows, hidden_size,
+      unpadded_hidden_size, k, num_experts_per_node, parallelism_config, enable_alltoall,
+      enable_pdl, stream);
 }
 
 template <class T, class WeightType, class OutputType, class InputType, class ScaleBiasType,

csrc/nv_internal/tensorrt_llm/kernels/cutlass_kernels/include/moe_util_kernels.h

@@ -72,10 +72,10 @@ void finalizeMoeRoutingKernelLauncher(
     GemmOutputType const* expanded_permuted_rows, OutputType* reduced_unpermuted_output,
     ScaleBiasType const* bias, float const* final_scales, int const* unpermuted_row_to_permuted_row,
     int const* permuted_row_to_unpermuted_row, int const* token_selected_experts,
-    int64_t const* expert_first_token_offset, int64_t const num_rows, int64_t const cols,
-    int64_t const experts_per_token, int64_t const num_experts_per_node,
-    MOEParallelismConfig parallelism_config, bool const enable_alltoall, bool enable_pdl,
-    cudaStream_t stream);
+    int64_t const* expert_first_token_offset, int64_t const num_rows, int64_t const padded_cols,
+    int64_t const unpadded_cols, int64_t const experts_per_token,
+    int64_t const num_experts_per_node, MOEParallelismConfig parallelism_config,
+    bool const enable_alltoall, bool enable_pdl, cudaStream_t stream);
 
 }  // namespace cutlass_kernels
 }  // namespace tensorrt_llm::kernels