update

microsoft · Feb 1, 2024 · b5cd733 · b5cd733
1 parent b7cb71b
commit b5cd733
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/onnxruntime/contrib_ops/rocm/bert/batched_gemm_softmax_gemm_permute_pipelines.cuh b/onnxruntime/contrib_ops/rocm/bert/batched_gemm_softmax_gemm_permute_pipelines.cuh
@@ -734,11 +734,11 @@ class GemmSoftmaxGemmPermuteTunableOp : public tunable::TunableOp<GemmSoftmaxGem
 
 template <typename U, typename V, typename T, bool USE_BIAS, bool USE_MASK>
 auto GetArgAndRunInvoker(const U& impl, const V& invoker, const GemmSoftmaxGemmPermuteParams<T>* params) {
+  constexpr const int kNumBiasBuffer = static_cast<int>(USE_BIAS) + static_cast<int>(USE_MASK);
+
   using Nop = ck::tensor_operation::element_wise::PassThrough;
   using Acc0ElementOp = internal::PreSoftmaxAttentionScoreOp;
 
-  constexpr const int kNumBiasBuffer = static_cast<int>(USE_BIAS) + static_cast<int>(USE_MASK);
-
   TUNABLE_OP_RETURN_UNSUPPORTED_ARGUMENT_IF(
       !GemmSoftmaxGemmPermuteTunableOp<T>::IsSupportedMode(params->attention),
       "attention mode is not supported, got ", params->attention->mode);