From b5cd7336301a245e6c7f6ac8f647ac3512e30481 Mon Sep 17 00:00:00 2001
From: peixuanzuo <peixuanzuo@microsoft.com>
Date: Wed, 27 Dec 2023 08:39:32 +0000
Subject: [PATCH] update

---
 .../rocm/bert/batched_gemm_softmax_gemm_permute_pipelines.cuh | 4 ++--
 1 file changed, 2 insertions(+), 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
index 691463ebd2259..54dda4bfa6d2c 100644
--- 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);