From c9b29fbd06aad25ba49f7e31c52a37f19b2f0977 Mon Sep 17 00:00:00 2001
From: KeDengMS <kedeng@microsoft.com>
Date: Thu, 25 Mar 2021 17:16:58 -0700
Subject: [PATCH] Disable MatmulTransposeFusion for CPU EP (#7135)

It causes convergence issue in BERT on CPU
---
 .../orttraining/core/optimizer/graph_transformer_utils.cc       | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/orttraining/orttraining/core/optimizer/graph_transformer_utils.cc b/orttraining/orttraining/core/optimizer/graph_transformer_utils.cc
index dd04537928..4b5353690c 100644
--- a/orttraining/orttraining/core/optimizer/graph_transformer_utils.cc
+++ b/orttraining/orttraining/core/optimizer/graph_transformer_utils.cc
@@ -171,7 +171,7 @@ std::vector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
       //transformers.emplace_back(onnxruntime::make_unique<ConstantFolding>(l1_execution_providers));
       transformers.emplace_back(onnxruntime::make_unique<MatMulAddFusion>(l1_execution_providers));
       transformers.emplace_back(onnxruntime::make_unique<FreeDimensionOverrideTransformer>(free_dimension_overrides));
-      transformers.emplace_back(onnxruntime::make_unique<MatmulTransposeFusion>(l1_execution_providers));
+      transformers.emplace_back(onnxruntime::make_unique<MatmulTransposeFusion>(cuda_rocm_execution_providers));
       transformers.emplace_back(onnxruntime::make_unique<BiasDropoutFusion>(cuda_rocm_execution_providers));
       transformers.emplace_back(onnxruntime::make_unique<BiasSoftmaxFusion>(l1_execution_providers));
       transformers.emplace_back(onnxruntime::make_unique<MatMulScaleFusion>(l1_execution_providers, weights_to_train));