diff --git a/.pipelines/nuget_config/x64/packages.config b/.pipelines/nuget_config/x64/packages.config
index 5dfeb0836f..8eef0b5bac 100644
--- a/.pipelines/nuget_config/x64/packages.config
+++ b/.pipelines/nuget_config/x64/packages.config
@@ -1,6 +1,6 @@
-
+
diff --git a/.pipelines/nuget_config/x86/packages.config b/.pipelines/nuget_config/x86/packages.config
index f8e63133f4..81f97948f1 100644
--- a/.pipelines/nuget_config/x86/packages.config
+++ b/.pipelines/nuget_config/x86/packages.config
@@ -1,6 +1,6 @@
-
+
diff --git a/cmake/external/dml.cmake b/cmake/external/dml.cmake
index 4dec75deca..1ee3a06b41 100644
--- a/cmake/external/dml.cmake
+++ b/cmake/external/dml.cmake
@@ -41,7 +41,7 @@ if (NOT onnxruntime_USE_CUSTOM_DIRECTML)
set(NUGET_CONFIG ${PROJECT_SOURCE_DIR}/../NuGet.config)
set(PACKAGES_CONFIG ${PROJECT_SOURCE_DIR}/../packages.config)
get_filename_component(PACKAGES_DIR ${CMAKE_CURRENT_BINARY_DIR}/../packages ABSOLUTE)
- set(DML_PACKAGE_DIR ${PACKAGES_DIR}/Microsoft.AI.DirectML.1.11.0)
+ set(DML_PACKAGE_DIR ${PACKAGES_DIR}/Microsoft.AI.DirectML.1.12.0)
# Restore nuget packages, which will pull down the DirectML redist package.
add_custom_command(
diff --git a/docs/OperatorKernels.md b/docs/OperatorKernels.md
index c23209d4fe..8843582fde 100644
--- a/docs/OperatorKernels.md
+++ b/docs/OperatorKernels.md
@@ -1204,6 +1204,7 @@ Do not modify directly.*
|DequantizeLinear|*in* x:**T1**
*in* x_scale:**T2**
*in* x_zero_point:**T1**
*out* y:**T2**|1+|**T1** = tensor(float)
**T2** = tensor(uint8)|
|EmbedLayerNormalization|*in* input_ids:**T1**
*in* segment_ids:**T1**
*in* word_embedding:**T**
*in* position_embedding:**T**
*in* segment_embedding:**T**
*in* gamma:**T**
*in* beta:**T**
*in* mask:**T1**
*in* position_ids:**T1**
*out* output:**T**
*out* mask_index:**T1**
*out* embedding_sum:**T**|1+|**T** = tensor(float), tensor(float16)|
|FusedMatMul|*in* A:**T**
*in* B:**T**
*out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
+|FusedMatMulActivation|*in* A:**T**
*in* B:**T**
*out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
|Gelu|*in* X:**T**
*out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
|GroupNorm|*in* X:**T**
*in* gamma:**M**
*in* beta:**M**
*out* Y:**T**|1+|**M** = tensor(float), tensor(float16)
**T** = tensor(float), tensor(float16)|
|NhwcConv|*in* X:**T**
*in* W:**T**
*in* B:**T**
*out* Y:**T**|1+|**T** = tensor(float), tensor(float16)|
diff --git a/onnxruntime/core/optimizer/graph_transformer_utils.cc b/onnxruntime/core/optimizer/graph_transformer_utils.cc
index 73403f2c62..7196465e5d 100644
--- a/onnxruntime/core/optimizer/graph_transformer_utils.cc
+++ b/onnxruntime/core/optimizer/graph_transformer_utils.cc
@@ -182,6 +182,7 @@ InlinedVector> GenerateTransformers(
#ifndef DISABLE_CONTRIB_OPS
const InlinedHashSet cpu_ep = {onnxruntime::kCpuExecutionProvider};
#endif
+ const InlinedHashSet dml_ep = {onnxruntime::kDmlExecutionProvider};
switch (level) {
case TransformerLevel::Level1: {
// RewriteRule optimizations are the simplest (they generally remove unnecessary nodes and are cheap to run)
@@ -308,6 +309,7 @@ InlinedVector> GenerateTransformers(
transformers.emplace_back(std::make_unique(cpu_cuda_dml_rocm_eps));
transformers.emplace_back(std::make_unique(cpu_cuda_dml_rocm_eps));
+ transformers.emplace_back(std::make_unique(dml_ep));
// GeluApproximation has side effects which may change results. It needs to be manually enabled,
// or alternatively the model can be updated offline using a model conversion script
diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp
index 667298d93c..0d555ed0f5 100644
--- a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp
+++ b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorRegistration.cpp
@@ -934,6 +934,7 @@ constexpr static OperatorRegistrationInformation operatorRegistrationInformation
{REG_INFO_MS( 1, Gelu, typeNameListDefault, supportedTypeListFloat16to32, DmlGraphSupport::Supported)},
{REG_INFO_MS( 1, BiasGelu, typeNameListDefault, supportedTypeListFloat16to32, DmlGraphSupport::Supported)},
{REG_INFO_MS( 1, FusedMatMul, typeNameListDefault, supportedTypeListFloat16to32, DmlGraphSupport::Supported)},
+ {REG_INFO_MS( 1, FusedMatMulActivation, typeNameListDefault, supportedTypeListFloat16to32, DmlGraphSupport::Supported)},
{REG_INFO_MS( 1, QLinearSigmoid, typeNameListDefault, supportedTypeListQLinearSigmoid, DmlGraphSupport::Supported, requiredConstantCpuInputs(), std::nullopt, QueryQLinearSigmoid)},
{REG_INFO_MS( 1, Attention, typeNameListAttention, supportedTypeListAttention, DmlGraphSupport::Supported, requiredConstantCpuInputs(), std::nullopt, QueryAttention)},
diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorUtility.cpp b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorUtility.cpp
index f83ad5a3a4..d8290bbdae 100644
--- a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorUtility.cpp
+++ b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/OperatorUtility.cpp
@@ -364,19 +364,10 @@ namespace Dml
int onnxAxis = HandleNegativeAxis(kernelInfo.GetOptionalAttribute(AttrName::GraphFusedAxis, -1), onnxDimCount);
auto dmlAdjustedAxis = GetDmlAdjustedAxis(onnxAxis, onnxDimCount, kernelInfo.GetTensorShapeDescription().GetInputTensorDimensionCount(0));
-
- // If the axis is supported by Softmax, use this version instead since it's more likely to be supported by metacommands
- if (dmlAdjustedAxis == onnxDimCount - 1)
- {
- activation.desc.activationType = DML_OPERATOR_ACTIVATION_SOFTMAX;
- }
- else
- {
- activation.desc.activationType = DML_OPERATOR_ACTIVATION_SOFTMAX1;
- activation.dmlAxes.push_back(dmlAdjustedAxis);
- activation.desc.params.softmax1.Axes = activation.dmlAxes.data();
- activation.desc.params.softmax1.AxisCount = gsl::narrow_cast(activation.dmlAxes.size());
- }
+ activation.desc.activationType = DML_OPERATOR_ACTIVATION_SOFTMAX1;
+ activation.dmlAxes.push_back(dmlAdjustedAxis);
+ activation.desc.params.softmax1.Axes = activation.dmlAxes.data();
+ activation.desc.params.softmax1.AxisCount = gsl::narrow_cast(activation.dmlAxes.size());
}
else
{
diff --git a/packages.config b/packages.config
index 97d03b55b6..b2c918c414 100644
--- a/packages.config
+++ b/packages.config
@@ -1,6 +1,6 @@
-
+
diff --git a/tools/nuget/generate_nuspec_for_native_nuget.py b/tools/nuget/generate_nuspec_for_native_nuget.py
index f935b025f8..e6d4759769 100644
--- a/tools/nuget/generate_nuspec_for_native_nuget.py
+++ b/tools/nuget/generate_nuspec_for_native_nuget.py
@@ -190,7 +190,7 @@ def generate_repo_url(line_list, repo_url, commit_id):
def generate_dependencies(xml_text, package_name, version):
- dml_dependency = ''
+ dml_dependency = ''
if package_name == "Microsoft.AI.MachineLearning":
xml_text.append("")