Merge

.gitattributes

@@ -1,4 +1,4 @@
-# This sets the default behaviour, overriding core.autocrlf
+# This sets the default behavior, overriding core.autocrlf
 * text=auto
 
 # All source files should have unix line-endings in the repository,

.pipelines/nuget_config/x64/packages.config

@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="utf-8"?>
 <packages>
   <package id="python" version="3.9.7" targetFramework="native" />
-  <package id="Microsoft.AI.DirectML" version="1.14.1" targetFramework="native" />
+  <package id="Microsoft.AI.DirectML" version="1.15.0" targetFramework="native" />
   <package id="Microsoft.Windows.CppWinRT" version="2.0.201201.7" targetFramework="native" />
 </packages>

.pipelines/nuget_config/x86/packages.config

@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="utf-8"?>
 <packages>
   <package id="pythonx86" version="3.9.7" targetFramework="native" />
-  <package id="Microsoft.AI.DirectML" version="1.14.1" targetFramework="native" />
+  <package id="Microsoft.AI.DirectML" version="1.15.0" targetFramework="native" />
   <package id="Microsoft.Windows.CppWinRT" version="2.0.201201.7" targetFramework="native" />
 </packages>

ThirdPartyNotices.txt

@@ -4820,7 +4820,7 @@ SOFTWARE.
 
 ----------------------------------------------------------------------------
 
-This is the MIT/Expat Licence. For more information see:
+This is the MIT/Expat License. For more information see:
 
 1. http://www.opensource.org/licenses/mit-license.php
 

cgmanifests/generated/cgmanifest.json

@@ -116,7 +116,7 @@
       "component": {
         "type": "git",
         "git": {
-          "commitHash": "344117638c8ff7e239044fd0fa7085839fc03021",
+          "commitHash": "a6ad7fbbdc2e14fab82bb8a6d27760d700198cbf",
           "repositoryUrl": "https://github.com/google/benchmark.git"
         },
         "comments": "google_benchmark"
@@ -136,7 +136,7 @@
       "component": {
         "type": "git",
         "git": {
-          "commitHash": "530d5c8c84abd2a46f38583ee817743c9b3a42b4",
+          "commitHash": "e39786088138f2749d64e9e90e0f9902daa77c40",
           "repositoryUrl": "https://github.com/google/googletest.git"
         },
         "comments": "googletest"
@@ -256,7 +256,7 @@
       "component": {
         "type": "git",
         "git": {
-          "commitHash": "3e9dfa2866941655c56877882565e7577de6fc7b",
+          "commitHash": "941f45bcb51457884fa1afd6e24a67377d70f75c",
           "repositoryUrl": "https://github.com/pybind/pybind11.git"
         },
         "comments": "pybind11"
@@ -266,7 +266,7 @@
       "component": {
         "type": "git",
         "git": {
-          "commitHash": "959002f82d7962a473d8bf301845f2af720e0aa4",
+          "commitHash": "ca678952a9a8eaa6de112d154e8e104b22f9ab3f",
           "repositoryUrl": "https://github.com/pytorch/cpuinfo.git"
         },
         "comments": "pytorch_cpuinfo"
@@ -276,7 +276,7 @@
       "component": {
         "type": "git",
         "git": {
-          "commitHash": "2b354c6ad0d0479dcff68dab23fb0d1143a482c2",
+          "commitHash": "6dcd83d60f7944926bfd308cc13979fc53dd69ca",
           "repositoryUrl": "https://github.com/google/re2.git"
         },
         "comments": "re2"

cmake/deps.txt

@@ -26,9 +26,9 @@ eigen;https://gitlab.com/libeigen/eigen/-/archive/e7248b26a1ed53fa030c5c459f7ea0
 flatbuffers;https://github.com/google/flatbuffers/archive/refs/tags/v23.5.26.zip;59422c3b5e573dd192fead2834d25951f1c1670c
 fp16;https://github.com/Maratyszcza/FP16/archive/0a92994d729ff76a58f692d3028ca1b64b145d91.zip;b985f6985a05a1c03ff1bb71190f66d8f98a1494
 fxdiv;https://github.com/Maratyszcza/FXdiv/archive/63058eff77e11aa15bf531df5dd34395ec3017c8.zip;a5658f4036402dbca7cebee32be57fb8149811e1
-google_benchmark;https://github.com/google/benchmark/archive/refs/tags/v1.8.3.zip;bf9870756ee3f8d2d3b346b24ee3600a41c74d3d
+google_benchmark;https://github.com/google/benchmark/archive/refs/tags/v1.8.5.zip;cd47d3d272faf353600c8cc2fdec2b52d6f69177
 google_nsync;https://github.com/google/nsync/archive/refs/tags/1.26.0.zip;5e7c00ef6bf5b787386fc040067903ec774e2752
-googletest;https://github.com/google/googletest/archive/530d5c8c84abd2a46f38583ee817743c9b3a42b4.zip;5e3a61db2aa975cfd0f97ba92c818744e7fa7034
+googletest;https://github.com/google/googletest/archive/refs/tags/v1.15.0.zip;9d2d0af8d77ac726ea55d44a8fa727ec98311349
 googlexnnpack;https://github.com/google/XNNPACK/archive/0da379fc4808f9601faef392352018c741c0f297.zip;663883491e380b628e0a5b162b5f2658032fae73
 json;https://github.com/nlohmann/json/archive/refs/tags/v3.10.5.zip;f257f8dc27c5b8c085dc887b40cddd18ae1f725c
 microsoft_gsl;https://github.com/microsoft/GSL/archive/refs/tags/v4.0.0.zip;cf368104cd22a87b4dd0c80228919bb2df3e2a14
@@ -48,9 +48,9 @@ protoc_linux_aarch64;https://github.com/protocolbuffers/protobuf/releases/downlo
 protoc_mac_universal;https://github.com/protocolbuffers/protobuf/releases/download/v21.12/protoc-21.12-osx-universal_binary.zip;23710c3d1c2036d8d65a6a22234372fa2d7af9ef
 psimd;https://github.com/Maratyszcza/psimd/archive/072586a71b55b7f8c584153d223e95687148a900.zip;1f5454b01f06f9656b77e4a5e2e31d7422487013
 pthreadpool;https://github.com/Maratyszcza/pthreadpool/archive/4fe0e1e183925bf8cfa6aae24237e724a96479b8.zip;07a0aa91dd9bf86f31b95497e00f31d8a261a4bd
-pybind11;https://github.com/pybind/pybind11/archive/refs/tags/v2.12.0.zip;8482f57ed55c7b100672815a311d5450858723fb
-pytorch_cpuinfo;https://github.com/pytorch/cpuinfo/archive/959002f82d7962a473d8bf301845f2af720e0aa4.zip;85da3caa60eb2b148613b443fbc2bfdc30689965
-re2;https://github.com/google/re2/archive/refs/tags/2024-05-01.tar.gz;206cfee5ee0b4c6844680ba66275e9e8faa77405
+pybind11;https://github.com/pybind/pybind11/archive/refs/tags/v2.13.1.zip;9255d5c8568debcc329dd42ed8f410ee139ac7b1
+pytorch_cpuinfo;https://github.com/pytorch/cpuinfo/archive/ca678952a9a8eaa6de112d154e8e104b22f9ab3f.zip;138bf57d2a110935330d1048dce6d7b82d17d377
+re2;https://github.com/google/re2/archive/refs/tags/2024-07-02.zip;646e1728269cde7fcef990bf4a8e87b047882e88
 safeint;https://github.com/dcleblanc/SafeInt/archive/refs/tags/3.0.28.zip;23f252040ff6cb9f1fd18575b32fa8fb5928daac
 tensorboard;https://github.com/tensorflow/tensorboard/archive/373eb09e4c5d2b3cc2493f0949dc4be6b6a45e81.zip;67b833913605a4f3f499894ab11528a702c2b381
 cutlass;https://github.com/NVIDIA/cutlass/archive/refs/tags/v3.5.0.zip;ae038931b9fc2c416c17d9cda91d9706b343f56d

cmake/external/abseil-cpp.cmake

@@ -50,46 +50,96 @@ endif()
 # TODO: since multiple ORT's dependencies depend on Abseil, the list below would vary from version to version.
 # We'd better to not manually manage the list.
 set(ABSEIL_LIBS
-absl::city
+absl::absl_log
+absl::log_internal_log_impl
+absl::log_internal_strip
+absl::log_internal_message
+absl::log_internal_format
+absl::synchronization
+absl::str_format
 absl::flags
-absl::flat_hash_map
-absl::flat_hash_set
+absl::log_internal_globals
+absl::kernel_timeout_internal
+absl::str_format_internal
 absl::hash
-absl::inlined_vector
-absl::low_level_hash
-absl::node_hash_map
-absl::node_hash_set
+absl::log_internal_append_truncated
+absl::absl_vlog_is_on
+absl::flags_commandlineflag
+absl::time
+absl::symbolize
+absl::graphcycles_internal
+absl::log_internal_conditions
+absl::strings
+absl::malloc_internal
+absl::demangle_internal
 absl::optional
-absl::raw_hash_set
+absl::stacktrace
+absl::base
+absl::demangle_rust
+absl::bad_optional_access
+absl::strings_internal
+absl::debugging_internal
+absl::int128
+absl::spinlock_wait
+absl::decode_rust_punycode
 absl::raw_logging_internal
-absl::str_format
-absl::str_format_internal
+absl::flat_hash_set
+absl::flat_hash_map
+absl::node_hash_map
+absl::node_hash_set
+absl::compare
+absl::base_internal
+absl::nullability
+absl::bounded_utf8_length_sequence
+absl::log_severity
+absl::type_traits
+absl::atomic_hook
 absl::bits
-absl::fixed_array
+absl::flags_commandlineflag_internal
+absl::hash_container_defaults
 absl::numeric_representation
-absl::utility
-absl::type_traits
-absl::string_view
+absl::node_slot_policy
 absl::core_headers
-absl::nullability
+absl::dynamic_annotations
+absl::utf8_for_code_point
+absl::errno_saver
+absl::absl_check
+absl::hash_function_defaults
+absl::function_ref
+absl::city
+absl::low_level_hash
+absl::fixed_array
+absl::variant
+absl::meta
+absl::log_internal_voidify
+absl::log_sink
+absl::log_internal_log_sink_set
+absl::log_sink_registry
+absl::log_entry
+absl::log_globals
+absl::log_internal_nullguard
+absl::examine_stack
+absl::inlined_vector
+absl::log_internal_proto
+absl::strerror
+absl::log_internal_config
+absl::raw_hash_map
+absl::raw_hash_set
+absl::container_memory
+absl::algorithm_container
 absl::span
-absl::config
-absl::synchronization
-absl::base
+absl::log_internal_nullstream
+absl::vlog_config_internal
+absl::flags_reflection
+absl::flags_internal
+absl::flags_config
+absl::fast_type_id
+absl::utility
+absl::time_zone
 absl::civil_time
-absl::debugging_internal
-absl::demangle_internal
-absl::graphcycles_internal
-absl::int128
-absl::kernel_timeout_internal
-absl::log_severity
-absl::malloc_internal
-absl::spinlock_wait
-absl::stacktrace
 absl::string_view
-absl::strings
-absl::strings_internal
-absl::symbolize
 absl::throw_delegate
-absl::time
-absl::time_zone)
+absl::memory
+absl::charset
+absl::endian
+absl::config)

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.1.14.1)
+  set(DML_PACKAGE_DIR ${PACKAGES_DIR}/Microsoft.AI.DirectML.1.15.0)
 
   # Restore nuget packages, which will pull down the DirectML redist package.
   add_custom_command(

cmake/onnxruntime.cmake

@@ -150,7 +150,7 @@ endif()
 
 if(CMAKE_SYSTEM_NAME STREQUAL "Android" AND onnxruntime_MINIMAL_BUILD)
   # target onnxruntime is a shared library, the dummy __cxa_demangle is only attach to it to avoid
-  # affecting downstream ort library users with the behaviour of dummy __cxa_demangle. So the dummy
+  # affecting downstream ort library users with the behavior of dummy __cxa_demangle. So the dummy
   # __cxa_demangle must not expose to libonnxruntime_common.a. It works as when the linker is
   # creating the DSO, our dummy __cxa_demangle always comes before libc++abi.a so the
   # __cxa_demangle in libc++abi.a is discarded, thus, huge binary size reduction.

cmake/onnxruntime_python.cmake

@@ -97,8 +97,12 @@ endif()
 
 onnxruntime_add_include_to_target(onnxruntime_pybind11_state Python::Module Python::NumPy)
 target_include_directories(onnxruntime_pybind11_state PRIVATE ${ONNXRUNTIME_ROOT} ${pybind11_INCLUDE_DIRS})
-if(onnxruntime_USE_CUDA AND onnxruntime_CUDNN_HOME)
-    target_include_directories(onnxruntime_pybind11_state PRIVATE ${onnxruntime_CUDNN_HOME}/include)
+if(onnxruntime_USE_CUDA)
+    target_include_directories(onnxruntime_pybind11_state PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+    # cudnn_home is optional for Window when cuda and cudnn are installed in the same directory.
+    if(onnxruntime_CUDNN_HOME)
+      target_include_directories(onnxruntime_pybind11_state PRIVATE ${onnxruntime_CUDNN_HOME}/include)
+    endif()
 endif()
 if(onnxruntime_USE_CANN)
     target_include_directories(onnxruntime_pybind11_state PRIVATE ${onnxruntime_CANN_HOME}/include)

cmake/patches/composable_kernel/Fix_Clang_Build.patch

@@ -44,7 +44,7 @@ index c23746e7f..bc326c8b5 100644
  find_package(HIP REQUIRED)
  # Override HIP version in config.h, if necessary.
 @@ -269,12 +248,6 @@ if( DEFINED CK_OVERRIDE_HIP_VERSION_PATCH )
-     message(STATUS "CK_HIP_VERSION_PATCH overriden with ${CK_OVERRIDE_HIP_VERSION_PATCH}")
+     message(STATUS "CK_HIP_VERSION_PATCH overridden with ${CK_OVERRIDE_HIP_VERSION_PATCH}")
  endif()
  message(STATUS "Build with HIP ${HIP_VERSION}")
 -link_libraries(hip::device)

cmake/patches/cpuinfo/9bb12d342fd9479679d505d93a478a6f9cd50a47.patch

@@ -1,22 +1,22 @@
 diff --git a/include/cpuinfo.h b/include/cpuinfo.h
-index c46b65e..8b83a64 100644
+index 03f2776..eaf6497 100644
 --- a/include/cpuinfo.h
 +++ b/include/cpuinfo.h
 @@ -18,7 +18,7 @@
- 	#define CPUINFO_ARCH_X86 1
+ #define CPUINFO_ARCH_X86 1
  #endif
 
 -#if defined(__x86_64__) || defined(__x86_64) || defined(_M_X64) || defined(_M_AMD64)
-+#if defined(__x86_64__) || (defined(_M_X64) && !defined(_M_ARM64EC)) || (defined(_M_AMD64) && !defined(_M_ARM64EC))
- 	#define CPUINFO_ARCH_X86_64 1
++#if defined(__x86_64__) || defined(__x86_64) || (defined(_M_X64) && !defined(_M_ARM64EC)) || (defined(_M_AMD64) && !defined(_M_ARM64EC))
+ #define CPUINFO_ARCH_X86_64 1
  #endif
 
 @@ -26,7 +26,7 @@
- 	#define CPUINFO_ARCH_ARM 1
+ #define CPUINFO_ARCH_ARM 1
  #endif
 
 -#if defined(__aarch64__) || defined(_M_ARM64)
 +#if defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
- 	#define CPUINFO_ARCH_ARM64 1
+ #define CPUINFO_ARCH_ARM64 1
  #endif
 

csharp/ApiDocs/_exported_templates/default/partials/title.tmpl.partial

@@ -39,7 +39,7 @@ Event {{name.0.value}}
 Operator {{name.0.value}}
 {{/inOperator}}
 {{#inEii}}
-Explict Interface Implementation {{name.0.value}}
+Explicit Interface Implementation {{name.0.value}}
 {{/inEii}}
 {{#inVariable}}
 Variable {{name.0.value}}

dockerfiles/README.md

@@ -32,7 +32,7 @@
   docker run -it onnxruntime-source
   ```
 
-The docker file supports both x86_64 and ARM64(aarch64). You may use docker's "--platform" parameter to explictly specify which CPU architecture you want to build. For example:
+The docker file supports both x86_64 and ARM64(aarch64). You may use docker's "--platform" parameter to explicitly specify which CPU architecture you want to build. For example:
 
 ```bash
   docker build --platform linux/arm64/v8 -f Dockerfile.source
@@ -274,7 +274,7 @@ Note: You may add --use_tensorrt and --tensorrt_home options if you wish to use
 Note: Resulting Docker image will have ONNX Runtime installed in /usr, and ONNX Runtime wheel copied to /onnxruntime directory.
 Nothing else from ONNX Runtime source tree will be copied/installed to the image.
 
-Note: When running the container you built in Docker, please either use 'nvidia-docker' command instead of 'docker', or use Docker command-line options to make sure NVIDIA runtime will be used and appropiate files mounted from host. Otherwise, CUDA libraries won't be found. You can also [set NVIDIA runtime as default in Docker](https://github.com/dusty-nv/jetson-containers#docker-default-runtime).
+Note: When running the container you built in Docker, please either use 'nvidia-docker' command instead of 'docker', or use Docker command-line options to make sure NVIDIA runtime will be used and appropriate files mounted from host. Otherwise, CUDA libraries won't be found. You can also [set NVIDIA runtime as default in Docker](https://github.com/dusty-nv/jetson-containers#docker-default-runtime).
 
 ## MIGraphX
 **Ubuntu 20.04, ROCm6.0, MIGraphX**

docs/ORTModule_Training_Guidelines.md

@@ -304,6 +304,16 @@ A classical usage of disabling the deep copy: when the deep copy before module e
 	export ORTMODULE_ENABLE_MEM_EFFICIENT_GRAD_MGMT=0 # Disable
 	```
 
+#### ORTMODULE_ATEN_SDPA_FALLBACK
+
+- **Feature Area**: *ORTMODULE/Optimizations*
+- **Description**: By default, this is disabled. This env var can be used for enabling pre-export attention fall back to PyTorch's [_scaled_dot_product_efficient_attention](https://github.com/pytorch/pytorch/blob/c12a4f2e65ad41b739aab1a261e2336b4a79fcfb/aten/src/ATen/native/native_functions.yaml#L14778) ATen kernel for execution when calling torch.nn.functional.scaled_dot_product_attention. NOTE: only use this feature if user model leverages memory efficient attention WITHOUT masking (ie. attn_mask=None). Utilize GPU profiling looks like NVIDIA Nsight Systems to identify if user model leverages memory efficient attention.
+
+    ```bash
+    export ORTMODULE_ATEN_SDPA_FALLBACK=1 # ENABLE
+    unset ORTMODULE_ATEN_SDPA_FALLBACK # DISABLE
+    ```
+
 ### 2.2 Memory Optimization
 
 Q: *Want to run a bigger batch size?*