Sync msft 25062026

.github/workflows/linux_cuda_plugin_ci.yml

@@ -141,3 +141,27 @@ jobs:
               cd /onnxruntime_src/onnxruntime/test/python/transformers
               python test_cuda_plugin_ep.py
             "
+
+      # --- Run the CUDA plugin EP C++ GoogleTest binary ---
+      # onnxruntime_provider_test is built into the artifact and links the plugin tests
+      # (gated by ORT_UNIT_TEST_HAS_CUDA_PLUGIN_EP). The user-stream + CUDA graph test
+      # registers the plugin .so via GetSharedLibraryFileName("onnxruntime_providers_cuda_plugin"),
+      # which returns the platform-specific filename without a directory component. Run from
+      # /build/Release/Release so that filename resolves to the plugin .so built there.
+      - name: Run CUDA Plugin EP C++ Tests
+        run: |
+          docker run --rm --gpus all \
+            -v ${{ github.workspace }}:/onnxruntime_src \
+            -v ${{ runner.temp }}/Release:/build/Release \
+            -e NVIDIA_VISIBLE_DEVICES=all \
+            ${{ steps.build_docker_image_step.outputs.full-image-name }} \
+            bash -c "
+              set -ex
+              export PATH=/opt/python/cp312-cp312/bin:\$PATH
+              # Make libcudart.so.13 (and the plugin's CUDA deps) findable; see note above.
+              export LD_LIBRARY_PATH=/build/Release/Release:/usr/local/cuda-13.0/lib64:\${LD_LIBRARY_PATH:-}
+
+              cd /build/Release/Release
+              ls -la onnxruntime_provider_test libonnxruntime_providers_cuda_plugin.so
+              ./onnxruntime_provider_test --gtest_filter='CudaPluginUserStreamGraphTest.*'
+            "

cmake/deps.txt

@@ -50,7 +50,7 @@ protoc_mac_universal;https://github.com/protocolbuffers/protobuf/releases/downlo
 psimd;https://github.com/Maratyszcza/psimd/archive/072586a71b55b7f8c584153d223e95687148a900.zip;1f5454b01f06f9656b77e4a5e2e31d7422487013
 pthreadpool;https://github.com/google/pthreadpool/archive/dcc9f28589066af0dbd4555579281230abbf74dd.zip;533a77943203ef15ca608bcd9dbe2c94da7451d2
 pybind11;https://github.com/pybind/pybind11/archive/refs/tags/v3.0.2.zip;a064e663b4d7a337ac291d1bef7337ef4e60a1ae
-pytorch_cpuinfo;https://github.com/pytorch/cpuinfo/archive/403d652dca4c1046e8145950b1c0997a9f748b57.zip;30b2a07fe4bae8574f89176e56274cacdd6d135b
+pytorch_cpuinfo;https://github.com/pytorch/cpuinfo/archive/4628dc060ce4e82345dc166bbac875609db4ff69.zip;e58d4b47c16a982111c897e669ae4f1821a393d7
 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

cmake/external/onnxruntime_external_deps.cmake

@@ -371,9 +371,7 @@ if (CPUINFO_SUPPORTED)
       PATCH_COMMAND
         ${Patch_EXECUTABLE} -p1 < ${PROJECT_SOURCE_DIR}/patches/cpuinfo/patch_cpuinfo_h_for_arm64ec.patch &&
         # https://github.com/pytorch/cpuinfo/pull/324
-        ${Patch_EXECUTABLE} -p1 < ${PROJECT_SOURCE_DIR}/patches/cpuinfo/patch_vcpkg_arm64ec_support.patch &&
-        # https://github.com/pytorch/cpuinfo/pull/348
-        ${Patch_EXECUTABLE} -p1 < ${PROJECT_SOURCE_DIR}/patches/cpuinfo/win_arm_fp16_detection_fallback.patch
+        ${Patch_EXECUTABLE} -p1 < ${PROJECT_SOURCE_DIR}/patches/cpuinfo/patch_vcpkg_arm64ec_support.patch
       FIND_PACKAGE_ARGS NAMES cpuinfo
     )
   elseif(CMAKE_SYSTEM_NAME STREQUAL "Linux")

cmake/onnxruntime_mlas.cmake

@@ -56,6 +56,7 @@ onnxruntime_add_static_library(onnxruntime_mlas
   ${MLAS_SRC_DIR}/sqnbitgemm_q8_block.h
   ${MLAS_SRC_DIR}/flashattn.cpp
   ${MLAS_SRC_DIR}/flashattn_qkv.cpp
+  ${MLAS_SRC_DIR}/flashattn_gqa.cpp
   ${MLAS_SRC_DIR}/qkv_quant.cpp
   ${MLAS_SRC_DIR}/cast.cpp
   ${MLAS_SRC_DIR}/layernorm.cpp

cmake/onnxruntime_providers_cuda_plugin.cmake

@@ -88,10 +88,11 @@ list(FILTER CUDA_PLUGIN_EP_CC_SRCS EXCLUDE REGEX ".*/tensor/sequence_op\\.cc$")
 # in the CPU provider and is not linked into the plugin.
 list(FILTER CUDA_PLUGIN_EP_CC_SRCS EXCLUDE REGEX ".*/tensor/size\\.cc$")
 
-# Permanently excluded — pure CPU ops, handled by GetCpuPreferredNodes.
-# shape_op.cc inherits from onnxruntime::OpKernel (framework)
-# which cannot convert to ep::adapter::OpKernel in the plugin build.
-list(FILTER CUDA_PLUGIN_EP_CC_SRCS EXCLUDE REGEX ".*/tensor/shape_op\\.cc$")
+# shape_op.cc is INCLUDED in the plugin build. It provides an adapter-based
+# Shape kernel under #ifdef BUILD_CUDA_EP_AS_PLUGIN (the CPU onnxruntime::Shape
+# class, which derives from the framework OpKernel, is only used in the
+# non-plugin build). Registering Shape on the EP keeps it off the CPU EP and
+# avoids Memcpy nodes that would otherwise break CUDA Graph capture.
 
 # Exclude contrib training ops (shrunken_gather depends on provider_api.h in header).
 list(FILTER CUDA_PLUGIN_EP_CC_SRCS EXCLUDE REGEX ".*/contrib_ops/cuda/tensor/shrunken_gather\\.cc$")

cmake/onnxruntime_unittests.cmake

@@ -1609,8 +1609,8 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
 
 endif()
 
-
-  if(onnxruntime_USE_QNN)
+  # Build ep_weight_sharing_ctx_gen for all supported EPs (QNN, TensorRT, OpenVINO, VitisAI)
+  if(onnxruntime_USE_QNN OR onnxruntime_USE_TENSORRT OR onnxruntime_USE_OPENVINO OR onnxruntime_USE_VITISAI)
     #qnn ctx generator
     set(ep_weight_sharing_ctx_gen_src_dir ${TEST_SRC_DIR}/ep_weight_sharing_ctx_gen)
     set(ep_weight_sharing_ctx_gen_src_patterns

cmake/patches/cpuinfo/fix_missing_sysfs_fallback.patch

@@ -1,10 +1,19 @@
 diff --git a/src/linux/processors.c b/src/linux/processors.c
-index 47bee76..d0c5569 100644
+index fd040a3..2ca8ec4 100644
 --- a/src/linux/processors.c
 +++ b/src/linux/processors.c
-@@ -2,0 +3 @@
+@@ -3,6 +3,7 @@
+ #include <stdio.h>
+ #include <stdlib.h>
+ #include <string.h>
 +#include <unistd.h>
-@@ -291,0 +293,22 @@
+
+ #if !defined(__ANDROID__)
+ /*
+@@ -289,6 +290,28 @@ static bool max_processor_number_parser(uint32_t processor_list_start, uint32_t
+ 	return true;
+ }
+
 +static uint32_t cpuinfo_linux_get_max_processor_from_sysconf(
 +	uint32_t max_processors_count,
 +	const char* processor_list_name) {
@@ -27,13 +36,31 @@ index 47bee76..d0c5569 100644
 +	return max_processor;
 +}
 +
-@@ -301 +324 @@
+ uint32_t cpuinfo_linux_get_max_possible_processor(uint32_t max_processors_count) {
+ 	uint32_t max_possible_processor = 0;
+ 	if (!cpuinfo_linux_parse_cpulist(
+@@ -298,7 +321,7 @@ uint32_t cpuinfo_linux_get_max_possible_processor(uint32_t max_processors_count)
+ #else
+ 		cpuinfo_log_warning("failed to parse the list of possible processors in %s", POSSIBLE_CPULIST_FILENAME);
+ #endif
 -		return UINT32_MAX;
 +		return cpuinfo_linux_get_max_processor_from_sysconf(max_processors_count, POSSIBLE_CPULIST_FILENAME);
-@@ -323 +346 @@
+ 	}
+ 	if (max_possible_processor >= max_processors_count) {
+ 		cpuinfo_log_warning(
+@@ -320,7 +343,7 @@ uint32_t cpuinfo_linux_get_max_present_processor(uint32_t max_processors_count)
+ #else
+ 		cpuinfo_log_warning("failed to parse the list of present processors in %s", PRESENT_CPULIST_FILENAME);
+ #endif
 -		return UINT32_MAX;
 +		return cpuinfo_linux_get_max_processor_from_sysconf(max_processors_count, PRESENT_CPULIST_FILENAME);
-@@ -357,0 +381,31 @@
+ 	}
+ 	if (max_present_processor >= max_processors_count) {
+ 		cpuinfo_log_warning(
+@@ -355,6 +378,37 @@ static bool detect_processor_parser(uint32_t processor_list_start, uint32_t proc
+ 	return true;
+ }
+
 +static bool cpuinfo_linux_detect_processors_from_sysconf(
 +	uint32_t max_processors_count,
 +	uint32_t* processor0_flags,
@@ -65,19 +92,34 @@ index 47bee76..d0c5569 100644
 +	return true;
 +}
 +
-@@ -373 +427,6 @@
+ bool cpuinfo_linux_detect_possible_processors(
+ 	uint32_t max_processors_count,
+ 	uint32_t* processor0_flags,
+@@ -370,7 +424,12 @@ bool cpuinfo_linux_detect_possible_processors(
+ 		return true;
+ 	} else {
+ 		cpuinfo_log_warning("failed to parse the list of possible processors in %s", POSSIBLE_CPULIST_FILENAME);
 -		return false;
 +		return cpuinfo_linux_detect_processors_from_sysconf(
 +			max_processors_count,
 +			processor0_flags,
 +			processor_struct_size,
 +			possible_flag,
 +			POSSIBLE_CPULIST_FILENAME);
-@@ -392 +451,6 @@
+ 	}
+ }
+
+@@ -389,7 +448,12 @@ bool cpuinfo_linux_detect_present_processors(
+ 		return true;
+ 	} else {
+ 		cpuinfo_log_warning("failed to parse the list of present processors in %s", PRESENT_CPULIST_FILENAME);
 -		return false;
 +		return cpuinfo_linux_detect_processors_from_sysconf(
 +			max_processors_count,
 +			processor0_flags,
 +			processor_struct_size,
 +			present_flag,
 +			PRESENT_CPULIST_FILENAME);
+ 	}
+ }
+

cmake/patches/cpuinfo/patch_vcpkg_arm64ec_support.patch

@@ -1,13 +1,13 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index aedc983..dab589e 100644
+index 072c987..e43d6ab 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
 @@ -72,6 +72,17 @@ IF(CMAKE_SYSTEM_NAME MATCHES "FreeBSD" AND CPUINFO_TARGET_PROCESSOR STREQUAL "am
  ENDIF()
  IF(IS_APPLE_OS AND CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64|arm64.*)$")
    SET(CPUINFO_TARGET_PROCESSOR "${CMAKE_OSX_ARCHITECTURES}")
 +ELSEIF(MSVC AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.10")
-+  # Use CMAKE_C_COMPILER_ARCHITECTURE_ID. MSVC values are documented as available since CMake 3.10.
++  # Use CMAKE_C_COMPILER_ARCHITECTURE_ID for non-VS generators (e.g. Ninja) with MSVC.
 +  IF(CMAKE_C_COMPILER_ARCHITECTURE_ID STREQUAL "X86")
 +    SET(CPUINFO_TARGET_PROCESSOR "x86")
 +  ELSEIF(CMAKE_C_COMPILER_ARCHITECTURE_ID STREQUAL "x64")

cmake/vcpkg-ports/cpuinfo/patch_vcpkg_arm64ec_support.patch

@@ -1,13 +1,13 @@
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index aedc983..dab589e 100644
+index 072c987..e43d6ab 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
 @@ -72,6 +72,17 @@ IF(CMAKE_SYSTEM_NAME MATCHES "FreeBSD" AND CPUINFO_TARGET_PROCESSOR STREQUAL "am
  ENDIF()
  IF(IS_APPLE_OS AND CMAKE_OSX_ARCHITECTURES MATCHES "^(x86_64|arm64.*)$")
    SET(CPUINFO_TARGET_PROCESSOR "${CMAKE_OSX_ARCHITECTURES}")
 +ELSEIF(MSVC AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.10")
-+  # Use CMAKE_C_COMPILER_ARCHITECTURE_ID. MSVC values are documented as available since CMake 3.10.
++  # Use CMAKE_C_COMPILER_ARCHITECTURE_ID for non-VS generators (e.g. Ninja) with MSVC.
 +  IF(CMAKE_C_COMPILER_ARCHITECTURE_ID STREQUAL "X86")
 +    SET(CPUINFO_TARGET_PROCESSOR "x86")
 +  ELSEIF(CMAKE_C_COMPILER_ARCHITECTURE_ID STREQUAL "x64")

cmake/vcpkg-ports/cpuinfo/portfile.cmake

@@ -6,13 +6,12 @@ endif()
 vcpkg_from_github(
     OUT_SOURCE_PATH SOURCE_PATH
     REPO pytorch/cpuinfo
-    REF 403d652dca4c1046e8145950b1c0997a9f748b57
-    SHA512 f7cd6dc44bd1120af610cae1337ed4c0f557ba78d2de9c73fed350fa3dfe9512643a1619ae55f5a540c6316a87d641856cca27297bb8766e48f39b7b7a59da1f
-    HEAD_REF master
+    REF 4628dc060ce4e82345dc166bbac875609db4ff69
+    SHA512 db7a93279f2f6daaf825fbd8552935d8ed671d276b65ad614e11f722b6a6848e663850d65180d33b554d67ef1a36aae842feb368699f90be8f21172a1af1924e
+    HEAD_REF main
     PATCHES
         patch_cpuinfo_h_for_arm64ec.patch
         patch_vcpkg_arm64ec_support.patch       # https://github.com/pytorch/cpuinfo/pull/324
-        win_arm_fp16_detection_fallback.patch   # https://github.com/pytorch/cpuinfo/pull/348
 )
 
 vcpkg_check_features(OUT_FEATURE_OPTIONS FEATURE_OPTIONS