Move Linux CUDA pipelines to H100 (#25523)

Author: snnn

.github/workflows/linux_cuda_ci.yml

@@ -28,7 +28,7 @@ jobs:
       dockerfile_path: tools/ci_build/github/linux/docker/Dockerfile.manylinux2_28_cuda
       docker_build_args: '--build-arg BASEIMAGE=onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda12_x64_ubi8_gcc12:20250124.1'
       docker_image_repo: onnxruntimecuda12manylinuxbuild
-      extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --enable_cuda_profiling --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=80 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
+      extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --enable_cuda_profiling --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=90 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
       python_path_prefix: 'PATH=/opt/python/cp310-cp310/bin:$PATH'
       run_tests: false            # <<< Do not run tests in this job
       upload_build_output: true   # <<< Upload the build/Release directory
@@ -41,7 +41,7 @@ jobs:
     needs: build-linux-cuda-x64-release
     runs-on:
       - self-hosted
-      - "1ES.Pool=Onnxruntime-github-Linux-GPU-A100-WUS3"
+      - "1ES.Pool=Onnxruntime-github-Linux-GPU-H100"
     permissions:
       contents: read
       packages: read
@@ -98,5 +98,5 @@ jobs:
           build_config: Release
           mode: 'test' # Set mode to test
           execution_providers: 'cuda'
-          extra_build_flags: '--use_binskim_compliant_compile_flags --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --enable_cuda_profiling --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=80 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
+          extra_build_flags: '--use_binskim_compliant_compile_flags --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --enable_cuda_profiling --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=90 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
           python_path_prefix: 'PATH=/opt/python/cp310-cp310/bin:$PATH'

.github/workflows/linux_tensorrt_ci.yml

@@ -28,7 +28,7 @@ jobs:
       dockerfile_path: tools/ci_build/github/linux/docker/Dockerfile.manylinux2_28_cuda
       docker_build_args: '--build-arg BASEIMAGE=onnxruntimebuildcache.azurecr.io/internal/azureml/onnxruntime/build/cuda12_x64_ubi8_gcc12:20250124.1 --build-arg TRT_VERSION=10.9.0.34-1.cuda12.8 --network=host'
       docker_image_repo: onnxruntimetensorrt86gpubuild
-      extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --use_tensorrt --tensorrt_home /usr  --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=80 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
+      extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --use_tensorrt --tensorrt_home /usr  --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=90 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
       python_path_prefix: 'PATH=/opt/python/cp310-cp310/bin:$PATH'
       run_tests: false            # <<< Do not run tests in this job
       upload_build_output: true   # <<< Upload the build/Release directory
@@ -41,7 +41,7 @@ jobs:
     needs: build-linux-TensorRT-x64-release
     runs-on:
       - self-hosted
-      - "1ES.Pool=Onnxruntime-github-Linux-GPU-A100-WUS3"
+      - "1ES.Pool=Onnxruntime-github-Linux-GPU-H100"
     permissions:
       contents: read
       packages: read
@@ -100,5 +100,5 @@ jobs:
           build_config: Release
           mode: 'test' # Set mode to test
           execution_providers: 'cuda tensorrt'
-          extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --use_tensorrt --tensorrt_home /usr  --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=80 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
+          extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.2 --cuda_home=/usr/local/cuda-12.2 --cudnn_home=/usr/local/cuda-12.2 --use_tensorrt --tensorrt_home /usr  --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=90 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
           python_path_prefix: 'PATH=/opt/python/cp310-cp310/bin:$PATH'

onnxruntime/test/providers/cuda/nhwc/nhwc_cuda_helper.h

@@ -46,7 +46,7 @@
   } else if (std::is_same<T, double>::value) {      \
     MAKE_PROVIDERS_EPS_EXT(2e-4, pad_to_nc1d)       \
   } else {                                          \
-    MAKE_PROVIDERS_EPS_EXT(2e-3, pad_to_nc1d)       \
+    MAKE_PROVIDERS_EPS_EXT(4e-3, pad_to_nc1d)       \
   }
 
 #define MAKE_PROVIDERS_EPS_TYPE(T) \