adjust nvcc_threads

Author: tianleiwu

.github/workflows/linux_cuda_ci.yml

@@ -29,7 +29,7 @@ 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_almalinux8_gcc14:20251017.1'
       docker_image_repo: onnxruntimecuda12manylinuxbuild
-      extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --cuda_version=12.8 --cuda_home=/usr/local/cuda-12.8 --cudnn_home=/usr/local/cuda-12.8 --enable_cuda_profiling --build_java --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=90 onnxruntime_BUILD_UNIT_TESTS=ON onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON'
+      extra_build_flags: '--use_binskim_compliant_compile_flags --build_wheel --parallel --nvcc_threads 1 --cuda_version=12.8 --cuda_home=/usr/local/cuda-12.8 --cudnn_home=/usr/local/cuda-12.8 --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

.github/workflows/windows_cuda.yml

@@ -115,7 +115,7 @@ jobs:
             exit $lastExitCode
           }
           # Execute the build process
-          python.exe ${{ github.workspace }}\tools\ci_build\build.py --update --build --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86  --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
+          python.exe ${{ github.workspace }}\tools\ci_build\build.py --update --build --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --nvcc_threads 1 --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86  --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
           if ($lastExitCode -ne 0) {
             exit $lastExitCode
           }
@@ -235,7 +235,7 @@ jobs:
             exit $lastExitCode
           }
 
-          python.exe ${{ github.workspace }}\tools\ci_build\build.py --test --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86  --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
+          python.exe ${{ github.workspace }}\tools\ci_build\build.py --test --config RelWithDebInfo --build_dir build --skip_submodule_sync --build_csharp --parallel --nvcc_threads 1 --use_binskim_compliant_compile_flags --cmake_generator "Visual Studio 17 2022" --build_shared_lib --build_wheel --build_java --use_cuda --cuda_home="$env:RUNNER_TEMP\v12.8" --enable_cuda_profiling --use_vcpkg --use_vcpkg_ms_internal_asset_cache --enable_transformers_tool_test --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86  --cmake_extra_defines onnxruntime_ENABLE_CUDA_EP_INTERNAL_TESTS=ON
           if ($lastExitCode -ne 0) {
             exit $lastExitCode
           }

.github/workflows/windows_tensorrt.yml

@@ -121,7 +121,7 @@ jobs:
             exit $lastExitCode
           }
           # Execute the build process
-          python ${{ github.workspace }}\tools\ci_build\build.py --config RelWithDebInfo --parallel --use_binskim_compliant_compile_flags             --build_dir build --skip_submodule_sync --build_shared_lib --build --update --cmake_generator "Visual Studio 17 2022" --build_wheel --enable_onnx_tests             --use_tensorrt --tensorrt_home="${{ runner.temp }}\TensorRT-10.14.1.48.Windows.win10.cuda-12.9"             --cuda_home="${{ runner.temp }}\v12.8" --use_vcpkg --use_vcpkg_ms_internal_asset_cache --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
+          python ${{ github.workspace }}\tools\ci_build\build.py --config RelWithDebInfo --parallel --nvcc_threads 1 --use_binskim_compliant_compile_flags --build_dir build --skip_submodule_sync --build_shared_lib --build --update --cmake_generator "Visual Studio 17 2022" --build_wheel --enable_onnx_tests --use_tensorrt --tensorrt_home="${{ runner.temp }}\TensorRT-10.14.1.48.Windows.win10.cuda-12.9" --cuda_home="${{ runner.temp }}\v12.8" --use_vcpkg --use_vcpkg_ms_internal_asset_cache --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
           if ($lastExitCode -ne 0) {
             exit $lastExitCode
           }
@@ -247,7 +247,7 @@ jobs:
             exit $lastExitCode
           }
 
-          python ${{ github.workspace }}\tools\ci_build\build.py --config RelWithDebInfo --parallel --use_binskim_compliant_compile_flags             --build_dir build --skip_submodule_sync --build_shared_lib --test --cmake_generator "Visual Studio 17 2022" --build_wheel --enable_onnx_tests             --use_tensorrt --tensorrt_home="${{ runner.temp }}\TensorRT-10.14.1.48.Windows.win10.cuda-12.9"             --cuda_home="${{ runner.temp }}\v12.8" --use_vcpkg --use_vcpkg_ms_internal_asset_cache --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
+          python ${{ github.workspace }}\tools\ci_build\build.py --config RelWithDebInfo --use_binskim_compliant_compile_flags --parallel --nvcc_threads 1 --build_dir build --skip_submodule_sync --build_shared_lib --test --cmake_generator "Visual Studio 17 2022" --build_wheel --enable_onnx_tests --use_tensorrt --tensorrt_home="${{ runner.temp }}\TensorRT-10.14.1.48.Windows.win10.cuda-12.9" --cuda_home="${{ runner.temp }}\v12.8" --use_vcpkg --use_vcpkg_ms_internal_asset_cache --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=86
           if ($lastExitCode -ne 0) {
             exit $lastExitCode
           }

tools/ci_build/build.py

@@ -217,16 +217,14 @@ def number_of_nvcc_threads(args):
 
         available_memory = psutil.virtual_memory().available
         if isinstance(available_memory, int) and available_memory > 0:
-            if available_memory > 60 * 1024 * 1024 * 1024:
+            if available_memory >= 64 * 1024 * 1024 * 1024:
                 # When available memory is large enough, chance of OOM is small.
-                nvcc_threads = 4
+                nvcc_threads = min(4, int(available_memory / (8 * 4 * 1024 * 1024 * 1024)))
             else:
-                # NVCC need a lot of memory to compile 8 flash attention cu files in Linux or 4 cutlass fmha cu files in Windows.
-                # Here we select number of threads to ensure each thread has enough memory (>= 4 GB). For example,
-                # Standard_NC4as_T4_v3 has 4 CPUs and 28 GB memory. When parallel=4 and nvcc_threads=2,
-                # total nvcc threads is 4 * 2, which is barely able to build in 28 GB memory so we will use nvcc_threads=1.
+                # NVCC need a lot of memory to compile 48 flash attention cu files.
+                # Here we select number of threads to ensure each thread has enough memory (>= 4 GB).
                 memory_per_thread = 4 * 1024 * 1024 * 1024
-                fmha_cu_files = 4 if is_windows() else 16
+                fmha_cu_files = 48
                 fmha_parallel_jobs = min(fmha_cu_files, number_of_parallel_jobs(args))
                 nvcc_threads = max(1, int(available_memory / (memory_per_thread * fmha_parallel_jobs)))
                 print(