Merge branch 'pytorch:main' into xu_switch_to_mkl_static

.github/scripts/validate_binaries.sh

@@ -69,6 +69,11 @@ else
         export PATH=${OLD_PATH}
     fi
 
+    # Use case CUDA_VISIBLE_DEVICES: https://github.com/pytorch/pytorch/issues/128819
+    if [[ ${MATRIX_GPU_ARCH_TYPE} == 'cuda' ]]; then
+        python3 -c "import torch;import os;print(torch.cuda.device_count(), torch.__version__);os.environ['CUDA_VISIBLE_DEVICES']='0';print(torch.empty(2, device='cuda'))"
+    fi
+
     # this is optional step
     if [[ ${TARGET_OS} != linux*  ]]; then
         conda deactivate

check_binary.sh

@@ -375,7 +375,7 @@ if [[ "$OSTYPE" == "msys" ]]; then
 fi
 
 # Test that CUDA builds are setup correctly
-if [[ "$DESIRED_CUDA" != 'cpu' && "$DESIRED_CUDA" != 'cpu-cxx11-abi' && "$DESIRED_CUDA" != *"rocm"* && "$(uname -m)" != "s390x" ]]; then
+if [[ "$DESIRED_CUDA" != 'cpu' && "$DESIRED_CUDA" != 'xpu' && "$DESIRED_CUDA" != 'cpu-cxx11-abi' && "$DESIRED_CUDA" != *"rocm"* && "$(uname -m)" != "s390x" ]]; then
   if [[ "$PACKAGE_TYPE" == 'libtorch' ]]; then
     build_and_run_example_cpp check-torch-cuda
   else

common/install_mkl.sh

@@ -2,7 +2,7 @@
 set -ex
 
 # MKL
-MKL_VERSION=2022.2.1
+MKL_VERSION=2024.2.0
 
 mkdir -p /opt/intel/
 pushd /tmp

manywheel/build_common.sh

@@ -437,19 +437,6 @@ for pkg in /$WHEELHOUSE_DIR/torch_no_python*.whl /$WHEELHOUSE_DIR/torch*linux*.w
         popd
     fi
 
-    if [[ "$USE_SPLIT_BUILD" == "true" ]]; then
-        # @sahanp todo: Remove this line
-        echo "current files in directory"
-        ls -l
-
-        echo "removing extraneous .so and .a files"
-        # todo @PaliC: Remove these .so and .a files before hand in the split build
-        rm *.so *.a || true
-
-        echo "Packaging the following files"
-        ls -alR
-    fi
-
     # zip up the wheel back
     zip -rq $(basename $pkg) $PREIX*
 

test/smoke_test/smoke_test.py

@@ -181,13 +181,6 @@ def smoke_test_cuda(package: str, runtime_error_check: str, torch_compile_check:
         if (sys.platform in ["linux", "linux2"]):
             print(f"torch nccl version: {torch.cuda.nccl.version()}")
 
-
-         # test settind CUDA_VISIBLE_DEVICES
-        print(f"torch.cuda.device_count: {torch.cuda.device_count()}")
-        os.environ['CUDA_VISIBLE_DEVICES']='0'
-        print(torch.empty(2, device='cuda'))
-        del os.environ['CUDA_VISIBLE_DEVICES']
-
         if runtime_error_check == "enabled":
             test_cuda_runtime_errors_captured()
 
@@ -323,6 +316,7 @@ def main() -> None:
     )
     options = parser.parse_args()
     print(f"torch: {torch.__version__}")
+    print(torch.__config__.parallel_info())
 
     check_version(options.package)
     smoke_test_conv2d()