This issue tracks the components / tests that are not working on Windows:

• test/test_distributed.py: currently disabled on Windows because Windows doesn't have fcntl and we need to look for substitutes
• torch/csrc/jit/fusion_compiler.cpp: Fuser is disabled on Windows because the current implementation uses symbols from Linux-specific headers dlfcn.h and unistd.h. We will need to find alternatives for Windows.
• test/test_torch.py: some parts of test_serialization and test_from_file are disabled because Windows doesn't support opening an already opened file (see discussion at https://github.com/pytorch/pytorch/pull/4065/files/4a7dfbd12c77850faa4171235b18b9b1540f3f0d#diff-9996665f82f52030836eb8657057cfadR4227 and https://github.com/pytorch/pytorch/pull/4065/files/4a7dfbd12c77850faa4171235b18b9b1540f3f0d#diff-9996665f82f52030836eb8657057cfadR4398)
• test/test_nccl: nccl currently doesn't work with Windows (Windows + Visual Studio port NVIDIA/nccl#31 is the porting diff)
• test/test_utils.py: test_cuda in TestBottleneck causes intermittent CUDA out-of-memory error on Windows
• test/test_utils.py, test/test_dataloader.py: DataLoader with multiple workers causes intermittent CUDA out-of-memory error on Windows.
• test/test_cpp_extensions.py: this test currently doesn't work on Windows
• test/test_dataloader.py, test/test_multiprocessing.py: various tests are disabled because there are issues when using multiprocessing on Windows (done in Fix multiprocessing and dataloader tests on Windows #4453)
• test/test_autograd.py: test_profiler is disabled because Windows doesn't have c++filt (see discussion at https://github.com/pytorch/pytorch/pull/4065/files/4a7dfbd12c77850faa4171235b18b9b1540f3f0d#diff-e7525582ab432c30821289d64cc52bc3R1653)
• Implement MM fusion (MM with add reduction tree) #4615 causes mysterious failures ("RuntimeError: Caught an unknown exception!") in 3 JIT tests
  • test_arg_configurations
  • test_multiuse_fn
  • test_output_unflatten
• test/test_torch.py, test/test_nn.py: CUDA HalfTensor currently doesn't work on Windows
• test/test_nn.py: test_Conv2d_deterministic_cudnn seems to be stuck on Windows
• test/test_cuda.py:
  • Some tests are turned off because CUDA HalfTensor currently doesn't work on Windows. Test logs to look at:
    https://ci.pytorch.org/jenkins/job/pytorch-builds/job/pytorch-win-ws2016-cuda9-cudnn7-py3-test/57/console
  • test_FloatTensor_qr_big and test_DoubleTensor_qr_big are known to stuck or have numerical errors intermittently

For more discussions, also see: #2941

cc @mruberry @peterjc123 @nbcsm @guyang3532