1.7.1 tracking

Issue	Status
Fix documentation to point to torch.overrides instead of _overrides	Complete
Core dump when checking that basic CNN works (Python 3.9) high priority module: autograd module: crash module: pybind triaged
Tensor-expression fuser bugfixes for 1.7.1
[v1.7.1] Various setup.py fixes
[v1.7.1] Enable Python 3.9 for Windows builds
[v1.7.1] Add Python 3.9 support (linux / macOS)
Add max supported SM for nvrtc-11.0
[torch][te] aten::type_as is unary, not binary
[pytorch][te] Handle negative axis in chunk
Constructing a ParameterDict raises a warning high priority module: nn
Incorrect info about overriding torch tensors in version 1.7.0 high priority
torch.arange numerics are different after 1.7 update on CPU high priority
[complex] torch.{sqrt, abs}: does not match numpy high priority
torch/utils/collect_env.py no longer works if pytorch is not installed
libtorch 1.5.0 libiomp5.dylib contains erroneous link to /DLC/torch/libiomp5.dylib instead of using @rpath
max_pool1d crashes (segfault)
Fix max_pool1d on discontiguous tensor
torch.version.debug returns True for release builds
Fix mul cuda for bool
RuntimeError: "mul_cuda" not implemented for 'Bool
Update pybind to 2.6.0
Fix torch.version.debug generation
[quant] Quantized AdaptivePool3d is much slower for ChannelsLast3d
Incorrect output loss value under specific CUDA version high priority
Error out when parameters() is called on replicated models
Pytorch 1.5.0 (installed from conda) errors with complaints about incompatibility between MKL and libgomp when using Pytorch's multiprocessing has workaround
PyTorch 1.7.0 CUDA driver warning
ProcessGroupNCCL NCCL lib version mismatch
Fix incorrect signatures in get_testing_overrides for 1.7 release
Fix output type of torch.max for Tensor subclasses