PyTorch 1.8.0 Dev Release Notes

release/packaging
1

PyTorch 1.8 release contains quite a few commits that are not user facing but are interesting to people compiling from source or developing low level extensions for PyTorch.
Here is a non-exhaustive list of the most important ones.

Python API

  • Remove PyCFunction casts as much as possible. (#46227)
  • Clean up use of Flake8 in GitHub CI (#46740)
  • Refactors test_torch.py to be fewer than 10k lines (#47356)
  • Refactored OpInfo testing to support custom SampleInputs, added addmm to op_db to test (#48627)
  • Review memory overlap checks for advanced indexing operations (#48651)
  • Remove unused six code for Python 2/3 compatibility (#48077)
  • Creation of test framework for Sparse Operators (#48488)
  • Internal gradcheck wrapper in testing._internal that changes the default value of some flags (#51133)
  • Fix assertEqual's handling of numpy array inputs (#48217)

Build

  • Switch PyTorch Selective Build (Custom Build) to use the SelectiveBuilder abstraction (#45722)
  • Set USE_DISTRIBUTED OFF when libuv is not installed (#45554)
  • Compress NVCC flags for Windows (#45842)
  • Conditional requirement for py3.6 only (#46932)
  • Improve libuv detection on Windows (#48571)
  • Fix Python.h discovery logic on some MacOS platforms (#51586)
  • Bring fast_nvcc.py to PyTorch OSS to speed up CUDA compilation (#48934)
  • Expose CXX_FLAGS at runtime via torch.__config__._cxx_flags()(#47861)

Distributed

  • Let RpcAgent::send() return JitFuture (#49906)
  • Use correct signatures for METH_NOARGS (#45528)
  • Move python-independent c10d implementations to torch/lib (#47309)
  • Completely remove FutureMessage type and its usage (#50029)
  • Use store based barrier in init_process_group (#49419)

CUDA

  • CUDA BFloat16 infrastructure (#44925)
  • [CUDA graphs] Cuda RNG-safe graph capture and replay bindings (#48875)
  • Move CUDA kernel check to c10 (#48277, #48615)
  • More about cudnn refactor (#50827)
  • Skip cuda test_cholesky_solve_batched_many_batches due to illegal memory access (#48999)
  • Bump up the CUDA OOM test memory size (#48029)
  • Refactor cudnn convolution (#49109)
  • Refactor CuFFTConfig to not use tensor objects (#46909)
  • Remove DataPtr extractor from CUDAFuture (#48840)
  • Allow ROCm CI to use non-default stream. (#48424)
  • Expand the test of torch.bmm on CUDA (#47124)
  • Use the latest philox_cuda_state API for stochastic rounding (#51004)

Dispatcher

  • Make duplicate def() calls an error in the dispatcher (#48098)
  • Support string argument defaults in native_functions.yaml (#45665)
  • Add alias dispatch key DefaultBackend (#45718)
  • Support DefaultBackend keyword in native_functions.yaml (#45719)
  • Rename legacy_dispatcher to native (#45974)
  • Refactor dispatcher and native to use Signature structure (#45990)
  • Update VariableTypeManual.cpp to not use catchAllKernel (#46353)
  • Remove catchAllKernel (#46354)
  • Delete TypeDefault call code generation logic in VariableType (#47000)
  • Faithful out arguments (#47712)
  • Add autograd data model to codegen (#48249)
  • Mostly removing unnecessary includes so that TensorIterator.h can be
    included from NativeFunctions.h without causing cycles (#48728)
  • Create a new key for threadLocalDebugInfo (#48762)
  • Unregister backward and requires_grad ops from Autograd backend key (#49613)
  • Move generator state APIs to ATen (#49589)
  • Split out RegisterDispatchKey to its own file (#51508)

Misc

  • [vmap] OpInfo-based tests now test that batched gradient computation with our vmap prototype works. (#50818)
  • [vmap] NewModuleTest and CriterionTest now test that batched gradient computation with our vmap prototype works. (#50739, #50740, #50744)
  • [TorchScript] Update JIT triage project board workflow (#45613)
  • [TorchScript] Reformat ivalue_inl.h and ivalue.h (#46174)
  • [Complex] Add hashing logic for c10::complex (#51441)
  • [Complex] Only run complex gradcheck in TestOpInfo based tests when complex is supported (#49018)
  • [Quantization] Add quantization triage bot script (#45622)
  • [ONNX] Fix onnx test-reports path in CI (#47315)
  • [Metal] Add MacOS unit tests for Metal ops (#50312)
1 Like