PyTorch Status
Status 06 January 2023
Status
Open PRs
Status 21 November 2022
Status
Open PRs
(Became stale and closed)
(Became stale and closed)
Closed PRs
is no longer an issue and it seems to be possible to build OpenBLAS with full LAPACK support. Now we need to test PyTorch with the relevant feature sets turned on.
Unable to build with VS2022 but https://developercommunity.visualstudio.com/t/Incorrect-result-for-std::complex-operat/10088405 is solved.
Status 13 September 2022
Status
test_tensorexpr is fixed
test_ops is fixed on latest version.
Further testing requires a LAPACK enabled BLAS library. In order to compile full LAPACK functionality with OpenBLAS a bug needs to be fixed on clang. It has been reported and is under investigation . We have tried using the C_BLAS option while trying to compile OpenBLAS but this version does not seem to have all functionalities required by PyTorch.
All patches are still under evaluation by the PyTorch maintainers, some of them have been approved by reviewers but nothing else has happened.
Those issues are currently blocking further progress.
Status
Build failures (In upstream review)
Fix c++ unit test failures without additional modules (In upstream review)
Enable and test with FBGEMM module (In progress)
Enable and test with OpenMP module (To do)
Enable and test with QNNPACK module (To do)
Enable and test with XNNPACK module (To do)
Debug Python test failures (To do)
CI/CD (To do)
Arm64 optimizations for PyTorch (To Do)
Solved issues awaiting review and merge
Building related PRs:
Failing tests related PRs:
This is fixes both a failing test and a compilation issue.
Microsoft related issues:
https://developercommunity.visualstudio.com/t/Incorrect-result-for-std::complex-operat/10088405 - Microsoft claims its fixed but we will only be able to check after Preview 2 is released.
Issues under consideration
The following list takes into account our current building procedure which disables OpenMP, FBGEMM, QNNPACK and XNNPACK so some tests are failing due to this being disabled and are not going to be mentioned here.
Tests for now are only encompassing C++ tests because they are easier to debug and Python tests are failing at generally related issues.
test_tensorexpr - Currently under investigation this test expects a fixed read/write history to evaluate the result. However the DAG can have more than one valid topological order and on Windows the other one gets generated hence generating an error.
test_ops - It seems to be a numerical inaccuracy of some sort, further investigation required.