You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Summary:
X-link: facebookresearch/FBGEMM#1481
Pull Request resolved: pytorch#4410
## Summary
This diff represents a second attempt to integrate a Manifold wrapper to facilitate reporter integration with the TBE forward operation, as referenced in D73927918. which failed a [test](https://www.internalfb.com/intern/test/281475120056413)where it expected all dependencies to be purely in Python.
- manifold_cli python library actually relies on C++ manifold cpp library, which breaks certain python packages/tests that cannot have a manfiold C++ dependency
- the workaround is to use manifold C++ and have all the dependency built (wrapped as Torch class) with fbgemm_gpu.
However, this causes issues that led to S532997.
- D76186007 updates `fb:utils` to use have manifold cpp dependency.
- many models/workflows use different packages for frontend (e.g., python) and backend (e.g., C++)
- this causes issue for frontend package that already include D76186007 but use backend package that does not include. D76186007
- fb:utils (in frontend package), expects this manifold cpp source code.
- backend packages does not contains this source code
- hence, it fails with
```libdeeplearning_fbgemm_fbgemm_gpu_fb_utils_cpp.so: cannot open shared object file: No such file or directory```
This diff ensures all the changes are built as separate targets, which are not included in any existing targets/usage.
We will carefully make changes to switch to use this targets to ensure backward compatibility in the next diff.
Reviewed By: spcyppt
Differential Revision: D77241003
fbshipit-source-id: c60a1ef75c4d7404e197d349f34debcedfee3d95
0 commit comments