[BE] use relative backwards references in torch.optim._multi_tensor #129132
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/129132
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit c86c6f5 with merge base 7178b4e ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
This pull request was exported from Phabricator. Differential Revision: D58815471 |
|
This pull request was exported from Phabricator. Differential Revision: D58815471 |
fbgheith
force-pushed
the
export-D58815471
branch
from
56ef5b4
to
cf9717a
Compare
…ytorch#129132) Summary: Pull Request resolved: pytorch#129132 This avoids some internal test failures caused by "infinite" recursion during import. Example: P1430048846 A bit of history: - PR pytorch#127703 introduced a circular dependency `torch/optim/__init__.py` imports `torch.optim._multi_tensor` and `torch.optim._multi_tensor/_init__.py` imports `torch.optim` This seemed to work fine (green signals everywhere) but caused some internal test failures after it landed; an infinite recursion during import. - PR pytorch#128875 attempted to fix this by removing the import from `torch/optim/__init__.py`. This seemed to work fine: green signals everywhere and the failing tests started passing but a smaller number of tests started failing; unable to import `torch.optim._multi_tensor` - This diff re-introduces the import but avoids the infinite recursion by using relative package paths Test Plan: CI signals Previously failing tests are passing: * https://www.internalfb.com/intern/testinfra/testrun/4785074840480058 * https://www.internalfb.com/intern/testinfra/testrun/281475349103659 * https://www.internalfb.com/intern/testinfra/testrun/11540474083575936 Differential Revision: D58815471
|
This pull request was exported from Phabricator. Differential Revision: D58815471 |
fbgheith
force-pushed
the
export-D58815471
branch
from
cf9717a
to
c86c6f5
Compare
| del _Adam | ||
| del _AdamW | ||
| del _NAdam | ||
| del _SGD | ||
| del _RAdam | ||
| del _RMSprop | ||
| del _Rprop | ||
| del _ASGD | ||
| del _Adamax | ||
| del _Adadelta | ||
| del _Adagrad |
There was a problem hiding this comment.
It's not strictly required. Just didn't want to corrupt the namespace.
|
|
||
|
|
||
| import torch.optim._multi_tensor |
There was a problem hiding this comment.
This was removed in #128875
What is the reason for adding it back?
There was a problem hiding this comment.
Removing it caused some test failures. The specific order in which imports were done caused some tests to not be able to resolve "torch.optim._multi_tensor".
| from .. import ( | ||
| Adadelta as _Adadelta, | ||
| Adagrad as _Adagrad, | ||
| Adam as _Adam, | ||
| Adamax as _Adamax, | ||
| AdamW as _AdamW, | ||
| ASGD as _ASGD, | ||
| NAdam as _NAdam, | ||
| RAdam as _RAdam, | ||
| RMSprop as _RMSprop, | ||
| Rprop as _Rprop, | ||
| SGD as _SGD, | ||
| ) |
There was a problem hiding this comment.
| from .. import ( | |
| Adadelta as _Adadelta, | |
| Adagrad as _Adagrad, | |
| Adam as _Adam, | |
| Adamax as _Adamax, | |
| AdamW as _AdamW, | |
| ASGD as _ASGD, | |
| NAdam as _NAdam, | |
| RAdam as _RAdam, | |
| RMSprop as _RMSprop, | |
| Rprop as _Rprop, | |
| SGD as _SGD, | |
| ) | |
| import torch |
Use torch.optim.XXX below.
There was a problem hiding this comment.
Circular imports with absolute paths are causing failures in our environment (infinite recursion). Using relative paths fixes that.
Summary:
This avoids some internal test failures caused by "infinite" recursion during import. Example: P1430048846
A bit of history:
PR [BE][Easy] export explicitly imported public submodules #127703 introduced a circular dependency
torch/optim/__init__.pyimportstorch.optim._multi_tensorandtorch.optim._multi_tensor/_init__.pyimportstorch.optimThis seemed to work fine (green signals everywhere) but caused some internal test failures after it landed; an infinite recursion during import.
PR Remove circular import #128875 attempted to fix this by removing the import from
torch/optim/__init__.py.This seemed to work fine: green signals everywhere and the failing tests started passing but a smaller number of tests started failing; unable to import
torch.optim._multi_tensorDifferential Revision: D58815471