-
-
Notifications
You must be signed in to change notification settings - Fork 984
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Update to using pytorch-1.0-rc #1431
Conversation
* Update JIT usage to PyTorch 0.4.1 * Add _broadcast_all() patch for jit exploration * Uncomment debug code
@neerajprad will this be ready to merge as soon as conflicts are resolved? |
Just merged with dev. This should be good to merge once tests pass. |
@neerajprad what's the status of the performance issues as of the 11/28 nightly build? Are they mostly resolved as suggested in pytorch/pytorch#12190 ? |
The 11/28 nightly build was not available when I checked late last night, but the issue seems to be mostly resolved in my local build. So either its a platform specific issue or the way the wheels are built. Once this PR merges, I will update to the 11/28 build separately and update on that issue. |
Do we want to merge this before double-checking? |
It is an external issue that mostly affects HMC more than other things. It will be nice to have this in dev, and be able to make other major refactoring changes. We expect the perf issues to be resolved either in the final release or in the pre release. But we shouldn't merge this if in the worst case, we don't want to release using pytorch-1.0 if the perf issues persist! |
Let me quickly check if 11/28 is available, but I would like to merge this soon so as not to accumulate more merge conflicts. |
The new wheel seems to throw an error saying |
I think we can already see from the unit tests that the slowdown still persists, and many tests (mostly HMC) are taking almost twice as long. Note that unit test stage on this branch is expected to take longer because we have more tests, but even individual tests are significantly slower. I have not been able to replicate this locally though (yet). |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Woo hoo!
This is simply reopening #1376 with an updated branch name. Note that we will keep this as the base PR until PyTorch 1.0 is released. We should merge any changes from dev directly, but any additional functionality that uses
pytorch-1.0
as a base branch should be a regular PR.Tasks
pyro.distributions
to remove wrappers forenumerate_support
,expand
NOTE: The slowdown that we observed and noted in pytorch/pytorch#12190 still continues, and the unit tests are slower. My hope is that this will be resolved when pytorch/pytorch#13420 merges.