Disable c10 op registration temporarily to reduce code size #26821
Conversation
[ghstack-poisoned]
pytorchbot
added
module: internals
Minimal revert of #26131 necessary to cutdown the codesize pre 1.3 until we figure out how to make templates better. Note the full revert doesn't work as many changes landed on top of it. On my linux machine libtorch.so (no cuda build) went from 98Mb to 75Mb [ghstack-poisoned]
smessmer
left a comment
smessmer
left a comment
There was a problem hiding this comment.
Looks good. Let's only land this into the release branch though, not to master.
Minimal revert of #26131 necessary to cutdown the codesize pre 1.3 until we figure out how to make templates better. Note the full revert doesn't work as many changes landed on top of it. On my linux machine libtorch.so (no cuda build) went from 98Mb to 75Mb Differential Revision: [D17576585](https://our.internmc.facebook.com/intern/diff/D17576585) [ghstack-poisoned]
pytorchbot
added
the
module: docs
|
|
smessmer
left a comment
smessmer
left a comment
There was a problem hiding this comment.
This PR is probably not necessary anymore because a lot of PRs landed fixing binary size issues. Feel free to override me though if you disagree.
|
based on @smessmer's comment and a discussion with @dzhulgakov, we are going to not land this. The stats are this would lower code size by 2 MB (out of ~90), but it's probably better to just keep this code as close as possible to master. |
Stack from ghstack:
Minimal revert of #26131 necessary to cutdown the codesize pre 1.3 until we figure out how to make templates better. Note the full revert doesn't work as many changes landed on top of it.
On my linux machine libtorch.so (no cuda build) went from 98Mb to 75Mb
Differential Revision: D17576585