-
Notifications
You must be signed in to change notification settings - Fork 25.2k
Refactored recomputable ops a bit and added a bunch more ops #85993
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
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/85993
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 FailuresAs of commit 20a413b: The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
[ghstack-poisoned]
pointwise_ops += [prims.div, prims.convert_element_type, aten.sign, aten.clone, aten._to_copy] # noqa: E501 | ||
misc_ops = [aten.to, aten.type_as, operator.getitem] | ||
recomputable_ops += [prims.div, prims.convert_element_type, aten.sign, aten.clone, aten._to_copy, aten.full_like, prims.var, prims.sum, aten.var, aten.std, prims.broadcast_in_dim, aten.select, aten.permute, aten._unsafe_view, aten.view, aten.expand, aten.slice, aten.reshape, aten.broadcast_tensors, aten.scalar_tensor, aten.ones, aten.new_zeros, aten.lift_fresh_copy, aten.minimum, aten.arange, aten.bitwise_and, aten.triu, aten.var_mean, aten.isinf, aten.any, aten.isnan, aten.full, aten.as_strided, aten.zeros, aten.argmax, aten.maximum, aten.bitwise_or, aten.logical_and, aten.logical_or] # noqa: E501 | ||
# Natalia said that we should allow recomputing indexing :) |
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.
It probably has to have a slightly higher multiplier on top of i/o bytes, but generally yes, it's a regular bw bound op, so recomputing it may be beneficial, subject to comparison of bytes read/written that mincut takes care of
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.
Yeah, we might need to do something smarter when it comes to "limiting amount of recomputation".
@pytorchbot merge -f "failing test is flaky OOM" |
@pytorchbot successfully started a merge job. Check the current status here. |
Hey @Chillee. |
Pull Request resolved: #85993 Approved by: https://github.com/ngimel
Stack from ghstack (oldest at bottom):