restore zero-filled storage when called (#21917) - #21917
Open
JakeStevens wants to merge 1 commit into
Open
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21917
Note: Links to docs will display an error until the docs builds have been completed. This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D116459495. |
This PR needs a
|
digantdesai
approved these changes
Aug 18, 2026
digantdesai
left a comment
digantdesai
left a comment
Contributor
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
meta-codesync
Bot
changed the title
JakeStevens
added a commit
to JakeStevens/executorch
that referenced
this pull request
Aug 18, 2026
Summary: To reduce GPU memory during compilation, ExecuTorch intentionally keeps shape metadata but frees its storage with resize_(0). The existing workaround restored the zero-filled storage only when called from Inductor’s _unlift_graph. PyTorch 2.13 introduced other cloning paths during AOTI compilation/autotuning. Those paths attempted to clone the shaped tensor while its storage was still zero bytes, producing: ``` required storage: 4194304 bytes actual storage: 0 bytes ``` The fix now detects any intentionally emptied KV-cache tensor before examining the caller and restores zero-filled storage Reviewed By: digantdesai Differential Revision: D116459495
JakeStevens
force-pushed
the
export-D116459495
branch
from
August 18, 2026 17:07
507e40e to
4e68d6d
Compare
Summary: To reduce GPU memory during compilation, ExecuTorch intentionally keeps shape metadata but frees its storage with resize_(0). The existing workaround restored the zero-filled storage only when called from Inductor’s _unlift_graph. PyTorch 2.13 introduced other cloning paths during AOTI compilation/autotuning. Those paths attempted to clone the shaped tensor while its storage was still zero bytes, producing: ``` required storage: 4194304 bytes actual storage: 0 bytes ``` The fix now detects any intentionally emptied KV-cache tensor before examining the caller and restores zero-filled storage Reviewed By: digantdesai Differential Revision: D116459495
JakeStevens
force-pushed
the
export-D116459495
branch
2 times, most recently
from
August 19, 2026 13:37
4e68d6d to
c9aabb1
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
To reduce GPU memory during compilation, ExecuTorch intentionally keeps shape metadata but frees its storage with resize_(0).
The existing workaround restored the zero-filled storage only when called from Inductor’s _unlift_graph. PyTorch 2.13 introduced other cloning paths during AOTI compilation/autotuning. Those paths attempted to clone the shaped tensor while its storage was still zero bytes, producing:
The fix now detects any intentionally emptied KV-cache tensor before examining the caller and restores zero-filled storage
Reviewed By: digantdesai
Differential Revision: D116459495