Use weight cache for quantized tensor scale data #14448
Merged
+33
−32
Conversation
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14448
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 1 Cancelled JobAs of commit 0ce8b1e with merge base 07d1092 ( NEW FAILURE - The following job has failed:
CANCELLED JOB - The following job was cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
@GregoryComer has exported this pull request. If you are a Meta employee, you can view the originating diff in D82862629. |
GregoryComer
force-pushed
the
export-D82862629
branch
from
2cb9258 to
9568209
Compare
|
@GregoryComer has imported this pull request. If you are a Meta employee, you can view this in D82862629. |
GregoryComer
added
the
release notes: none
|
Adding to 1.0 release blockers, as weight cache is enabled by default in several presets (including Android). QC8 and QB4 linear is broken when it's enabled, so we'll want to cherry pick this fix. |
digantdesai
approved these changes
Sep 19, 2025
digantdesai
approved these changes
Sep 19, 2025
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
Sep 19, 2025
Summary: When enabling the XNNPACK weight cache and running a model with qb4 or qc8-quantized linear weights, it triggers an assertion that is intended to make sure all data is in the weight cache. This can be reproduced by running the XNNPACK backend linear op tests with weight cache enabled. The root cause appears to be that tensor scale data was bypassing the weight cache - likely an oversight in initial implementation. This isn't a correctness issue, but does cause the aforementioned assert to fail and uses marginally more memory than it otherwise needs to. This PR updates the XNNPACK compileModel call to use the weight cache for scale data (instead of putting it in the unpacked_buffers list). With this change, the linear op tests pass with weight cache enabled. Test Plan: ``` buck test -c executorch.xnnpack_weights_cache=1 fbcode//executorch/backends/xnnpack/test:test_xnnpack_ops -- linear ``` Reviewed By: digantdesai Differential Revision: D82862629 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D82862629
branch
from
9568209 to
391c7b3
Compare
|
@GregoryComer has exported this pull request. If you are a Meta employee, you can view the originating diff in D82862629. |
lucylq
approved these changes
Sep 19, 2025
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
Sep 19, 2025
Summary: When enabling the XNNPACK weight cache and running a model with qb4 or qc8-quantized linear weights, it triggers an assertion that is intended to make sure all data is in the weight cache. This can be reproduced by running the XNNPACK backend linear op tests with weight cache enabled. The root cause appears to be that tensor scale data was bypassing the weight cache - likely an oversight in initial implementation. This isn't a correctness issue, but does cause the aforementioned assert to fail and uses marginally more memory than it otherwise needs to. This PR updates the XNNPACK compileModel call to use the weight cache for scale data (instead of putting it in the unpacked_buffers list). With this change, the linear op tests pass with weight cache enabled. Test Plan: ``` buck test -c executorch.xnnpack_weights_cache=1 fbcode//executorch/backends/xnnpack/test:test_xnnpack_ops -- linear ``` Reviewed By: lucylq, digantdesai Differential Revision: D82862629 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D82862629
branch
from
391c7b3 to
41fc9c4
Compare
|
@GregoryComer has exported this pull request. If you are a Meta employee, you can view the originating diff in D82862629. |
Summary: When enabling the XNNPACK weight cache and running a model with qb4 or qc8-quantized linear weights, it triggers an assertion that is intended to make sure all data is in the weight cache. This can be reproduced by running the XNNPACK backend linear op tests with weight cache enabled. The root cause appears to be that tensor scale data was bypassing the weight cache - likely an oversight in initial implementation. This isn't a correctness issue, but does cause the aforementioned assert to fail and uses marginally more memory than it otherwise needs to. This PR updates the XNNPACK compileModel call to use the weight cache for scale data (instead of putting it in the unpacked_buffers list). With this change, the linear op tests pass with weight cache enabled. Test Plan: ``` buck test -c executorch.xnnpack_weights_cache=1 fbcode//executorch/backends/xnnpack/test:test_xnnpack_ops -- linear ``` Reviewed By: lucylq, digantdesai Differential Revision: D82862629 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D82862629
branch
from
41fc9c4 to
0ce8b1e
Compare
|
@GregoryComer has imported this pull request. If you are a Meta employee, you can view this in D82862629. |
|
@pytorchbot cherry-pick --onto release/1.0 -c fixnewfeature |
Cherry picking #14448The cherry pick PR is at #14455 and it is recommended to link a fixnewfeature cherry pick PR with an issue. The following tracker issues are updated: Details for Dev Infra teamRaised by workflow job |
StrycekSimon
pushed a commit
to nxp-upstream/executorch
that referenced
this pull request
Sep 23, 2025
Differential Revision: D82862629 Pull Request resolved: pytorch#14448
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
6 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:
When enabling the XNNPACK weight cache and running a model with qb4 or qc8-quantized linear weights, it triggers an assertion that is intended to make sure all data is in the weight cache. This can be reproduced by running the XNNPACK backend linear op tests with weight cache enabled.
The root cause appears to be that tensor scale data is bypassing the weight cache - likely an oversight. This isn't a correctness issue, but does cause the aforementioned assert to fail and uses marginally more memory than it otherwise needs to.
This PR updates the XNNPACK compileModel call to use the weight cache for scale data (instead of putting it in the unpacked_buffers list). With this change, the linear op tests pass with weight cache enabled.
Differential Revision: D82862629