Add pass to remove unused parameters in to_edge #10484
Merged
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
GregoryComer
requested review from
JacobSzwejbka,
larryliu0820 and
tarun292
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/10484
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit 6cc9b48 with merge base 3064308 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
CLA Signed
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D73654202 |
GregoryComer
added
module: exir
GregoryComer
force-pushed
the
export-D73654202
branch
from
c518552 to
3087578
Compare
Contributor
|
@GregoryComer has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
Contributor
Can we run it in to_edge instead? We should remove unused things as soon as possible imo. |
Member
Author
Sure, I'll move it to there. |
JacobSzwejbka
approved these changes
Apr 28, 2025
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
May 5, 2025
Summary: Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior. This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example). This PR adds a new pass (`remove_unused_parameters_pass`) to strip unused parameters from the `ExportedProgram`. It is run as part of `to_executorch`. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph. As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this. cc JacobSzwejbka angelayi Reviewed By: digantdesai, JacobSzwejbka Differential Revision: D73654202 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D73654202
branch
from
3087578 to
e433b63
Compare
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
May 5, 2025
Summary: Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior. This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example). This PR adds a new pass (`remove_unused_parameters_pass`) to strip unused parameters from the `ExportedProgram`. It is run as part of `to_executorch`. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph. As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this. cc JacobSzwejbka angelayi Reviewed By: digantdesai, JacobSzwejbka Differential Revision: D73654202 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D73654202
branch
from
e433b63 to
5f941b1
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D73654202 |
JacobSzwejbka
approved these changes
May 5, 2025
GregoryComer
changed the title
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D73654202 |
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
May 5, 2025
Summary: Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior. This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example). This PR adds a new pass (`remove_unused_parameters_pass`) to strip unused parameters from the `ExportedProgram`. It is run as part of `to_executorch`. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph. As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this. cc JacobSzwejbka angelayi Pull Request resolved: pytorch#10484 Reviewed By: digantdesai, JacobSzwejbka Differential Revision: D73654202 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D73654202
branch
from
5f941b1 to
cc38cbd
Compare
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
May 5, 2025
Summary: Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior. This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example). This PR adds a new pass (`remove_unused_parameters_pass`) to strip unused parameters from the `ExportedProgram`. It is run as part of `to_executorch`. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph. As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this. cc JacobSzwejbka angelayi Pull Request resolved: pytorch#10484 Reviewed By: digantdesai, JacobSzwejbka Differential Revision: D73654202 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D73654202
branch
from
cc38cbd to
98d742d
Compare
GregoryComer
added a commit
to GregoryComer/executorch
that referenced
this pull request
May 5, 2025
Summary: Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior. This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example). This PR adds a new pass (`remove_unused_parameters_pass`) to strip unused parameters from the `ExportedProgram`. It is run as part of `to_executorch`. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph. As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this. cc JacobSzwejbka angelayi Pull Request resolved: pytorch#10484 Reviewed By: digantdesai, JacobSzwejbka Differential Revision: D73654202 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D73654202
branch
from
98d742d to
1067754
Compare
Summary: Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior. This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example). This PR adds a new pass (`remove_unused_parameters_pass`) to strip unused parameters from the `ExportedProgram`. It is run as part of `to_executorch`. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph. As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this. cc JacobSzwejbka angelayi Pull Request resolved: pytorch#10484 Reviewed By: digantdesai, JacobSzwejbka Differential Revision: D73654202 Pulled By: GregoryComer
GregoryComer
force-pushed
the
export-D73654202
branch
from
1067754 to
6cc9b48
Compare
Contributor
|
@GregoryComer has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
phaiting
pushed a commit
that referenced
this pull request
May 6, 2025
Differential Revision: D73654202 Pull Request resolved: #10484
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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:
Currently, ExecuTorch will serialize any parameters in the exported program, regardless of whether they are actually used. Exporting with strict=True will remove unused parameters, but strict=False will not. Export recently switched to non-strict as the default behavior.
This causes PTE bloat when doing pt2e quantization (unquantized weights are left in the graph) or sometimes when exporting multiple methods (encode and decoder, for example).
This PR adds a new pass (
remove_unused_parameters_pass) to strip unused parameters from theExportedProgram. It is run as part ofto_executorch. Parameters are considered unused if there are no uses of the placeholder node. Parameters are removed by stripping them from the state_dict, input specs, and graph.As a question for reviewers, should we run this pass earlier, as part of to_edge? My rationale for running as part of to_executorch was that it could theoretically clean up anything else left by partitioning and lowering, but I'm not aware of any concrete use cases for this.
Differential Revision: D73654202
cc @JacobSzwejbka @angelayi