Better Shape Function Registration#3237
Closed
jwfromm wants to merge 1 commit into
Closed
Conversation
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
jwfromm
force-pushed
the
export-D64147797
branch
from
d0c968e to
5f0bfb1
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
jwfromm
force-pushed
the
export-D64147797
branch
from
5f0bfb1 to
7f1aca2
Compare
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Oct 17, 2024
Summary: X-link: facebookresearch/FBGEMM#338 With the increasing use of torch.compile both internally and in open source, I wanted to take some time to think about how to best make our FBGEMM custom ops compile compatible. This diff attempts to implement a testable and scalable way of registering custom shape functions easily. `torch.compile` requires that custom operators have shape functions registered, as they are needed for tracing. In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators. Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use `register_fake`. It turns out `register_fake` is the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph. This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors. Reviewed By: jianyuh Differential Revision: D64147797
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Oct 17, 2024
Summary: X-link: facebookresearch/FBGEMM#338 With the increasing use of torch.compile both internally and in open source, I wanted to take some time to think about how to best make our FBGEMM custom ops compile compatible. This diff attempts to implement a testable and scalable way of registering custom shape functions easily. `torch.compile` requires that custom operators have shape functions registered, as they are needed for tracing. In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators. Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use `register_fake`. It turns out `register_fake` is the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph. This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors. Reviewed By: jianyuh, jiawenliu64 Differential Revision: D64147797
jwfromm
force-pushed
the
export-D64147797
branch
from
7f1aca2 to
5c48315
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Oct 17, 2024
Summary: X-link: facebookresearch/FBGEMM#338 With the increasing use of torch.compile both internally and in open source, I wanted to take some time to think about how to best make our FBGEMM custom ops compile compatible. This diff attempts to implement a testable and scalable way of registering custom shape functions easily. `torch.compile` requires that custom operators have shape functions registered, as they are needed for tracing. In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators. Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use `register_fake`. It turns out `register_fake` is the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph. This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors. Reviewed By: jianyuh, jiawenliu64 Differential Revision: D64147797
jwfromm
force-pushed
the
export-D64147797
branch
from
5c48315 to
d0a6304
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Oct 17, 2024
Summary: X-link: facebookresearch/FBGEMM#338 With the increasing use of torch.compile both internally and in open source, I wanted to take some time to think about how to best make our FBGEMM custom ops compile compatible. This diff attempts to implement a testable and scalable way of registering custom shape functions easily. `torch.compile` requires that custom operators have shape functions registered, as they are needed for tracing. In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators. Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use `register_fake`. It turns out `register_fake` is the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph. This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors. Reviewed By: jianyuh, jiawenliu64 Differential Revision: D64147797
jwfromm
force-pushed
the
export-D64147797
branch
from
d0a6304 to
d0dd9b8
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
jwfromm
added a commit
to jwfromm/FBGEMM
that referenced
this pull request
Oct 17, 2024
Summary: X-link: facebookresearch/FBGEMM#338 With the increasing use of torch.compile both internally and in open source, I wanted to take some time to think about how to best make our FBGEMM custom ops compile compatible. This diff attempts to implement a testable and scalable way of registering custom shape functions easily. `torch.compile` requires that custom operators have shape functions registered, as they are needed for tracing. In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators. Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use `register_fake`. It turns out `register_fake` is the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph. This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors. Reviewed By: jianyuh, jiawenliu64 Differential Revision: D64147797
jwfromm
force-pushed
the
export-D64147797
branch
from
d0dd9b8 to
237c0a4
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
Summary: X-link: facebookresearch/FBGEMM#338 With the increasing use of torch.compile both internally and in open source, I wanted to take some time to think about how to best make our FBGEMM custom ops compile compatible. This diff attempts to implement a testable and scalable way of registering custom shape functions easily. `torch.compile` requires that custom operators have shape functions registered, as they are needed for tracing. In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators. Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use `register_fake`. It turns out `register_fake` is the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph. This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors. Reviewed By: jianyuh, jiawenliu64 Differential Revision: D64147797
jwfromm
force-pushed
the
export-D64147797
branch
from
237c0a4 to
bc3254c
Compare
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D64147797 |
Contributor
|
This pull request has been merged in 4ba523c. |
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:
torch.compilerequires that custom operators have shape functions registered, as they are needed for tracing.In FBGEMM, we have inconsistently registered such shape functions. This diff attempts to clean up and register a good number of commonly used operators.
Notably, pytorch allows two methods of registering shape functions for custom ops. In CPP, you can use a Meta function, or in python you can use
register_fake. It turns outregister_fakeis the recommended and more powerful approach. For example, it is needed for ops that cross devices (such as the car ops) and is needed when exporting a traced graph.This diff thus focuses on the register_fake method and converts a handful of Meta registrations to it. My hope is that this can provide an easily extensible way of registering shape functions for other kernel authors.
Differential Revision: D64147797