Skip to content

[executorch][runtime] Introduce CoreDataMap for weight sharing #8848

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

Closed
wants to merge 2 commits into from

Conversation

lucylq
Copy link
Contributor

@lucylq lucylq commented Feb 28, 2025

Stack from ghstack (oldest at bottom):

CoreDataMap is the NamedDataMap that will live in the runtime. It is used to give delegates access to opaque named data stored in the PTE file. Open to alternative naming suggestions, maybe 'PTEDataMap' or 'ProgramDataMap'?

Usage
The CoreDataMap is owned by the program, and instantiated at program load time if named_data exists in the PTE file. We introduce usage of 'std::optional' here. I think we can also use executorch::aten::optional to avoid adding standard lib ?

When initializing delegates, the CoreDataMap is given to delegate_init. Delegates can retrieve opaque delegate data by key using 'get_data'. This gives them a FreeableBuffer that they can free later.

Testing
This test uses the C++ flatbuffer API to build a fake program containing named data. We also creates a temp file with sample data that the data loader can wrap around.

TODO: e2e test once delegate aot is ready and we can generate a file with named data.

Note
As the CoreDataMap wraps around flatbuffer constructs, the Program must outlive the CoreDataMap.

CoreDataMap does not implement

  • get_metadata; currently, all data stored is opaque. Later, we can implement get_metadata if a backend stores plain tensor data.
  • load_into; this is mostly used for the training case, and isn't used by delegates, at least not at the moment.

Differential Revision: D70213646

CoreDataMap is the NamedDataMap that will live in the runtime. It is used to give delegates access to opaque named data stored in the PTE file. Open to alternative naming suggestions, maybe 'PTEDataMap' or 'ProgramDataMap'?

**Usage**
The CoreDataMap is owned by the program, and instantiated at program load time if named_data exists in the PTE file. We introduce usage of 'std::optional' here. I think we can also use executorch::aten::optional to avoid adding standard lib ?

When initializing delegates, the CoreDataMap is given to delegate_init. Delegates can retrieve opaque delegate data by key using 'get_data'. This gives them a FreeableBuffer that they can free later.

**Testing**
This test uses the C++ flatbuffer API to build a fake program containing named data. We also creates a temp file with sample data that the data loader can wrap around.

TODO: e2e test once delegate aot is ready and we can generate a file with named data.

**Note**
As the CoreDataMap wraps around flatbuffer constructs, the Program must outlive the CoreDataMap.

CoreDataMap does not implement
- get_metadata; currently, all data stored is opaque. Later, we can implement get_metadata if a backend stores plain tensor data.
- load_into; this is mostly used for the training case, and isn't used by delegates, at least not at the moment.

Differential Revision: [D70213646](https://our.internmc.facebook.com/intern/diff/D70213646/)

[ghstack-poisoned]
Copy link

pytorch-bot bot commented Feb 28, 2025

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/8848

Note: Links to docs will display an error until the docs builds have been completed.

❌ 36 New Failures

As of commit 02362f1 with merge base 4df0ade (image):

NEW FAILURES - The following jobs have failed:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

lucylq added a commit that referenced this pull request Feb 28, 2025
CoreDataMap is the NamedDataMap that will live in the runtime. It is used to give delegates access to opaque named data stored in the PTE file. Open to alternative naming suggestions, maybe 'PTEDataMap' or 'ProgramDataMap'?

**Usage**
The CoreDataMap is owned by the program, and instantiated at program load time if named_data exists in the PTE file. We introduce usage of 'std::optional' here. I think we can also use executorch::aten::optional to avoid adding standard lib ?

When initializing delegates, the CoreDataMap is given to delegate_init. Delegates can retrieve opaque delegate data by key using 'get_data'. This gives them a FreeableBuffer that they can free later.

**Testing**
This test uses the C++ flatbuffer API to build a fake program containing named data. We also creates a temp file with sample data that the data loader can wrap around.

TODO: e2e test once delegate aot is ready and we can generate a file with named data.

**Note**
As the CoreDataMap wraps around flatbuffer constructs, the Program must outlive the CoreDataMap.

CoreDataMap does not implement
- get_metadata; currently, all data stored is opaque. Later, we can implement get_metadata if a backend stores plain tensor data.
- load_into; this is mostly used for the training case, and isn't used by delegates, at least not at the moment.

Differential Revision: [D70213646](https://our.internmc.facebook.com/intern/diff/D70213646/)

ghstack-source-id: 269080307
Pull Request resolved: #8848
@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Feb 28, 2025
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D70213646

@lucylq lucylq added the release notes: runtime Changes related to the core runtime which loads the program methods, initializes delegates, and runs label Feb 28, 2025
…ring"

CoreDataMap is the NamedDataMap that will live in the runtime. It is used to give delegates access to opaque named data stored in the PTE file. Open to alternative naming suggestions, maybe 'PTEDataMap' or 'ProgramDataMap'?

**Usage**
The CoreDataMap is owned by the program, and instantiated at program load time if named_data exists in the PTE file. We introduce usage of 'std::optional' here. I think we can also use executorch::aten::optional to avoid adding standard lib ?

When initializing delegates, the CoreDataMap is given to delegate_init. Delegates can retrieve opaque delegate data by key using 'get_data'. This gives them a FreeableBuffer that they can free later.

**Testing**
This test uses the C++ flatbuffer API to build a fake program containing named data. We also creates a temp file with sample data that the data loader can wrap around.

TODO: e2e test once delegate aot is ready and we can generate a file with named data.

**Note**
As the CoreDataMap wraps around flatbuffer constructs, the Program must outlive the CoreDataMap.

CoreDataMap does not implement
- get_metadata; currently, all data stored is opaque. Later, we can implement get_metadata if a backend stores plain tensor data.
- load_into; this is mostly used for the training case, and isn't used by delegates, at least not at the moment.

Differential Revision: [D70213646](https://our.internmc.facebook.com/intern/diff/D70213646/)

[ghstack-poisoned]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. fb-exported release notes: runtime Changes related to the core runtime which loads the program methods, initializes delegates, and runs
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants