Repository navigation
Prepack stored inputs of dtype conversions - #24435
JakeStevens wants to merge 2 commits into
Conversation
Summary: Originally, the example supported only mv2. This PR extends that support to all models within torchvision. to be more specific, it allows you to export any torchvision model via a script-- it does *not* imply all export successfully or will run successfully. These will be addressed in follow-on PRs. Reviewed By: SS-JIA Differential Revision: D128684681
Summary: ## Ulterior Motive Unblock stored-constant dtype conversions in native MobileNetV3 graphs. ## Rationale **What**: Add `aten._to_copy.default` to the existing stored-input prepack list. **Why**: A stored constant resolves to a TensorRef, but the conversion shader needs a bound tensor resource. Passing the reference through leaves its input descriptor unwritten. ## Details - Reuse the existing prepack insertion and consumer replacement. - Retain the existing stored-binding and mutation guards. - Add a structural prepack test and a numerical int64-to-float conversion test. Authored with AI using Metamate on agentcloud. Differential Revision: D129088168
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24435
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 9d7306d with merge base cec2e73 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D129088168. |
This PR needs a
|
Summary:
Ulterior Motive
Unblock stored-constant dtype conversions in native MobileNetV3 graphs.
Rationale
What: Add
aten._to_copy.defaultto the existing stored-input prepack list.Why: A stored constant resolves to a TensorRef, but the conversion shader needs a bound tensor resource. Passing the reference through leaves its input descriptor unwritten.
Details
Authored with AI using Metamate on agentcloud.
Differential Revision: D129088168