AOSP related issues

[kirklandsign]

Using this issue as the main thread for discussing AOSP related issues.

Original discussion: 51cef8c#diff-757cd72756ef920fd2357653c76c087315969557e257a0fc8532ed2c6dca1649R105

Known issues:

1. java test coverage <- WIP
2. [high priority] use fd in data loader like https://cs.android.com/android/platform/superproject/main/+/main:external/federated-compute/fcp/tensorflow/file_descriptor_filesystem.cc
3. java stability
4. question about HybridData, do we really need it?
   executorch/extension/android/src/main/java/org/pytorch/executorch/Tensor.java <- Yes
5. when port executorch v0.3.0, got error "This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository." It means this release is not cut from main branch and qiaoli31 can't port into aosp. Is there a way to fix this tag and make it cut from main branch. In addition, could we make sure we cut release from main branch in the future
6. [high priority] could we expose a serializable format of EValue and Tensor class? The usage is partner can use this class to create Object[] for inferenceInput and ODP can use it to convert byte[] back to these data class when run executorch. We have deadline mid Nov including code complete and API review, so this is request is a little bit urgent. Let me know if this request makes sense or not.
   https://cs.android.com/android/platform/superproject/main/+/main:packages/modules/OnDevicePersonalization/framework/java/android/adservices/ondevicepersonalization/InferenceInput.java;l=56?q=inferenceinput

[qiaoli31]

6. [high priority] could we expose a serializable format of EValue and Tensor class? The usage is partner can use this class to create Object[] for inferenceInput and ODP can use it to convert byte[] back to these data class when run executorch. We have deadline mid Nov including code complete and API review, so this is request is a little bit urgent. Let me know if this request makes sense or not.
   https://cs.android.com/android/platform/superproject/main/+/main:packages/modules/OnDevicePersonalization/framework/java/android/adservices/ondevicepersonalization/InferenceInput.java;l=56?q=inferenceinput

[kirklandsign]

Hi @qiaoli31, I guess for tensor, it's easier. I'm not sure about EValue use case other than tensor. However, if we serialize it for AOSP only, do we need to guarantee stability?

[qiaoli31]

we need the change in main branch not only AOSP. yes, we need this class to be stable because it will be used by both partner and ODP.

btw for request 1-5 as well, we need it in main branch, otherwise we need resolve conflict or manually edit every time because we are not allowed to port "random" branch. This aosp branch is only for testing purpose for us as well.

[kirklandsign]

6. [high priority] could we expose a serializable format of EValue and Tensor class? The usage is partner can use this class to create Object[] for inferenceInput and ODP can use it to convert byte[] back to these data class when run executorch. We have deadline mid Nov including code complete and API review, so this is request is a little bit urgent. Let me know if this request makes sense or not.
   https://cs.android.com/android/platform/superproject/main/+/main:packages/modules/OnDevicePersonalization/framework/java/android/adservices/ondevicepersonalization/InferenceInput.java;l=56?q=inferenceinput

Oh is this for Java layer only?

[qiaoli31]

EValue is quite straightforward as long as Tensor is serializable, all other types are primitive types.
https://github.com/pytorch/executorch/blob/eea172de0d4c16ff2a831604c49efb9f47cc8df6/extension/android/src/main/java/org/pytorch/executorch/EValue.java

[qiaoli31]

Java class only. More specifically is these two classes.
https://github.com/pytorch/executorch/blob/eea172de0d4c16ff2a831604c49efb9f47cc8df6/extension/android/src/main/java/org/pytorch/executorch/EValue.java
https://github.com/pytorch/executorch/blob/eea172de0d4c16ff2a831604c49efb9f47cc8df6/extension/android/src/main/java/org/pytorch/executorch/Tensor.java

[kirklandsign]

Sounds good. Basically we put a header for type and for tensor, the sizes and dim order. I will try to define one and ask our team for a review.

[qiaoli31]

Do we have ETA for item (6) which is the highest priority?

[kirklandsign]

Hi @qiaoli31 I will ramp up another engineer soon. Meanwhile I will work on (6)

[qiaoli31]

sounds good, thank you!