Fix Android app S3 path with tiktoken

[huydhn]

After #4203, the app is now uploaded to artifact/tiktoken_${{ matrix.tiktoken }} instead with tiktoken can be either ON or OFF. https://github.com/pytorch/executorch/blob/main/examples/demo-apps/android/LlamaDemo/README.md#alternative-2-build-from-local-machine mentions that tiktoken is only for Llama3. So, we can export it later in another archive like https://ossci-assets.s3.amazonaws.com/executorch-android-llama2-7b.zip when this is updated to run Llama3.

Note that the test is still unstable because the app runs OOM on S22 (#3344)

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/4289

• 📄 Preview Python docs built from this PR

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

❌ 1 New Failure

As of commit 13f29f0 with merge base 1d7d71d ():

NEW FAILURE - The following job has failed:

• Android / test-llama-app (OFF) / mobile-job (android) (gh)
  Process completed with exit code 1.

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

[facebook-github-bot]

@huydhn has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[guangy10]

Thanks for the quick fix @huydhn. I think it's better to temporarily disable/remove it from this android job. This android job runs on the PR and should only be used for quick validation of the android demo app build and run with a tiny model. It also runs in the GitHub action to prompt commit from main to stable, the instability would affect the stable branch health and release as we would prefer doing branch-cut off stable if possible. So the better place is to run the benchmarking job in a non-critical path.

I have #4287 to remove/disable the uploading and benchmarking portion off the android job
I also have #4288 starting refactor the android job to build different flavors of android demo apps, which will be used for benchmarking later. After merged those, I will start re-enable the benchmarking job for various of supported LLMs in a non-critical path.

[guangy10]

I think the reason of #4203 updating this to include tiktoken tokenizer is to ensure the demo app can build for Llama3. So as mentioned that this job should focus on validating different flavors of android demo app builds. If we remove tiktoken here, there will be a tcc regression.

[huydhn]

Yeah, this covers only llama2 without tiktoken atm. To cover llama3, for example, we would need to zip the exported model and the tokenizer it uses and upload the archive onto S3, i.e. https://github.com/pytorch/executorch/blob/main/.github/workflows/android.yml#L122. Then the archive can be passed to the job as its extra-data parameter to be copied to the device.