CI for runner-et

[metascroy]

Attempt CI for runner-et with new ./scripts/install_et.sh

[guangy10]

How is it different from

python generate.py --checkpoint-path ${MODEL_PATH} --temperature 0 --pte-path ${PWD}/${MODEL_NAME}.pte  > ${PWD}/output_et
          cat ${PWD}/output_et

Instead of adding a new worklfow for mac, can you add the et-runner changes to et.yml? You can extend the job to run both on linux and mac. I hope we can soon moving all the setup steps to some scripts and setup the same model tests we have in ExecuTorch so developers can really focus on enabling models w/ what configs (xnnpack, quant, aoti, etc.) in llama-fast but no worry about to run the model across all the platforms.

[metascroy]

The difference is that python generate.py calls ET using pybindings, and ./build/cmake-out/runner_et directly uses the ET C++ APIs (see https://github.com/pytorch-labs/llama-fast/tree/main/runner-et). There is currently no CI testing the runner-et, and I'd like it to not break.

I'd love to have these unified, but I cannot get them to run on the same runner. I cannot install the executorch python library on macos-14-xlarge because of this issue (https://fb.workplace.com/groups/osssupport/permalink/26054350164186919/), and I cannot get runner-et to build on linux. So this script installs executorch without pybindings (so that it works on macos-14-xlarge), and then builds (runner-et).