Fix conda build

[leshikus]

No-basekit requires more work

[leshikus]

@pbchekin Regarding your last offline comments that reusable action becomes too big with this patch, how do you think it is better to approach the problem?

Should I modify the existing actions instead of installing pre-built wheels?

[pbchekin]

@pbchekin Regarding your last offline comments that reusable action becomes too big with this patch, how do you think it is better to approach the problem?

Should I modify the existing actions instead of installing pre-built wheels?

Roughly 50% of steps in build-test-reusable have a check for env_manager. IMHO, it is too much. At this point it is much cleaner just to create something like conda-build-test-reusable. I would expect some code duplication, but the code will be much simpler.

[leshikus]

@pbchekin the separation is completed, please take a look

btw, I wonder why conda workflow starts automatically on this PR - how it is implemented? is it some gh magic in some other workflow?

[pbchekin]

It starts because the workflow has

  pull_request:
    branches:
      - main

[pbchekin]

Do we need the new files in scripts such as run-{base,conda}.sh?

[leshikus]

you are correct, we no longer need run-base; we need run-conda, because I have also run-no-basekit in other branch

[leshikus]

It starts because the workflow has

@pbchekin this has to be reverted to the scheduled run, is it correct?

[pbchekin]

It starts because the workflow has

@pbchekin this has to be reverted to the scheduled run, is it correct?

Yes.

[pbchekin]

For some reason the original important comments are missing:

          # TODO: Find the fastest Hugging Face model
          # The script above always returns 0, so we need an additional check to see if the accuracy test passed

[leshikus]

# The script above always returns 0, so we need an additional check to see if the accuracy test passed

the script is no longer called, thus the note about its return status are no longer relevant

I will restore TODO

[leshikus]

Fixed

[pbchekin]

Suggested change

	INSTRUMENTATION_LIB_DIR=$(ls -1d $TRITON_PROJ/python/build/*lib*/triton/instrumentation) || SHARED_LIB_DIR=$(ls -1d $TRITON_PROJ/python/triton/_C) || err "Could not find $TRITON_PROJ/python/build/*lib*/triton/instrumentation, build Triton first"
	INSTRUMENTATION_LIB_DIR=$(ls -1d $TRITON_PROJ/python/build/*lib*/triton/instrumentation) || INSTRUMENTATION_LIB_DIR=$(ls -1d $TRITON_PROJ/python/triton/_C) || err "Could not find $TRITON_PROJ/python/build/*lib*/triton/instrumentation, build Triton first"

[pbchekin]

Also looks like the error message is no longer consistent with the verified locations. Also, conda workflow also builds triton, why $TRITON_PROJ/python/build/*lib*/triton/instrumentation does not work?

[leshikus]

The location is different for installed wheels. From the other side the instrumentation test do not work for installed wheels, so this change is optional anyway.

[pbchekin]

The location is different for installed wheels. From the other side the instrumentation test do not work for installed wheels, so this change is optional anyway.

I see. Let's undo this change for now.

[leshikus]

Here is passing conda test run, https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11575780630