Use AMD CI workflow defined in hf-workflows #35058
Conversation
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
LGTM. But is it still in draft mode? Also, does everything run smooth with this new |
Yes, I kept it in draft until the PR in hf-workflows was merged so that I could test 👍 |
There was a problem hiding this comment.
OK. Thanks.
I would suggest however using the push event to test if everything is working fine before merge however. By fine I means the call is working and we know the resources could be accessed etc, not meaning the jobs / reports themselves.
(There are line ending issues which was CRLF, it was my bad, and it's good to change to LF here)
utils/notification_service.py
Outdated
| @@ -920,6 +922,8 @@ def prepare_reports(title, header, reports, to_truncate=True): | |||
|
|
|||
| if __name__ == "__main__": | |||
| SLACK_REPORT_CHANNEL_ID = os.environ["SLACK_REPORT_CHANNEL"] | |||
| REPORT_REPO_ID = os.environ.get("REPORT_REPO_ID", "hf-internal-testing/transformers_daily_ci") | |||
There was a problem hiding this comment.
hmm, IIRC, REPORT_REPO_ID will exist as you define it in slack-report.yml, despite it could be empty string if not from AMD (i.e. for Nvidia CI). And we need to change the condition here.
(just like what you did one line below for UPLOAD_REPORT_SUMMARY)
There was a problem hiding this comment.
Right, have to check for empty string as well.
utils/notification_service.py
Outdated
| @@ -533,11 +535,11 @@ def payload(self) -> str: | |||
| commit_info = api.upload_file( | |||
| path_or_fileobj=file_path, | |||
| path_in_repo=f"{datetime.datetime.today().strftime('%Y-%m-%d')}/ci_results_{job_name}/new_model_failures.txt", | |||
| repo_id="hf-internal-testing/transformers_daily_ci", | |||
| repo_id=self.repo_id, | |||
There was a problem hiding this comment.
the new_model_failures stuff depends on prev_ci_artifacts which is defined by get_last_daily_ci_reports (called in this file toward the end of this file).
It's hard coded to fetch the previous run of Nvidia scheduled CI runs.
If AMD also want to have these files uploaded, we have to adjust the definition of get_last_daily_ci_reports
There was a problem hiding this comment.
Great. I'll add support for differentiating them.
utils/notification_service.py
Outdated
| @@ -1220,7 +1227,8 @@ def prepare_reports(title, header, reports, to_truncate=True): | |||
| os.makedirs(os.path.join(os.getcwd(), f"ci_results_{job_name}")) | |||
|
|
|||
| target_workflow = "huggingface/transformers/.github/workflows/self-scheduled-caller.yml@refs/heads/main" | |||
| is_scheduled_ci_run = os.environ.get("CI_WORKFLOW_REF") == target_workflow | |||
| amd_target_workflow = "huggingface/transformers/.github/workflows/self-scheduled-amd-caller.yml@refs/heads/main" | |||
| is_scheduled_ci_run = os.environ.get("CI_WORKFLOW_REF") in [target_workflow, amd_target_workflow] | |||
There was a problem hiding this comment.
I am a bit uncertain here
CI_WORKFLOW_REF: ${{ github.workflow_ref }}
when it comes to AMD CI (called from hf-workflows ), what github.workflow_ref would be ?
There was a problem hiding this comment.
oh my bad.
The triggered workflow run would be .github/workflows/self-scheduled-amd-caller.yml, and hf-workflows ones are being called
* Use AMD CI workflow defined in hf-workflows
What does this PR do?
Uses external workflows defined in hf-workflows for self hosted AMD runners. This removes a possible attack vector on self hosted CI runners on github.