Run linux_job_v2 workflow called by ROCm job only on cron schedule and when push to main

[akashveramd]

In this PR, added the condition in linux_job_v2.yml to run workflow called by ROCm job only on cron schedule and when push to main. This feature is optional and shouldn't affect other workflows. The calling ROCm workflow must set the input 'check-run-condition' to true for using this condition.
This effort is to limit the PR runs on ROCm runner due to limited CI resources and is part of torchtitan PR pytorch/torchtitan#2018

[vercel]

@akashveramd is attempting to deploy a commit to the Meta Open Source Team on Vercel.

A member of the Team first needs to authorize it.

[akashveramd]

@huydhn @malfet: Can I please get a review for this PR?

[seemethere]

No, I don't want to add a rocm specific conditional to the generic job

[akashveramd]

No, I don't want to add a rocm specific conditional to the generic job

I see. However, I think we already have rocm specific conditionals https://github.com/pytorch/test-infra/blob/main/.github/workflows/linux_job_v2.yml#L169 in the existing file?

[seemethere]

No, I don't want to add a rocm specific conditional to the generic job

I see. However, I think we already have rocm specific conditionals main/.github/workflows/linux_job_v2.yml#L169 in the existing file?

That conditional makes sense, the one you are introducing unfortunately does not.

If you want this conditional make it upstream in the torchtitan repository.

[huydhn]

Echo @seemethere point here is that this would not fly on linux_job_v2 because the if condition is too specific to ROCm use case. You could try to move that line here to the caller from torchtitan https://github.com/pytorch/torchtitan/blob/main/.github/workflows/integration_test_8gpu_features.yaml#L29.

Also, it's important to note that GHA if wouldn't be able to check for value in the matrix like rocm or cuda directly, but there is way to generate the matrix dynamically like this example https://github.com/pytorch/pytorch-integration-testing/blob/main/.github/workflows/vllm-benchmark.yml#L74. Maybe it's easier to just return early inside the job if it's rocm and not a push or a schedule