split single_gpu and multi_gpu

[ydshieh]

What does this PR do?

Fix the scheduled CI issue caused by the 256 limits (jobs generated from matrix).

Note that the workflow run page has a graph that has no single-gpu and multi-gpu on it. But on the left side, the job names have matrix mentioned.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[LysandreJik]

Thanks for your PR! Did you launch it as a trial to see if it works? I see the following should be completed, on line 303:

[setup, run_tests_gpu, run_examples_gpu, run_pipelines_tf_gpu, run_pipelines_torch_gpu, run_all_tests_torch_cuda_extensions_gpu]

[ydshieh]

Thanks for your PR! Did you launch it as a trial to see if it works? I see the following should be completed, on line 303:

[setup, run_tests_gpu, run_examples_gpu, run_pipelines_tf_gpu, run_pipelines_torch_gpu, run_all_tests_torch_cuda_extensions_gpu]

You are right, that line should be changed. I haven't launched it (just tried with a dummy example). I will launch it now.

[LysandreJik]

You can launch it with only 1-2 models in each run, for example by updating this line:

          echo "::set-output name=matrix::$(python3 -c 'import os; tests = os.getcwd(); model_tests = os.listdir(os.path.join(tests, "models")); d1 = sorted(list(filter(os.path.isdir, os.listdir(tests)))); d2 = sorted(list(filter(os.path.isdir, [f"models/{x}" for x in model_tests]))); d1.remove("models"); d = d2 + d1; print(d)')"

to

          echo "::set-output name=matrix::$(python3 -c 'import os; tests = os.getcwd(); model_tests = os.listdir(os.path.join(tests, "models"))[:2]; d1 = sorted(list(filter(os.path.isdir, os.listdir(tests)))); d2 = sorted(list(filter(os.path.isdir, [f"models/{x}" for x in model_tests]))); d1.remove("models"); d = d2 + d1; print(d)')"

This way you'll test the full behavior without having 12-hour long iterations.

[ydshieh]

It took sometime, but the run looks good.

https://github.com/huggingface/transformers/actions/runs/2276209307

[LysandreJik]

Looks good, thanks @ydshieh!