PR GPU CI skips new-model GPU tests: Cosmos3 memory tests failed on an unrelated PR

[sayakpaul]

Summary

GPU-only model tests added in a PR can silently never run on that PR, then surface as failures on an unrelated PR later. This just happened with the Cosmos3 transformer: two TestCosmos3OmniTransformerMemory tests failed in the ace-step-lora PR's CI (run), even though ace-step-lora has nothing to do with Cosmos3.

This is relevant to the ongoing work on selecting which tests to run for a given PR — the fix here is really "make the PR GPU workflow run the tests that the PR's changes actually affect."

What happened

PR	Added	Ran its Cosmos3 GPU tests?
#13818 "Adding Cosmos 3"	transformer_cosmos3.py + AE + pipeline + docs + examples, but zero tests/ files	Workflow triggered (via examples/**), but no transformer tests existed yet
#14181 "Cosmos3 edge support"	First added tests/models/transformers/test_models_transformer_cosmos3.py incl. TestCosmos3OmniTransformerMemory	No — GPU workflow never triggered (see below)
#14193 "ace-step lora"	Touched a loader file	GPU workflow triggered → ran the full model suite → Cosmos3 memory tests ran for the first time → failed, blocking an unrelated PR

Root cause — two compounding gaps

1. pr_tests_gpu.yml's paths: filter is out of sync with push_tests.yml's.

• push_tests.yml (Fast GPU Tests on main) gates on src/diffusers/**.py, examples/**.py, tests/**.py.
• pr_tests_gpu.yml (Fast GPU Tests on PR) gates only on a narrow list of core loader/modeling files + examples/**/*.py. It excludes src/diffusers/models/** and tests/**.

So a PR that adds a new model + its tests (touching only src/diffusers/models/** and tests/models/**, as #14181 did) never triggers GPU CI on the PR. The tests first execute post-merge on main, or accidentally when some unrelated PR trips the narrow filter.

Net: GPU-only tests were introduced in a PR that structurally couldn't run them.

Suggested fixes (Claude)

1. Align the PR trigger with the push trigger — add src/diffusers/models/**.py, src/diffusers/pipelines/**.py, tests/models/**.py, tests/pipelines/**.py to pr_tests_gpu.yml's paths:.
2. Scope the pytest run to changed files so broadening the trigger doesn't run the entire tests/models suite (~17 min) on every model PR. Derive changed test files from git diff origin/main...HEAD and pass those (still combined with the mixin -k pattern). This overlaps directly with the "fetch tests to run per PR" work.
3. Stopgap: until the above lands, run /diffusers-bot pytest <new test file> (via pr_comment_gpu_tests.yml) on any new-model / new-transformer PR, and make it a review-checklist item.

@DN6 mentioning you since 2. above is probably also best implemented after your take on improving the test fetcher?