[Test] Add vllm_ut test

[pranavprakash-amd]

Motivation

The rocm-tests framework validates the full ROCm stack up through ML frameworks, but had no coverage for vLLM, a critical inference engine in the
ROCm ML ecosystem. Upstream vLLM ships its own pytest suite, but it exists as a standalone shell script (run_unit_tests) outside our framework —
meaning it can't participate in marker-based selection, scheduling, Allure reporting, or the remote-fleet/container executor infrastructure.

This PR ports that suite into rocm-tests as a first-class, marker-compliant e2e.stack test under a new tests/e2e/frameworks/vllm/ area, giving us
continuous vLLM-on-ROCm signal in nightly (focused) and weekly (full) CI gates.

Technical Details

New test area — tests/e2e/frameworks/{,vllm/} with package init.py files and a CATEGORY_PROFILE entry in framework/markers/taxonomy.py
supplying hw.gpu / layer.runtime / ci.weekly / e2e.stack / os.linux by default.

test_vllm_ut.py — two test functions driven by a ContainerExecutor:

• test_vllm_ut — parametrized over 7 upstream sub-suites (core, test_regression.py, engine trio, tokenization, test_logits_processor.py,
  tool_use, AWQ-Triton kernel). Marked ci.weekly + runtime.soak (multi-hour, downloads gated HF models).
• test_vllm_ut_int4_awq_triton — runs only the AWQ-Triton quantization kernel (source's --int4 fast path). Overrides the profile to ci.nightly +
  runtime.medium for a lighter daily signal.

Test Plan

Marker/collection validation (no GPU) — confirms profile injection & parametrization

pytest tests/e2e/frameworks/vllm/ --collect-only -q --no-gpu

Nightly focused case on a real GPU + vLLM container

HF_TOKEN= pytest tests/e2e/frameworks/vllm/
-m "hw.gpu and ci.nightly" --container-mode -v

Full weekly suite

HF_TOKEN= pytest tests/e2e/frameworks/vllm/
-m "hw.gpu and ci.weekly" --container-mode --alluredir=build/allure-results -v

Test Result

====================================================== ROCm Test Suite Summary ======================================================
Test Directory PASS FAIL SKIP ERROR Duration
─────────────────────────────────────────────────────────────
tests/frameworks/vllm 0 8 0 0 1726.2 s
─────────────────────────────────────────────────────────────
TOTAL 8 tests │ 0 passed │ 8 failed │ 0 skipped │ 0 error │ 1726.2 s

Note: The failures are part of the pytest tests that are triggered inside the vllm. Those failures are not because of the current test as its just invoking them inside the container.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[y1k2chaitanya]

Automated review — 1 blocker (required CI check failing):

• pre-commit (black --check) is FAILING → black would reformat both new files, so this required gate is red and blocks merge. Run black tests/frameworks/vllm/ and push. Specifics:
  • tests/frameworks/vllm/conftest.py:96 — needs a blank line after import logging, the any(...) generator collapsed to one line (~L102), and the final pytest.skip(...) collapsed (~L115).
  • tests/frameworks/vllm/test_vllm_ut.py:167 — collapse the split apt-get ... "curl libsodium23" string, and reformat the three commands.extend([...]) blocks (~L177-L208) and the tests-tree commands.extend([...]) (~L214) to black's style.

Non-blocking nits (no action required to approve): module docstrings reference the path tests/e2e/frameworks/vllm/ but the files live at tests/frameworks/vllm/.

All imports (framework.reporting.allure_reporter.report_metric), the container_executor fixture, probe()/run(), and the --no-gpu/--container-image options resolve, and the dotted markers match the registered taxonomy — so once black passes this is otherwise mergeable.

[pranavprakash-amd]

Run result::

====================================================== ROCm Test Suite Summary ======================================================
Test Directory PASS FAIL SKIP ERROR Duration
─────────────────────────────────────────────────────────────
tests/frameworks/vllm 0 8 0 0 1726.2 s
─────────────────────────────────────────────────────────────
TOTAL 8 tests │ 0 passed │ 8 failed │ 0 skipped │ 0 error │ 1726.2 s

Failed tests:
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[core]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[test_regression_py]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[engine_test_sequence_py_test_config_py_test_logger_py]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[tokenization]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[test_logits_processor_py]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[tool_use]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut[kernels_quantization_test_awq_triton_py]
• tests/frameworks/vllm/test_vllm_ut.py::test_vllm_ut_int4_awq_triton

Note: The failures are part of the pytest tests that are triggered inside the vllm. Those failures are not because of the current test as its just invoking them inside the container.