[CI/Build] Cut L3 entrypoint BAGEL cold starts via L4 small-model sleep suites (#6851) (#7359)

Signed-off-by: wangyu <410167048@qq.com>
Signed-off-by: [Your Name] <your.email@example.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
wangyuandCursor authored and GitHub committed 2026-09-23 22:06:58 -07:00
1 parent db82a55162
commit 3d43571b94
21 files changed
+1012 -487

No files matched your search

@@ -537,7 +537,10 @@ class TestQwen3OmniRealtimeWebSocket:
# These regressions use the engine-owned Qwen duplex plugin. The older class
# above intentionally exercises the separate turn/STT deployment.
_DUPLEX_REPEATS = int(os.environ.get("VLLM_OMNI_TEST_QWEN_DUPLEX_REPEATS", "20"))
# ``advanced_model`` (merge) uses 5 interruptions; ``full_model`` (nightly)
# keeps the 20-repeat race hunt. Marks select the case, not ``--run-level``.
_DUPLEX_REPEATS_ADVANCED = 5
_DUPLEX_REPEATS_FULL = 20
_DUPLEX_LONG_INSTRUCTIONS = "Answer the user's question in English using at least eight complete sentences."
_DUPLEX_SHORT_INSTRUCTIONS = "Answer the user's question in one short English sentence."
@@ -906,12 +909,12 @@ async def _run_duplex_history_pruning(runtime, log_path: Path) -> None:
_duplex_assert_completed_audio(events, response_id)
@pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "H100"}, num_cards=2)
@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=2)
class TestQwen3OmniDuplexPlayback:
@pytest.mark.advanced_model
@pytest.mark.parametrize("generation_finished", [False, True], ids=["generating", "queued-playback"])
@pytest.mark.parametrize("repetition", range(_DUPLEX_REPEATS))
@pytest.mark.parametrize("repetition", range(_DUPLEX_REPEATS_ADVANCED))
def test_speech_interrupts_playback(
self, qwen_duplex_server, tmp_path: Path, generation_finished: bool, repetition: int
):
@@ -922,5 +925,19 @@ class TestQwen3OmniDuplexPlayback:
)
)
@pytest.mark.full_model
@pytest.mark.parametrize("generation_finished", [False, True], ids=["generating", "queued-playback"])
@pytest.mark.parametrize("repetition", range(_DUPLEX_REPEATS_FULL))
def test_speech_interrupts_playback_full(
self, qwen_duplex_server, tmp_path: Path, generation_finished: bool, repetition: int
):
"""Same contract as ``test_speech_interrupts_playback``, 20 nightly repeats."""
asyncio.run(
_run_duplex_playback_interruption(
qwen_duplex_server, tmp_path / f"events-{repetition}.json", generation_finished=generation_finished
)
)
@pytest.mark.advanced_model
def test_mixed_image_audio_history_pruning(self, qwen_duplex_server, tmp_path: Path):
asyncio.run(_run_duplex_history_pruning(qwen_duplex_server, tmp_path / "events.json"))