Fit weekly serving fixtures to assigned GPUs and initialize accuracy groups

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/dfx/reliability/invalid_param_test/test_invalid_omni_chat.py

@@ -216,7 +216,7 @@ def test_video_chat_stream_invalid_requests(
 # WS /v1/realtime
 # ─────────────────────────────────────────────────────────────────────────────
 
-# Resolved in ``test_realtime_invalid_requests`` (skip ``session.created`` handshake).
+# Resolved in ``test_realtime_invalid_requests`` after the initial handshake.
 _REALTIME_WS_MODEL_MISMATCH = object()
 _REALTIME_WS_INVALID_AUDIO_APPEND = object()
 
@@ -286,7 +286,7 @@ def test_realtime_invalid_requests(
             "send_frames": send_frames,
             "timeout": 120,
             "ws_max_size": None,
-            "ws_skip_types": ["session.created"],
+            "ws_skip_types": ["session.created", "conversation.created"],
             "ws_json_type": "error",
             "ws_error_code": ws_error_code,
             "err_message": err_message,

tests/e2e/accuracy/sana_video/conftest.py

@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+import pytest
+
+
+@pytest.fixture
+def sana_model_parallel(tmp_path):
+    """Construct native parallel layers with real single-rank GPU groups."""
+    from vllm_omni.diffusion.distributed.parallel_state import (
+        destroy_distributed_env,
+        init_distributed_environment,
+        initialize_model_parallel,
+    )
+
+    init_distributed_environment(
+        world_size=1,
+        rank=0,
+        local_rank=0,
+        distributed_init_method=f"file://{tmp_path / 'sana-distributed'}",
+    )
+    try:
+        initialize_model_parallel()
+        yield
+    finally:
+        destroy_distributed_env()

tests/e2e/accuracy/sana_video/test_sana_video_golden.py

@@ -202,6 +202,7 @@ def test_sana_video_transformer_matches_frozen_golden(
     variant: str,
     _hardware,
     accuracy_artifact_root: Path,
+    sana_model_parallel,
 ) -> None:
     del _hardware
     from vllm.config import LoadConfig

tests/e2e/accuracy/sana_video/test_sana_video_pipeline_alignment.py

@@ -189,6 +189,7 @@ def _cfg_prediction(
 def test_sana_video_quick_pipeline_stages_match_diffusers(
     variant: str,
     _hardware,
+    sana_model_parallel,
 ) -> None:
     """Compare the first denoising step without replacing native attention.
 

tests/e2e/online_serving/test_flux_kontext_expansion.py

@@ -23,7 +23,8 @@ BASE_FEATURE_MARKS = hardware_marks(res={"cuda": "L4"})
 def _get_diffusion_feature_cases(model: str):
     return [
         pytest.param(
-            OmniServerParams(model=model),
+            # The unsharded transformer alone exceeds the L4's 22 GiB.
+            OmniServerParams(model=model, server_args=["--enable-cpu-offload"]),
             id="base",
             marks=BASE_FEATURE_MARKS,
         ),

tests/e2e/online_serving/test_moss_tts_expansion.py

@@ -28,7 +28,7 @@ import pytest
 
 from tests.helpers.mark import hardware_test
 from tests.helpers.runtime import OmniServerParams
-from tests.helpers.stage_config import get_deploy_config_path
+from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
 
 pytestmark = [
     pytest.mark.advanced_model,
@@ -84,7 +84,12 @@ tts_server_params = [
     pytest.param(
         OmniServerParams(
             model=MODEL,
-            stage_config_path=get_deploy_config_path("moss_tts_realtime.yaml"),
+            stage_config_path=modify_stage_config(
+                get_deploy_config_path("moss_tts_realtime.yaml"),
+                # Leave room on the single L4 for the codec weights, reference
+                # encoder, and speech-serving activations in addition to KV.
+                updates={"stages": {0: {"gpu_memory_utilization": 0.45}}},
+            ),
             server_args=["--disable-log-stats"],
         ),
         id="moss_tts_realtime",