Update realtime invalid-request fixtures and reserve Moss encoder memory

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

tests/dfx/reliability/invalid_param_test/test_invalid_omni_chat.py

@@ -216,44 +216,39 @@ def test_video_chat_stream_invalid_requests(
 # WS /v1/realtime
 # ─────────────────────────────────────────────────────────────────────────────
 
-# Resolved in ``test_realtime_invalid_requests`` after the initial handshake.
-_REALTIME_WS_MODEL_MISMATCH = object()
-_REALTIME_WS_INVALID_AUDIO_APPEND = object()
-
-
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
 @pytest.mark.parametrize(
     "send_frames_spec, ws_error_code, err_message",
     [
-        pytest.param("{not-json", "invalid_json", "Invalid JSON", id="invalid_json"),
+        pytest.param("{not-json", "invalid_event", "Invalid or unrecognized client event", id="invalid_json"),
         pytest.param(
             json.dumps({"type": "session.update"}),
             "invalid_event",
-            ("model", "Missing required field"),
-            id="session_update_missing_model",
+            "Invalid or unrecognized client event",
+            id="session_update_missing_session",
         ),
         pytest.param(
-            _REALTIME_WS_MODEL_MISMATCH,
-            "model_not_found",
-            ("model", "does not exist"),
-            id="session_update_model_mismatch",
+            json.dumps({"type": "session.update", "session": []}),
+            "invalid_event",
+            "Invalid or unrecognized client event",
+            id="session_update_invalid_session",
         ),
         pytest.param(
-            json.dumps({"type": "input_audio_buffer.commit", "final": False}),
-            "model_not_validated",
-            ("session.update", "validate the model"),
-            id="commit_without_model_validation",
+            json.dumps({"type": "input_audio_buffer.commit"}),
+            "invalid_request_error",
+            "Input audio buffer is empty",
+            id="commit_without_audio",
         ),
         pytest.param(
             json.dumps({"type": "unknown.event"}),
-            "unknown_event",
-            "Unknown event type",
+            "invalid_event",
+            "Invalid or unrecognized client event",
             id="unknown_event_type",
         ),
         pytest.param(
-            _REALTIME_WS_INVALID_AUDIO_APPEND,
-            "invalid_audio",
-            "Invalid audio data",
+            json.dumps({"type": "input_audio_buffer.append", "audio": "not-valid-base64!!!"}),
+            "invalid_request_error",
+            "Invalid base64 audio data",
             id="invalid_audio_append",
         ),
     ],
@@ -266,24 +261,9 @@ def test_realtime_invalid_requests(
     ws_error_code: str,
     err_message: str | tuple[str, ...],
 ) -> None:
-    send_frames: Any
-    if send_frames_spec is _REALTIME_WS_MODEL_MISMATCH:
-        send_frames = json.dumps(
-            {
-                "type": "session.update",
-                "model": "this-model-is-not-served-on-this-instance",
-            }
-        )
-    elif send_frames_spec is _REALTIME_WS_INVALID_AUDIO_APPEND:
-        send_frames = [
-            json.dumps({"type": "session.update", "model": omni_server.model}),
-            json.dumps({"type": "input_audio_buffer.append", "audio": "not-valid-base64!!!"}),
-        ]
-    else:
-        send_frames = send_frames_spec
     online_client.send_realtime_ws_request(
         {
-            "send_frames": send_frames,
+            "send_frames": send_frames_spec,
             "timeout": 120,
             "ws_max_size": None,
             "ws_skip_types": ["session.created", "conversation.created"],

tests/e2e/online_serving/test_moss_tts_expansion.py

@@ -88,7 +88,7 @@ tts_server_params = [
                 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}}},
+                updates={"stages": {0: {"gpu_memory_utilization": 0.35}}},
             ),
             server_args=["--disable-log-stats"],
         ),