[CI/Build][MiniCPM-o] Add MiniCPM-o 4.5 accuracy and performance nightly coverage (#5710)

Signed-off-by: psv666 <2693925048@qq.com>
Signed-off-by: psv666 <150513104+psv666@users.noreply.github.com>
Signed-off-by: amy-why-3459 <wuhaiyan17@huawei.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>

tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py

@@ -39,10 +39,7 @@ _MAX_SEED_TTS_MEAN_WER = 0.05
 # Match the validated Daily-Omni client body from daily_omni_bench.sh.
 _DAILY_EXTRA_BODY = {
     "modalities": ["text"],
-    "chat_template_kwargs": {
-        "enable_thinking": False,
-        "use_tts_template": True,
-    },
+    "chat_template_kwargs": {"enable_thinking": False},
 }
 _SEED_EXTRA_BODY = {
     "modalities": ["text", "audio"],
@@ -151,6 +148,7 @@ def test_minicpmo_4_5_seed_tts_wer_bench(omni_server) -> None:
         omni_server,
         skip_seed=False,
         skip_daily=True,
+        max_concurrency=4,
     )
     argv.extend(
         [

tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py

@@ -963,12 +963,23 @@ async def _send_clean_turn(
         if outcome == "listen":
             return None, "listen"
     else:
-        await _wait_for(
-            state,
-            lambda: state.count("response.created") > before_created,
-            timeout_s=timeout_s,
-            label=f"{transcript} response.created",
-        )
+        try:
+            await _wait_for(
+                state,
+                lambda: state.count("response.created") > before_created,
+                timeout_s=timeout_s,
+                label=f"{transcript} response.created",
+            )
+        except TimeoutError:
+            # ``auto_response`` sessions only get ``response.created`` once the model
+            # decides to speak, so report the listen decisions that took its place.
+            raise TimeoutError(
+                f"Timed out waiting for {transcript} response.created after "
+                f"{state.count('input_audio_buffer.committed')} committed input turns; "
+                f"model_listen={state.model_listen_count - before_model_listen} "
+                f"buffering_listen={state.buffering_listen_count} "
+                f"errors={state.count('error')}"
+            ) from None
         await _wait_for(
             state,
             lambda: eligible_response_id() is not None,

tests/e2e/online_serving/test_minicpmo_4_5_duplex.py

@@ -133,6 +133,10 @@ def test_duplex_single_session_response_required(omni_server, model_prefix: str,
         output_dir=tmp_path / "single_session",
     )
     args.turns = 2
+    # Every turn replays the same active-speech window as the first one. The default
+    # shorter follow-up window is a different mid-utterance slice, which the native
+    # duplex model may legitimately answer with "listen" instead of a response.
+    args.turn_duration_ms = [args.first_turn_ms] * args.turns
     result = asyncio.run(run_demo(args))
     assert result["ok"] is True
     assert result["audio_delta_count"] > 0

tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py

@@ -53,7 +53,11 @@ def test_duplex_admission_and_expiry_reaper(omni_server, model_prefix: str, tmp_
     assert result["admission"]["overflow_error_code"] == "resource_exhausted"
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+# CUDA-only for now: this contract asserts the model speaks while the user is
+# still talking, which only holds when the duplex pipeline sustains real-time
+# throughput. The current NPU stack runs several times slower than real time,
+# so it never reaches a mid-stream decision point.
+@hardware_test(res={"cuda": "H100"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
 def test_duplex_soft_interrupt(omni_server, model_prefix: str, tmp_path: Path) -> None:
     input_wav = validated_soft_interrupt_wav()