[Bugfix] Fix MiniCPM-o 4.5 audio placeholder pool step (#5116)

Signed-off-by: a-kaa <siruhe666@gmail.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>

tests/model_executor/models/minicpmo_4_5/test_audio_pool_step.py

@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Regression tests for MiniCPM-o 4.5 audio placeholder pooling."""
+
+from __future__ import annotations
+
+import math
+from types import SimpleNamespace
+
+import pytest
+
+from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_llm import (
+    MiniCPMO45OmniLLMProcessingInfo,
+    MiniCPMOConfig,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+
+class _FakeProcessor:
+    """Minimal processor surface used by ``get_audio_placeholder``."""
+
+    def __init__(self) -> None:
+        # Reproduce the cached vLLM processor's legacy initial state.
+        self.pool_step = 2
+        self.audio_processor = SimpleNamespace(hop_length=160)
+        self.image_processor = SimpleNamespace(mean=[0.0], std=[1.0])
+
+    def get_audio_placeholder(
+        self,
+        audio_lens: int,
+        chunk_input: bool = True,
+        chunk_length: int = 1,
+    ) -> str:
+        del chunk_input, chunk_length
+        feature_lens = math.ceil(audio_lens / self.audio_processor.hop_length)
+        cnn_feature_lens = (feature_lens - 1) // 2 + 1
+        output_lens = (cnn_feature_lens - self.pool_step) // self.pool_step + 1
+        return "<unk>" * output_lens
+
+
+def _processing_info(pool_step: int, processor: _FakeProcessor) -> MiniCPMO45OmniLLMProcessingInfo:
+    info = object.__new__(MiniCPMO45OmniLLMProcessingInfo)
+    info.ctx = SimpleNamespace(
+        tokenizer=object(),
+        get_hf_config=lambda: SimpleNamespace(audio_pool_step=pool_step),
+        get_hf_processor=lambda **kwargs: processor,
+    )
+    return info
+
+
+@pytest.mark.parametrize(
+    ("configured_pool_step", "expected_placeholders"),
+    [(5, 50), (2, 125)],
+)
+def test_audio_placeholders_follow_checkpoint_pool_step(
+    configured_pool_step: int,
+    expected_placeholders: int,
+) -> None:
+    """The prompt path must use the same pool step as the audio encoder."""
+    processor = _FakeProcessor()
+    info = _processing_info(configured_pool_step, processor)
+
+    placeholder = info.get_audio_placeholder(5 * 16_000)
+
+    assert processor.pool_step == configured_pool_step
+    assert placeholder.count("<unk>") == expected_placeholders
+
+
+def test_minicpmo_4_5_pool_step_defaults_to_checkpoint_value() -> None:
+    assert MiniCPMOConfig().audio_pool_step == 5

vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_llm.py

@@ -328,7 +328,7 @@ class MiniCPMOConfig(Qwen2Config):
         tts_config=None,
         use_image_id=True,
         vision_batch_size=16,
-        audio_pool_step=2,
+        audio_pool_step=5,
         audio_chunk_length=1.0,
         stream_input=False,
         init_vision=True,
@@ -2839,6 +2839,9 @@ class MiniCPMO45OmniLLMProcessingInfo(BaseProcessingInfo):
     ) -> str:
         hf_processor = self.get_hf_processor()
 
+        # Keep prompt placeholders aligned with the encoder's checkpoint config.
+        hf_processor.pool_step = self.get_default_audio_pool_step()
+
         return hf_processor.get_audio_placeholder(
             audio_lens,
             chunk_input=chunk_input,
@@ -2846,7 +2849,7 @@ class MiniCPMO45OmniLLMProcessingInfo(BaseProcessingInfo):
         )
 
     def get_default_audio_pool_step(self) -> int:
-        return getattr(self.get_hf_config(), "audio_pool_step", 2)
+        return getattr(self.get_hf_config(), "audio_pool_step", 5)
 
     def get_default_audio_sampling_rate(self) -> int:
         return 16000