[Bugfix] Drop librosa reintroduced by Step-Audio2 (#6467)

Signed-off-by: Nick Cao <ncao@redhat.com>

vllm_omni/model_executor/models/step_audio2/step_audio2_thinker.py

@@ -13,7 +13,6 @@ from __future__ import annotations
 from collections.abc import Iterable, Mapping, Sequence
 from typing import TYPE_CHECKING, Any, TypedDict
 
-import librosa  # noqa: TID251
 import numpy as np
 import torch
 import torch.nn as nn
@@ -52,6 +51,7 @@ from vllm_omni.model_executor.models.step_audio2.step_audio2_constants import (
     DEFAULT_TOKEN_CONFIG,
     STEP_AUDIO2_AUDIO_PATCH_TOKEN_ID,
 )
+from vllm_omni.utils.audio import mel_filter_bank
 
 if TYPE_CHECKING:
     pass
@@ -65,7 +65,7 @@ if TYPE_CHECKING:
 def _mel_filters(n_mels: int) -> torch.Tensor:
     """Generate mel filter banks"""
     assert n_mels in {80, 128}, f"Unsupported n_mels: {n_mels}"
-    return torch.from_numpy(librosa.filters.mel(sr=16000, n_fft=400, n_mels=n_mels))
+    return mel_filter_bank(sr=16000, n_fft=400, n_mels=n_mels)
 
 
 def _normalize_audio(audio: Any) -> torch.Tensor:

vllm_omni/model_executor/models/step_audio2/step_audio2_token2wav.py

@@ -1,14 +1,15 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import io
 import os
 from collections.abc import Iterable
 from dataclasses import dataclass, field
 from typing import Any
 
-import librosa  # noqa: TID251
 import numpy as np
 import onnxruntime
 import s3tokenizer
-import soundfile as sf
 import torch
 import torch.nn as nn
 import torchaudio
@@ -23,6 +24,8 @@ from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.model_executor.models.interfaces import SupportsPP
 from vllm.model_executor.models.utils import WeightsMapper
+from vllm.multimodal.audio import AudioResampler
+from vllm.multimodal.media.audio import load_audio
 from vllm.sequence import IntermediateTensors
 from vllm.v1.outputs import SamplerOutput
 from vllm.v1.sample.metadata import SamplingMetadata
@@ -231,19 +234,11 @@ class StepAudio2Token2WavCore(nn.Module):
 
     def _prepare_prompt(self, prompt_wav: str):
         """Prepare prompt audio for conditioning"""
-        # Prefer soundfile/librosa path to avoid torchaudio->torchcodec runtime coupling.
-        try:
-            audio_np_raw, sample_rate = sf.read(prompt_wav, dtype="float32", always_2d=False)
-        except Exception:
-            audio_np_raw, sample_rate = librosa.load(prompt_wav, sr=None, mono=True)
-        if isinstance(audio_np_raw, np.ndarray) and audio_np_raw.ndim > 1:
-            audio_np_raw = np.mean(audio_np_raw, axis=-1)
+        audio_np_raw, sample_rate = load_audio(prompt_wav, sr=None, mono=True)
         sample_rate = int(sample_rate)
 
         # 16 kHz branch: s3tokenizer + speaker embedding
-        audio_16k = audio_np_raw.astype(np.float32)
-        if sample_rate != 16000:
-            audio_16k = librosa.resample(y=audio_16k, orig_sr=sample_rate, target_sr=16000)
+        audio_16k = AudioResampler(target_sr=16000).resample(audio_np_raw.astype(np.float32), orig_sr=sample_rate)
         audio = torch.from_numpy(audio_16k)
         mels = s3tokenizer.log_mel_spectrogram(audio)
         mels, mels_lens = s3tokenizer.padding([mels])
@@ -257,9 +252,7 @@ class StepAudio2Token2WavCore(nn.Module):
 
         # 24 kHz branch: mel spectrogram for flow model conditioning
         # Must resample from the ORIGINAL audio, not the 16 kHz version.
-        audio_24k = audio_np_raw.astype(np.float32)
-        if sample_rate != 24000:
-            audio_24k = librosa.resample(y=audio_24k, orig_sr=sample_rate, target_sr=24000)
+        audio_24k = AudioResampler(target_sr=24000).resample(audio_np_raw.astype(np.float32), orig_sr=sample_rate)
         audio_24k_t = torch.from_numpy(audio_24k).unsqueeze(0)  # [1, T]
         prompt_mel = (
             mel_spectrogram(