[Refactor] Drop direct dependency on librosa (#39079)

Signed-off-by: Nick Cao <ncao@redhat.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Nick Cao
2026-04-18 06:55:38 +00:00
committed by GitHub
co-authored by Claude
parent 87518c3027
commit 153ba7f0f3
13 changed files with 40 additions and 44 deletions
@@ -13,7 +13,6 @@ import io
import time
from statistics import mean, median
import librosa
import pytest
import soundfile
import torch
@@ -21,6 +20,7 @@ from datasets import load_dataset
from evaluate import load
from transformers.models.whisper.english_normalizer import EnglishTextNormalizer
from vllm.multimodal.audio import get_audio_duration
from vllm.tokenizers import get_tokenizer
from ....models.registry import HF_EXAMPLE_MODELS
@@ -84,7 +84,7 @@ async def process_dataset(model, client, data, concurrent_request):
trust_remote_code=model_info.trust_remote_code,
)
# Warmup call as the first `librosa.load` server-side is quite slow.
# Warmup call as the first `load_audio` server-side is quite slow.
audio, sr = data[0]["audio"]["array"], data[0]["audio"]["sampling_rate"]
_ = await bound_transcribe(sem, client, tokenizer, (audio, sr), "")
@@ -118,7 +118,7 @@ def print_performance_metrics(results, total_time):
def add_duration(sample):
y, sr = sample["audio"]["array"], sample["audio"]["sampling_rate"]
sample["duration_ms"] = librosa.get_duration(y=y, sr=sr) * 1000
sample["duration_ms"] = get_audio_duration(y=y, sr=sr) * 1000
return sample
@@ -5,7 +5,6 @@ import asyncio
import json
import warnings
import librosa
import numpy as np
import pybase64 as base64
import pytest
@@ -14,6 +13,7 @@ import websockets
from tests.entrypoints.openai.conftest import add_attention_backend
from tests.utils import ROCM_ENV_OVERRIDES, ROCM_EXTRA_ARGS, RemoteOpenAIServer
from vllm.assets.audio import AudioAsset
from vllm.multimodal.media.audio import load_audio
# Increase engine iteration timeout for ROCm where first-use JIT compilation
# can exceed the default 60s, causing a silent deadlock in feed_tokens.
@@ -56,7 +56,7 @@ async def send_event(ws, event: dict) -> None:
def mary_had_lamb_audio_chunks() -> list[str]:
"""Audio split into ~1 second chunks for streaming."""
path = AudioAsset("mary_had_lamb").get_local_path()
audio, _ = librosa.load(str(path), sr=16000, mono=True)
audio, _ = load_audio(str(path), sr=16000, mono=True)
# Split into ~0.1 second chunks (1600 samples at 16kHz)
chunk_size = 1600
@@ -6,7 +6,6 @@ import asyncio
import io
import json
import librosa
import numpy as np
import openai
import pytest
@@ -14,6 +13,7 @@ import pytest_asyncio
import soundfile as sf
from tests.utils import RemoteOpenAIServer
from vllm.multimodal.media.audio import load_audio
from vllm.platforms import current_platform
MODEL_NAME = "openai/whisper-large-v3-turbo"
@@ -134,7 +134,7 @@ async def test_bad_requests(mary_had_lamb, whisper_client):
@pytest.mark.asyncio
async def test_long_audio_request(mary_had_lamb, whisper_client):
mary_had_lamb.seek(0)
audio, sr = librosa.load(mary_had_lamb)
audio, sr = load_audio(mary_had_lamb)
# Add small silence after each audio for repeatability in the split process
audio = np.pad(audio, (0, 1600))
repeated_audio = np.tile(audio, 10)
@@ -7,7 +7,6 @@ import io
import json
import httpx
import librosa
import numpy as np
import openai
import pytest
@@ -17,6 +16,7 @@ import soundfile as sf
from tests.entrypoints.openai.conftest import add_attention_backend
from tests.utils import RemoteOpenAIServer
from vllm.logger import init_logger
from vllm.multimodal.media.audio import load_audio
logger = init_logger(__name__)
@@ -264,7 +264,7 @@ async def test_long_audio_request(foscolo, client_and_model):
if model_name == "google/gemma-3n-E2B-it":
pytest.skip("Gemma3n does not support long audio requests")
foscolo.seek(0)
audio, sr = librosa.load(foscolo)
audio, sr = load_audio(foscolo)
repeated_audio = np.tile(audio, 2)
# Repeated audio to buffer
buffer = io.BytesIO()