[OmniVoice] Add two-stage TTS serving support (#2463)

Signed-off-by: linyueqian <linyueqian@outlook.com>
This commit is contained in:
Yueqian Lin authored and GitHub committed 2026-04-03 20:41:20 +00:00
1 parent 10db95f9a9
commit 0e83ebe1d4
24 files changed
+2737 -1

No files matched your search

@@ -0,0 +1,84 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
E2E offline tests for OmniVoice TTS model with text input and audio output.
Uses GPUGenerationWorker for both stages (iterative unmasking + DAC decoder).
"""
import os
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
from pathlib import Path
import numpy as np
import pytest
from tests.utils import hardware_test
MODEL = "k2-fsa/OmniVoice"
def get_stage_config():
return str(
Path(__file__).parent.parent.parent.parent / "vllm_omni" / "model_executor" / "stage_configs" / "omnivoice.yaml"
)
@pytest.mark.advanced_model
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}, num_cards=1)
def test_omnivoice_text_to_audio() -> None:
"""
Test OmniVoice text-to-audio generation via offline Omni runner.
Deploy Setting: omnivoice.yaml (enforce_eager=true)
Input Modal: text
Output Modal: audio
"""
from vllm_omni.entrypoints.omni import Omni
omni = Omni(
model=MODEL,
stage_configs_path=get_stage_config(),
trust_remote_code=True,
log_stats=True,
)
try:
prompts = {"prompt": "Hello, this is a test for text to audio."}
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
sampling_params_list = [OmniDiffusionSamplingParams()]
outputs = list(omni.generate(prompts, sampling_params_list=sampling_params_list))
assert len(outputs) > 0, "No outputs generated"
# Check final output has audio
final_output = outputs[-1]
ro = final_output.request_output
assert ro is not None, "No request_output"
mm = getattr(ro, "multimodal_output", None)
if not mm and ro.outputs:
mm = getattr(ro.outputs[0], "multimodal_output", None)
assert mm is not None, "No multimodal_output"
assert "audio" in mm, f"No 'audio' key in multimodal_output: {mm.keys()}"
audio = mm["audio"]
if isinstance(audio, np.ndarray):
audio_np = audio
else:
audio_np = audio.cpu().numpy().squeeze()
assert audio_np.size > 0, "Audio output is empty"
rms = np.sqrt(np.mean(audio_np**2))
assert rms > 0.01, f"Audio RMS too low ({rms:.4f}), likely silence"
print(f"Generated audio: {len(audio_np) / 24000:.2f}s, rms={rms:.4f}")
finally:
omni.close()
@@ -0,0 +1,84 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
E2E Online tests for OmniVoice TTS model via /v1/audio/speech endpoint.
Tests verify that the OmniVoice model generates valid audio when
accessed through the standard OpenAI-compatible speech API.
"""
import os
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
os.environ["VLLM_TEST_CLEAN_GPU_MEMORY"] = "0"
from pathlib import Path
import httpx
import pytest
from tests.conftest import OmniServerParams
from tests.utils import hardware_test
MODEL = "k2-fsa/OmniVoice"
STAGE_CONFIG = str(
Path(__file__).parent.parent.parent.parent / "vllm_omni" / "model_executor" / "stage_configs" / "omnivoice.yaml"
)
EXTRA_ARGS = [
"--trust-remote-code",
"--disable-log-stats",
]
TEST_PARAMS = [
OmniServerParams(
model=MODEL,
stage_config_path=STAGE_CONFIG,
server_args=EXTRA_ARGS,
)
]
MIN_AUDIO_BYTES = 5000
def make_speech_request(
host: str,
port: int,
text: str,
timeout: float = 180.0,
) -> httpx.Response:
"""Make a request to the /v1/audio/speech endpoint for OmniVoice."""
url = f"http://{host}:{port}/v1/audio/speech"
payload = {"input": text}
with httpx.Client(timeout=timeout) as client:
return client.post(url, json=payload)
def verify_wav_audio(content: bytes) -> bool:
"""Verify that content is valid WAV audio data."""
if len(content) < 44:
return False
return content[:4] == b"RIFF" and content[8:12] == b"WAVE"
@pytest.mark.parametrize("omni_server", TEST_PARAMS, indirect=True)
class TestOmniVoiceTTS:
"""E2E tests for OmniVoice TTS model."""
@pytest.mark.core_model
@pytest.mark.omni
@hardware_test(res={"cuda": "L4"}, num_cards=1)
def test_speech_auto_voice(self, omni_server) -> None:
"""Test auto voice TTS generation (text only, no reference audio)."""
response = make_speech_request(
host=omni_server.host,
port=omni_server.port,
text="Hello, this is a test of the OmniVoice text to speech system.",
)
assert response.status_code == 200, f"Request failed: {response.text}"
assert response.headers.get("content-type") == "audio/wav"
assert verify_wav_audio(response.content), "Response is not valid WAV audio"
assert len(response.content) > MIN_AUDIO_BYTES, (
f"Audio too small ({len(response.content)} bytes), expected > {MIN_AUDIO_BYTES}"
)