Merge origin/main into codex-vae-spatial-decode-auto

Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
Hongsheng Liu committed 2026-08-13 23:49:11 +08:00
commit 089851e55a
244 files changed
+16669 -2175

No files matched your search

@@ -2156,12 +2156,12 @@ def test_direct_listen_decision_survives_inner_completion_metadata():
)
assert decision is not None
output = attach_duplex_output_decision(
OmniRequestOutput(
OmniRequestOutput.from_stage_output(
inner_output,
request_id="duplex-direct-listen",
finished=True,
stage_id=0,
final_output_type=decision.final_output_type,
request_output=inner_output,
metrics={
"stage_metrics": {
"0": {
@@ -3049,12 +3049,12 @@ async def test_minicpmo_auto_response_listen_without_response_does_not_defer_com
)
assert decision is not None
listen_output = attach_duplex_output_decision(
OmniRequestOutput(
OmniRequestOutput.from_stage_output(
inner_output,
request_id=request_id,
finished=True,
stage_id=0,
final_output_type=decision.final_output_type,
request_output=inner_output,
),
decision,
)
@@ -119,7 +119,6 @@ class MockGenerationResult:
def __init__(self, images):
self.images = images
self.request_output = SimpleNamespace(images=images)
self.stage_durations = {}
self.peak_memory_mb = 0.0
@@ -137,10 +136,7 @@ class MockStageResult:
outputs = [SimpleNamespace(text=text, index=0)]
else:
outputs = []
self.request_output = SimpleNamespace(
outputs=outputs,
images=self.images,
)
self.outputs = outputs
self.stage_durations = {}
self.peak_memory_mb = 0.0
@@ -2080,22 +2076,20 @@ def test_extract_images_from_result():
assert all(isinstance(img, Image.Image) for img in images)
assert all(img.size == (64, 64) for img in images)
# Test dict path: result.request_output["images"]
# Test result with "images" attribute set in __init__
class DictRequestOutput:
def __init__(self):
self.request_output = {"images": [np.random.randint(0, 255, (64, 64, 3), dtype=np.uint8)]}
self.images = [np.random.randint(0, 255, (64, 64, 3), dtype=np.uint8)]
result = DictRequestOutput()
images = _extract_images_from_result(result)
assert len(images) == 1
assert isinstance(images[0], Image.Image)
# Test attribute path: result.request_output.images
# Test result with "images" attribute from an inner object
class AttrRequestOutput:
def __init__(self):
self.request_output = type(
"obj", (), {"images": [np.random.randint(0, 255, (32, 32, 3), dtype=np.uint8)]}
)()
self.images = [np.random.randint(0, 255, (32, 32, 3), dtype=np.uint8)]
result = AttrRequestOutput()
images = _extract_images_from_result(result)
@@ -180,10 +180,10 @@ def create_mock_audio_output_for_test(
audio_tensor = torch.sin(torch.linspace(0, 440 * 2 * torch.pi, num_samples))
mock_request_output = MockRequestOutput(request_id=request_id, audio_tensor=audio_tensor)
return OmniRequestOutput(
return OmniRequestOutput.from_stage_output(
mock_request_output,
stage_id=0,
final_output_type="audio",
request_output=mock_request_output,
)
@@ -680,7 +680,7 @@ class TestSpeechAPI:
@pytest.mark.asyncio
async def test_create_diffusion_speech_extra_params(self, mocker: MockerFixture):
"""Test that extra_params are correctly applied to sampling_params_list in diffusion mode."""
"""Test public diffusion speech success and extra_params propagation."""
# Mock the engine client
mock_engine = mocker.MagicMock()
@@ -704,7 +704,11 @@ class TestSpeechAPI:
req = OpenAICreateSpeechRequest(input="Hello", extra_params={"new_arg": 123, "existing_arg": "new_value"})
await server._create_diffusion_speech(req)
response = await server.create_speech(req)
assert response.status_code == 200
assert response.media_type == "audio/wav"
assert response.body == b"dummy"
# Verify generate was called
mock_engine.generate.assert_called_once()
@@ -737,6 +741,19 @@ class TestTTSMethods:
yield server
server.shutdown()
def test_diffusion_audio_encode_speed_skips_ar_adapter(self, mocker: MockerFixture):
"""Pure-diffusion speech keeps generic speed handling outside AR adapters."""
server = OmniOpenAIServingSpeech.for_diffusion(
diffusion_engine=mocker.MagicMock(),
model_name="test-model",
)
resolve_adapter = mocker.patch.object(serving_speech_module, "resolve_adapter")
speed = server._audio_encode_speed(OpenAICreateSpeechRequest(input="Hello", speed=1.25))
assert speed == 1.25
resolve_adapter.assert_not_called()
def test_is_tts_detection_no_stage(self, speech_server):
"""Test TTS model detection when no TTS stage exists."""
# Fixture creates server with stage_configs = [] -> _is_tts should be False
@@ -2292,10 +2309,10 @@ class TestStreamingResponse:
self.prompt_logprobs = None
self.kv_transfer_params = None
return OmniRequestOutput(
return OmniRequestOutput.from_stage_output(
MockRequestOutput(audio_tensor=chunk),
stage_id=0,
final_output_type="audio",
request_output=MockRequestOutput(audio_tensor=chunk),
finished=finished,
)
@@ -3441,10 +3458,10 @@ class TestWAVStreaming:
self.prompt_logprobs = None
self.kv_transfer_params = None
return OmniRequestOutput(
return OmniRequestOutput.from_stage_output(
MockRequestOutput(audio_tensor=chunk),
stage_id=0,
final_output_type="audio",
request_output=MockRequestOutput(audio_tensor=chunk),
finished=finished,
)
@@ -3,6 +3,7 @@ import base64
from types import SimpleNamespace
import pytest
import torch
from fastapi import FastAPI, WebSocket
from pytest_mock import MockerFixture
from starlette.testclient import TestClient
@@ -11,11 +12,24 @@ from starlette.websockets import WebSocketDisconnect
from vllm_omni.entrypoints.openai import serving_speech_stream as streaming_speech_module
from vllm_omni.entrypoints.openai.serving_speech import OmniOpenAIServingSpeech
from vllm_omni.entrypoints.openai.serving_speech_stream import OmniStreamingSpeechHandler
from vllm_omni.utils.forced_aligner import WordTimestamp
from vllm_omni.model_executor.stage_input_processors.forced_aligner import ALIGNER_WORDS_KEY
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
def _fake_aligner_res(pairs, words):
"""Build a stand-in for the forced-aligner stage's pooling output.
Mirrors what rides the generator in production: ``res.outputs.data`` is a
``[n_words, 2]`` int32 tensor of ``[start_ms, end_ms]`` and the word strings
travel in ``additional_information``. Decoded by ``extract_word_timestamps``.
"""
return SimpleNamespace(
outputs=SimpleNamespace(data=torch.tensor(pairs, dtype=torch.int32), additional_information=None),
additional_information={ALIGNER_WORDS_KEY: list(words)},
)
def _build_test_app(
speech_service=None,
*,
@@ -28,9 +42,9 @@ def _build_test_app(
speech_service = mocker.MagicMock(spec=OmniOpenAIServingSpeech)
speech_service._generate_audio_bytes = mocker.AsyncMock(return_value=(b"RIFF" + b"\x00" * 32, "audio/wav"))
speech_service._prepare_speech_generation = mocker.AsyncMock(return_value=("req-1", object(), {}))
speech_service.forced_aligner_config = None
speech_service.forced_aligner_enabled = False
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False):
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False, collect=None):
for chunk in (b"\x01\x02", b"\x03\x04\x05"):
yield (chunk, 24000) if include_sample_rate else chunk
@@ -235,7 +249,7 @@ class TestStreamingSpeechWebSocket:
speech_service._generate_audio_bytes = mocker.AsyncMock(return_value=(b"", "audio/wav"))
speech_service.engine_client = mocker.MagicMock()
speech_service.engine_client.abort = mocker.AsyncMock()
speech_service.forced_aligner_config = None
speech_service.forced_aligner_enabled = False
async def mock_prepare_speech_generation(request):
captured_requests.append(request)
@@ -311,13 +325,13 @@ class TestStreamingSpeechWebSocket:
assert error["type"] == "error"
assert "without --forced-aligner" in error["message"]
def test_word_timestamps_emit_sidecar_json_frame(self, mocker: MockerFixture):
def test_word_timestamps_emit_pipeline_json_frame(self, mocker: MockerFixture):
captured_requests = []
speech_service = mocker.MagicMock(spec=OmniOpenAIServingSpeech)
speech_service._generate_audio_bytes = mocker.AsyncMock(return_value=(b"", "audio/wav"))
speech_service.engine_client = mocker.MagicMock()
speech_service.engine_client.abort = mocker.AsyncMock()
speech_service.forced_aligner_config = SimpleNamespace(model="aligner")
speech_service.forced_aligner_enabled = True
async def mock_prepare_speech_generation(request):
captured_requests.append(request)
@@ -328,19 +342,15 @@ class TestStreamingSpeechWebSocket:
first_chunk = b"\x01" * 1000
second_chunk = b"\x02" * 1000
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False):
# The forced-aligner stage rides the same generator: its pooling output
# is surfaced via the ``collect`` channel once the audio has streamed.
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False, collect=None):
for chunk in (first_chunk, second_chunk):
yield (chunk, 1000) if include_sample_rate else chunk
if collect is not None:
collect["aligner_res"] = _fake_aligner_res([[0, 200], [200, 900]], ["Hello", "world"])
speech_service._generate_pcm_chunks = mock_generate_pcm_chunks
# Sentence-level: aligner runs once over the whole sentence audio.
mock_align = mocker.AsyncMock(
return_value=[
WordTimestamp("Hello", 0, 200),
WordTimestamp("world", 200, 900),
]
)
mocker.patch.object(streaming_speech_module, "forced_align", mock_align)
app, _ = _build_test_app(speech_service)
with TestClient(app) as client:
@@ -406,7 +416,6 @@ class TestStreamingSpeechWebSocket:
}
assert captured_requests[0].word_timestamps is True
assert mock_align.await_count == 1
def test_word_timestamps_emit_word_dicts(self, mocker: MockerFixture):
# The streaming layer forwards the aligner's (already monotonic,
@@ -415,21 +424,16 @@ class TestStreamingSpeechWebSocket:
speech_service._generate_audio_bytes = mocker.AsyncMock(return_value=(b"", "audio/wav"))
speech_service.engine_client = mocker.MagicMock()
speech_service.engine_client.abort = mocker.AsyncMock()
speech_service.forced_aligner_config = SimpleNamespace(model="aligner")
speech_service.forced_aligner_enabled = True
speech_service._prepare_speech_generation = mocker.AsyncMock(return_value=("req", object(), {}))
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False):
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False, collect=None):
chunk = b"\x01" * 1000
yield (chunk, 1000) if include_sample_rate else chunk
if collect is not None:
collect["aligner_res"] = _fake_aligner_res([[0, 1000], [1000, 1200]], ["Hello", "world"])
speech_service._generate_pcm_chunks = mock_generate_pcm_chunks
mock_align = mocker.AsyncMock(
return_value=[
WordTimestamp("Hello", 0, 1000),
WordTimestamp("world", 1000, 1200),
]
)
mocker.patch.object(streaming_speech_module, "forced_align", mock_align)
app, _ = _build_test_app(speech_service)
with TestClient(app) as client:
@@ -572,7 +576,7 @@ class TestStreamingSpeechWebSocket:
speech_service._generate_pcm_chunks = mocker.AsyncMock()
speech_service.engine_client = mocker.MagicMock()
speech_service.engine_client.abort = mocker.AsyncMock()
speech_service.forced_aligner_config = None
speech_service.forced_aligner_enabled = False
app, _ = _build_test_app(speech_service)
with TestClient(app) as client:
@@ -602,7 +606,7 @@ class TestStreamingSpeechWebSocket:
speech_service._prepare_speech_generation = mocker.AsyncMock(return_value=("req-stream-err", object(), {}))
speech_service.engine_client = mocker.MagicMock()
speech_service.engine_client.abort = mocker.AsyncMock()
speech_service.forced_aligner_config = None
speech_service.forced_aligner_enabled = False
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False):
yield b"\x01\x02"
@@ -696,7 +700,7 @@ class TestStreamingSpeechWebSocket:
speech_service._prepare_speech_generation = mocker.AsyncMock(return_value=("req-abort", object(), {}))
speech_service.engine_client = mocker.MagicMock()
speech_service.engine_client.abort = mocker.AsyncMock()
speech_service.forced_aligner_config = None
speech_service.forced_aligner_enabled = False
async def mock_generate_pcm_chunks(_generator, _request_id, *, include_sample_rate=False):
yield b"\x01\x02"
@@ -47,7 +47,7 @@ def _text_result(text: str) -> OmniRequestOutput:
output.text = text
request_output = RequestOutput()
request_output.outputs = [output]
return OmniRequestOutput(final_output_type="text", request_output=request_output)
return OmniRequestOutput.from_stage_output(request_output, final_output_type="text")
def _audio_result(audio_data: Any) -> OmniRequestOutput:
@@ -61,7 +61,7 @@ def _audio_result(audio_data: Any) -> OmniRequestOutput:
output.multimodal_output = {"audio": audio_data}
request_output = RequestOutput()
request_output.outputs = [output]
return OmniRequestOutput(final_output_type="audio", request_output=request_output)
return OmniRequestOutput.from_stage_output(request_output, final_output_type="audio")
class MockWebSocket:
+1 -1
View File
@@ -364,7 +364,7 @@ async def test_omni_generate_request_id():
Text modality only; asserts caller-visible ids are preserved across the
multi-stage orchestrator path on H100.
"""
engine = AsyncOmni(model=OMNI_MODEL, stage_configs_path=OMNI_STAGE_CONFIG)
engine = AsyncOmni(model=OMNI_MODEL, deploy_config=OMNI_STAGE_CONFIG)
try:
for req_id in _OMNI_REQ_IDS:
async for output in engine.generate(
@@ -478,12 +478,11 @@ def test_resolve_stage_configs_injects_additional_config_into_diffusion_stage(mo
)
engine = AsyncOmniEngine.__new__(AsyncOmniEngine)
engine._strip_single_engine_args = lambda kwargs: kwargs
_, stage_configs = engine._resolve_stage_configs(
"dummy-model",
{
"stage_configs_path": "dummy.yaml",
"deploy_config": "dummy.yaml",
"additional_config": {"torchair_graph_config": {"enabled": True}},
},
trust_remote_code=False,
@@ -516,12 +515,11 @@ def test_resolve_stage_configs_injects_quantization_config_into_diffusion_stage(
)
engine = AsyncOmniEngine.__new__(AsyncOmniEngine)
engine._strip_single_engine_args = lambda kwargs: kwargs
_, stage_configs = engine._resolve_stage_configs(
"dummy-model",
{
"stage_configs_path": "dummy.yaml",
"deploy_config": "dummy.yaml",
"quantization_config": {"method": "bitsandbytes"},
},
trust_remote_code=False,
+4 -6
View File
@@ -480,11 +480,9 @@ def test_async_omni_duplex_direct_output_prefers_outer_control_metadata():
OmniRequestOutput(
request_id="duplex-sid-e0-stage0",
stage_id=0,
request_output=SimpleNamespace(
outputs=[],
multimodal_output=SimpleNamespace(kind="processed-payload"),
_custom_output={"inner": "completion"},
),
outputs=[],
_multimodal_output={"kind": "processed-payload"},
_custom_output={"inner": "completion"},
),
decision,
)
@@ -557,7 +555,7 @@ async def test_async_omni_duplex_collect_wraps_raw_response_stage_output():
assert outputs[0].finished is False
assert outputs[0].stage_id == 1
assert outputs[0].final_output_type == "audio"
assert outputs[0].request_output is raw_stage1_output
assert outputs[0].peak_memory_mb == raw_stage1_output.peak_memory_mb
def test_async_omni_duplex_request_info_includes_response_stage():
+9 -10
View File
@@ -72,8 +72,8 @@ def make_output_msg(
final_output_type=final_output_type,
images=images or [],
stage_durations={},
outputs=[SimpleNamespace(text=payload, index=0)],
)
engine_output.payload = payload
return OutputMessage(
request_id=request_id,
stage_id=stage_id,
@@ -474,7 +474,7 @@ async def test_async_omni_yields_only_final_stage_outputs(monkeypatch: pytest.Mo
app.shutdown()
assert [output.stage_id for output in outputs] == [2]
assert [output.request_output.payload for output in outputs] == ["final"]
assert [output.outputs[0].text for output in outputs] == ["final"]
assert "req-1" not in app.request_states
@@ -492,7 +492,7 @@ async def test_async_omni_accepts_multiple_final_stage_streams(monkeypatch: pyte
app.shutdown()
assert [output.stage_id for output in outputs] == [0, 0, 0, 2, 2, 2]
assert [output.request_output.payload for output in outputs] == [
assert [output.outputs[0].text for output in outputs] == [
"req-1-stage0-0",
"req-1-stage0-1",
"req-1-stage0-2",
@@ -517,7 +517,7 @@ async def test_async_omni_stops_on_final_stage_finished(monkeypatch: pytest.Monk
finally:
app.shutdown()
assert [output.request_output.payload for output in outputs] == [
assert [output.outputs[0].text for output in outputs] == [
"req-1-stage0",
"req-1-stage2-final",
]
@@ -544,7 +544,7 @@ async def test_async_omni_diffusion_only_yields_single_image_output(monkeypatch:
assert outputs[0].stage_id == 0
assert outputs[0].final_output_type == "image"
assert outputs[0].images == ["req-1-image"]
assert outputs[0].request_output.payload == "req-1-diffusion-final"
assert outputs[0].outputs[0].text == "req-1-diffusion-final"
@pytest.mark.asyncio
@@ -565,7 +565,7 @@ async def test_async_omni_llm_diffusion_yields_text_stream_then_image(monkeypatc
assert [output.stage_id for output in outputs] == [0, 0, 0, 1]
assert [output.final_output_type for output in outputs] == ["text", "text", "text", "image"]
assert [output.request_output.payload for output in outputs] == [
assert [output.outputs[0].text for output in outputs] == [
"req-1-text-0",
"req-1-text-1",
"req-1-text-2",
@@ -695,7 +695,7 @@ def test_omni_generate_py_generator_yields_final_outputs_for_each_request(monkey
assert len(outputs) == 4
assert [output.stage_id for output in outputs] == [0, 2, 0, 2]
assert [output.request_output.payload for output in outputs] == [
assert [output.outputs[0].text for output in outputs] == [
f"{engine.submitted[0]['request_id']}-stage0-0",
f"{engine.submitted[0]['request_id']}-stage2-final",
f"{engine.submitted[1]['request_id']}-stage0-0",
@@ -739,7 +739,7 @@ def test_omni_generate_diffusion_only_yields_single_image_per_request(monkeypatc
assert len(outputs) == 2
assert [output.stage_id for output in outputs] == [0, 0]
assert [output.final_output_type for output in outputs] == ["image", "image"]
assert [output.request_output.payload for output in outputs] == [
assert [output.outputs[0].text for output in outputs] == [
f"{engine.submitted[0]['request_id']}-diffusion-final",
f"{engine.submitted[1]['request_id']}-diffusion-final",
]
@@ -767,7 +767,7 @@ def test_omni_generate_llm_diffusion_yields_final_text_then_image_per_request(
assert len(outputs) == 4
assert [output.stage_id for output in outputs] == [0, 1, 0, 1]
assert [output.final_output_type for output in outputs] == ["text", "image", "text", "image"]
assert [output.request_output.payload for output in outputs] == [
assert [output.outputs[0].text for output in outputs] == [
f"{engine.submitted[0]['request_id']}-text-0",
f"{engine.submitted[0]['request_id']}-image-final",
f"{engine.submitted[1]['request_id']}-text-0",
@@ -1046,7 +1046,6 @@ def _enqueue_stage_error(
if kill_engine:
engine._alive = False
engine_output = OmniRequestOutput.from_error(msg["request_id"], error_text)
engine_output.payload = ""
engine.output_q.put_nowait(
OutputMessage(
request_id=msg["request_id"],
+1 -1
View File
@@ -205,7 +205,7 @@ def test_run_headless_parses_and_forwards_stage_overrides(mocker: MockerFixture)
with pytest.raises(ValueError, match="No stage config found for stage_id=0"):
run_headless(args)
assert captured["args"][1] is None
assert len(captured["args"]) == 2
assert captured["deploy_config_path"] == "/tmp/deploy.yaml"
assert captured["stage_overrides"] == {"0": {"devices": "0,1"}, "1": {"devices": "2"}}
assert captured["strategy_config_path"] == "/tmp/strategy.yaml"
@@ -69,10 +69,10 @@ def _make_text_omni_output(
],
finished=finish_reason is not None,
)
return OmniRequestOutput(
return OmniRequestOutput.from_stage_output(
res,
request_id=request_id,
final_output_type="text",
request_output=res,
finished=finish_reason is not None,
)
@@ -108,12 +108,12 @@ def _make_audio_omni_output(
outputs=[completion],
finished=True,
)
return OmniRequestOutput(
return OmniRequestOutput.from_stage_output(
res,
request_id=request_id,
stage_id=stage_id,
replica_id=replica_id,
final_output_type="audio",
request_output=res,
finished=True,
)
@@ -151,7 +151,7 @@ def _build_serving_chat():
)
instance._create_audio_choice = MagicMock(
side_effect=lambda omni_res, role, request, stream=False: _mock_audio_choices(
index=omni_res.request_output.outputs[0].index,
index=omni_res.outputs[0].index,
role=role,
)
)
@@ -489,7 +489,7 @@ async def test_audio_chunk_without_waveform_keeps_stream_alive():
request = _make_request(modalities=["text", "audio"])
empty_audio = _make_audio_omni_output()
empty_audio.request_output.outputs[0].multimodal_output = {"audio": []}
empty_audio.outputs[0].multimodal_output = {"audio": []}
def create_audio_choice(omni_res, role, request, stream=False):
return OmniOpenAIServingChat._create_audio_choice(serving_chat, omni_res, role, request, stream=stream)
+9 -7
View File
@@ -381,7 +381,6 @@ class TestLoadAndResolveStageConfigs:
config_path, stage_configs, _ = load_and_resolve_stage_configs(
model="black-forest-labs/FLUX.2-klein-4B",
stage_configs_path=None,
kwargs=kwargs,
trust_remote_code=False,
default_stage_cfg_factory=lambda: AsyncOmniEngine._create_default_diffusion_stage_cfg(kwargs),
@@ -390,9 +389,7 @@ class TestLoadAndResolveStageConfigs:
assert len(stage_configs) == 1
assert "dtype" in stage_configs[0]["engine_args"]
def test_stage_configs_path_promotes_new_deploy_yaml_without_expanding_replicas(
self, tmp_path, mocker: MockerFixture
):
def test_deploy_config_preserves_cli_overrides_and_replicas(self, tmp_path, mocker: MockerFixture):
deploy_path = tmp_path / "qwen3_multi.yaml"
deploy_path.write_text(
'stages:\n - stage_id: 0\n devices: "0"\n - stage_id: 1\n devices: "1,2,3"\n num_replicas: 3\n',
@@ -413,18 +410,23 @@ class TestLoadAndResolveStageConfigs:
"vllm_omni.entrypoints.utils.load_stage_configs_from_model",
return_value=(returned_stage_configs, None),
)
cli_overrides = {
"interleave_mm_strings": True,
"media_io_kwargs": {"video": {"fps": 1, "num_frames": 128}},
"gpu_memory_utilization": 0.55,
}
config_path, stage_configs, _ = load_and_resolve_stage_configs(
model="dummy-model",
stage_configs_path=str(deploy_path),
kwargs={},
kwargs=cli_overrides,
trust_remote_code=True,
deploy_config_path=str(deploy_path),
)
load_stage_configs.assert_called_once_with(
"dummy-model",
trust_remote_code=True,
base_engine_args={},
base_engine_args=cli_overrides,
deploy_config_path=str(deploy_path),
stage_overrides=None,
strategy_config_path=None,