Compare commits

..
Author SHA1 Message Date
tzhouam 0dbee5684b [Test] Match invalid realtime requests to the current endpoint contract
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
2026-10-04 03:19:52 +08:00
4 changed files with 109 additions and 84 deletions

No files matched your search

+34 -45
View File
@@ -1,46 +1,35 @@
# Focused weekly comparison against unchanged Omni main and official vLLM 0.30.0.
steps: steps:
- label: ':docker: Build image' - label: ":docker: Build image"
key: image-build key: image-build
commands: commands:
- bash .buildkite/cuda/scripts/build-ci-image.sh - "bash .buildkite/cuda/scripts/build-ci-image.sh"
agents: agents:
queue: cpu_queue_premerge queue: "cpu_queue_premerge"
- label: Upload selected weekly baseline
depends_on: image-build # L2 Test
key: upload-weekly-pipeline - label: "Upload Ready Pipeline"
commands: depends_on: image-build
- | key: upload-ready-pipeline
python3 - <<'PYBASELINE' commands:
import tempfile - python3 .buildkite/common/scripts/upload_pipeline.py --upload --all .buildkite/cuda/test-ready.yml
from pathlib import Path agents:
import yaml queue: "cpu_queue_premerge"
labels = set(['TTS · Audex · CosyVoice · Qwen3-TTS · L4', 'TTS · H100', 'Diffusion · Bagel · H100 · Single-GPU', 'Diffusion · SenseNova-U1 · H100 · Single-GPU', 'Diffusion · DreamZero · H100 · 2-GPU', 'Diffusion · H100 · 4-GPU', 'Diffusion · MAGI-2 Preview Serving Smoke · H100 · 4-GPU'])
source = yaml.safe_load(Path('.buildkite/cuda/test-weekly.yml').read_text()) # L3 Test
selected = set() - label: "Upload Merge Pipeline"
guard = 'python3 -c \'import vllm, torch, flashinfer, vllm._C_stable_libtorch; assert vllm.__version__ == \'"\'"\'0.30.0\'"\'"\'; assert torch.cuda.is_available(); print(\'"\'"\'Verified baseline runtime:\'"\'"\', vllm.__version__, torch.__version__)\'' depends_on: image-build
def select(steps): key: upload-merge-pipeline
kept = [] commands:
for step in steps: - python3 .buildkite/common/scripts/upload_pipeline.py --upload --all .buildkite/cuda/test-merge.yml
if not isinstance(step, dict): agents:
continue queue: "cpu_queue_premerge"
if 'steps' in step:
children = select(step['steps']) # L4 Test
if children: - label: "Upload Nightly Pipeline"
kept.append(dict(step, steps=children)) depends_on: image-build
elif step.get('label') in labels: key: upload-nightly-pipeline
item = dict(step) if: build.env("NIGHTLY") == "1"
commands = item['commands'] commands:
item['commands'] = [guard] + ([commands] if isinstance(commands, str) else commands) - python3 .buildkite/common/scripts/upload_pipeline.py --upload --all .buildkite/cuda/test-nightly.yml
kept.append(item) agents:
selected.add(item['label']) queue: "cpu_queue_premerge"
return kept
source['steps'] = select(source['steps'])
assert selected == labels, (selected, labels)
path = Path(tempfile.gettempdir()) / 'vllm-omni-selected-weekly-baseline.yml'
path.write_text(yaml.safe_dump(source, sort_keys=False, allow_unicode=True))
print('Selected baseline suites:', sorted(selected), flush=True)
PYBASELINE
python3 .buildkite/common/scripts/upload_pipeline.py --upload --all /tmp/vllm-omni-selected-weekly-baseline.yml
agents:
queue: cpu_queue_premerge
+1 -2
View File
@@ -36,8 +36,7 @@ RUN apt-get update && \
# for CI # for CI
COPY pyproject.toml setup.py ./ COPY pyproject.toml setup.py ./
COPY requirements/ requirements/ COPY requirements/ requirements/
# Keep source-built dependency versions independent of the Omni CI version. ENV SETUPTOOLS_SCM_PRETEND_VERSION=0.0.0
ENV VLLM_OMNI_VERSION_OVERRIDE=0.0.0
# Editable install: deps + project metadata are cached in this layer, while # Editable install: deps + project metadata are cached in this layer, while
# the package itself is linked to APP_DIR (populated later by COPY --link). # the package itself is linked to APP_DIR (populated later by COPY --link).
# Do NOT use a non-editable install here — that copies an empty stub into # Do NOT use a non-editable install here — that copies an empty stub into
@@ -216,44 +216,40 @@ def test_video_chat_stream_invalid_requests(
# WS /v1/realtime # WS /v1/realtime
# ───────────────────────────────────────────────────────────────────────────── # ─────────────────────────────────────────────────────────────────────────────
# Resolved in ``test_realtime_invalid_requests`` (skip ``session.created`` handshake).
_REALTIME_WS_MODEL_MISMATCH = object()
_REALTIME_WS_INVALID_AUDIO_APPEND = object()
@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2) @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
@pytest.mark.parametrize( @pytest.mark.parametrize(
"send_frames_spec, ws_error_code, err_message", "send_frames_spec, ws_error_code, err_message",
[ [
pytest.param("{not-json", "invalid_json", "Invalid JSON", id="invalid_json"), pytest.param("{not-json", "invalid_event", "Invalid or unrecognized client event", id="invalid_json"),
pytest.param( pytest.param(
json.dumps({"type": "session.update"}), json.dumps({"type": "session.update"}),
"invalid_event", "invalid_event",
("model", "Missing required field"), "Invalid or unrecognized client event",
id="session_update_missing_model", id="session_update_missing_session",
), ),
pytest.param( pytest.param(
_REALTIME_WS_MODEL_MISMATCH, json.dumps({"type": "session.update", "session": []}),
"model_not_found", "invalid_event",
("model", "does not exist"), "Invalid or unrecognized client event",
id="session_update_model_mismatch", id="session_update_invalid_session",
), ),
pytest.param( pytest.param(
json.dumps({"type": "input_audio_buffer.commit", "final": False}), json.dumps({"type": "input_audio_buffer.commit"}),
"model_not_validated", "invalid_request_error",
("session.update", "validate the model"), "Input audio buffer is empty",
id="commit_without_model_validation", id="commit_without_audio",
), ),
pytest.param( pytest.param(
json.dumps({"type": "unknown.event"}), json.dumps({"type": "unknown.event"}),
"unknown_event", "invalid_event",
"Unknown event type", "Invalid or unrecognized client event",
id="unknown_event_type", id="unknown_event_type",
), ),
pytest.param( pytest.param(
_REALTIME_WS_INVALID_AUDIO_APPEND, json.dumps({"type": "input_audio_buffer.append", "audio": "not-valid-base64!!!"}),
"invalid_audio", "invalid_request_error",
"Invalid audio data", "Invalid base64 audio data",
id="invalid_audio_append", id="invalid_audio_append",
), ),
], ],
@@ -266,27 +262,12 @@ def test_realtime_invalid_requests(
ws_error_code: str, ws_error_code: str,
err_message: str | tuple[str, ...], err_message: str | tuple[str, ...],
) -> None: ) -> None:
send_frames: Any
if send_frames_spec is _REALTIME_WS_MODEL_MISMATCH:
send_frames = json.dumps(
{
"type": "session.update",
"model": "this-model-is-not-served-on-this-instance",
}
)
elif send_frames_spec is _REALTIME_WS_INVALID_AUDIO_APPEND:
send_frames = [
json.dumps({"type": "session.update", "model": omni_server.model}),
json.dumps({"type": "input_audio_buffer.append", "audio": "not-valid-base64!!!"}),
]
else:
send_frames = send_frames_spec
online_client.send_realtime_ws_request( online_client.send_realtime_ws_request(
{ {
"send_frames": send_frames, "send_frames": send_frames_spec,
"timeout": 120, "timeout": 120,
"ws_max_size": None, "ws_max_size": None,
"ws_skip_types": ["session.created"], "ws_skip_types": ["session.created", "conversation.created"],
"ws_json_type": "error", "ws_json_type": "error",
"ws_error_code": ws_error_code, "ws_error_code": ws_error_code,
"err_message": err_message, "err_message": err_message,
@@ -4,7 +4,10 @@
from __future__ import annotations from __future__ import annotations
import asyncio
import base64 import base64
import json
from types import SimpleNamespace
import numpy as np import numpy as np
import pytest import pytest
@@ -84,3 +87,56 @@ class TestAsyncOmniStreamingParamsValidation:
p = SamplingParams(n=1, stop=["\n"], output_kind=RequestOutputKind.DELTA) p = SamplingParams(n=1, stop=["\n"], output_kind=RequestOutputKind.DELTA)
with pytest.raises(ValueError, match="Input streaming"): with pytest.raises(ValueError, match="Input streaming"):
AsyncOmni._validate_streaming_input_sampling_params(p) AsyncOmni._validate_streaming_input_sampling_params(p)
@pytest.mark.parametrize(
"frame,code,message",
[
("{not-json", "invalid_event", "Invalid or unrecognized client event"),
(json.dumps({"type": "session.update"}), "invalid_event", "Invalid or unrecognized client event"),
(
json.dumps({"type": "session.update", "session": []}),
"invalid_event",
"Invalid or unrecognized client event",
),
(json.dumps({"type": "input_audio_buffer.commit"}), "invalid_request_error", "Input audio buffer is empty"),
(json.dumps({"type": "unknown.event"}), "invalid_event", "Invalid or unrecognized client event"),
(
json.dumps({"type": "input_audio_buffer.append", "audio": "not-valid-base64!!!"}),
"invalid_request_error",
"Invalid base64 audio data",
),
],
)
def test_full_duplex_invalid_request_contract(frame, code, message):
from starlette.websockets import WebSocketDisconnect
from vllm_omni.entrypoints.openai.realtime.connection import OpenAIFullDuplexConnection
class Socket:
def __init__(self):
self.sent = []
self.frames = iter([frame])
async def accept(self):
pass
async def receive_text(self):
value = next(self.frames, None)
if value is None:
raise WebSocketDisconnect()
return value
async def send_text(self, payload):
self.sent.append(json.loads(payload))
async def run():
socket = Socket()
connection = OpenAIFullDuplexConnection(socket, SimpleNamespace(), "test-model", None)
await connection.handle_connection()
assert [event["type"] for event in socket.sent] == ["session.created", "conversation.created", "error"]
error = socket.sent[-1]["error"]
assert error["code"] == code
assert error["message"] == message
asyncio.run(run())