Compare commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
82aa524c8e |
No files matched your search
@@ -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
|
|
||||||
@@ -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
|
||||||
|
|||||||
@@ -13,6 +13,7 @@ from vllm import SamplingParams
|
|||||||
from vllm.config.compilation import CUDAGraphMode
|
from vllm.config.compilation import CUDAGraphMode
|
||||||
from vllm.v1.core.sched.output import SchedulerOutput
|
from vllm.v1.core.sched.output import SchedulerOutput
|
||||||
from vllm.v1.worker.gpu.model_runner import GPUModelRunner
|
from vllm.v1.worker.gpu.model_runner import GPUModelRunner
|
||||||
|
from vllm.v1.worker.gpu.model_states.default import DefaultModelState
|
||||||
from vllm.v1.worker.gpu.sample.sampler import Sampler
|
from vllm.v1.worker.gpu.sample.sampler import Sampler
|
||||||
|
|
||||||
from vllm_omni.config.model import OmniModelConfig
|
from vllm_omni.config.model import OmniModelConfig
|
||||||
@@ -163,6 +164,28 @@ def test_init_model_state_factory_dispatches_omni_only(monkeypatch, flag, expect
|
|||||||
assert state is upstream.return_value
|
assert state is upstream.return_value
|
||||||
|
|
||||||
|
|
||||||
|
@pytest.mark.parametrize("omni_stage", [False, True])
|
||||||
|
@pytest.mark.parametrize("specialized", [False, True])
|
||||||
|
def test_auxiliary_default_state_has_omni_lifecycle_but_preserves_upstream_states(monkeypatch, omni_stage, specialized):
|
||||||
|
class SpecializedState(DefaultModelState):
|
||||||
|
pass
|
||||||
|
|
||||||
|
state_cls = SpecializedState if specialized else DefaultModelState
|
||||||
|
upstream_state = object.__new__(state_cls)
|
||||||
|
monkeypatch.setattr("vllm_omni.worker_v2.model_states._upstream_init_model_state", lambda *_args: upstream_state)
|
||||||
|
monkeypatch.setattr(OmniModelState, "__init__", lambda *_args: None)
|
||||||
|
model_config = object.__new__(OmniModelConfig) if omni_stage else SimpleNamespace()
|
||||||
|
state = init_omni_model_state(
|
||||||
|
SimpleNamespace(model_config=model_config), SimpleNamespace(), None, torch.device("cpu")
|
||||||
|
)
|
||||||
|
if omni_stage and not specialized:
|
||||||
|
assert isinstance(state, OmniModelState)
|
||||||
|
assert callable(state.run_preprocess)
|
||||||
|
assert callable(state.postprocess_model_output)
|
||||||
|
else:
|
||||||
|
assert state is upstream_state
|
||||||
|
|
||||||
|
|
||||||
def test_finish_requests_notifies_model_and_cleans_only_known_slots(monkeypatch):
|
def test_finish_requests_notifies_model_and_cleans_only_known_slots(monkeypatch):
|
||||||
runner = _make_runner()
|
runner = _make_runner()
|
||||||
calls = []
|
calls = []
|
||||||
|
|||||||
@@ -4,7 +4,8 @@
|
|||||||
"""Omni-aware ``init_model_state`` factory.
|
"""Omni-aware ``init_model_state`` factory.
|
||||||
|
|
||||||
Extends the upstream v2 factory with Omni architecture dispatch.
|
Extends the upstream v2 factory with Omni architecture dispatch.
|
||||||
Non-Omni architectures fall through to the upstream ``init_model_state``.
|
Non-Omni architectures use upstream dispatch; its default state acquires
|
||||||
|
Omni lifecycle support when owned by an Omni stage.
|
||||||
"""
|
"""
|
||||||
|
|
||||||
from __future__ import annotations
|
from __future__ import annotations
|
||||||
@@ -16,8 +17,11 @@ from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
|
|||||||
from vllm.v1.worker.gpu.model_states import (
|
from vllm.v1.worker.gpu.model_states import (
|
||||||
init_model_state as _upstream_init_model_state,
|
init_model_state as _upstream_init_model_state,
|
||||||
)
|
)
|
||||||
|
from vllm.v1.worker.gpu.model_states.default import DefaultModelState
|
||||||
from vllm.v1.worker.gpu.model_states.interface import ModelState
|
from vllm.v1.worker.gpu.model_states.interface import ModelState
|
||||||
|
|
||||||
|
from vllm_omni.config.model import OmniModelConfig
|
||||||
|
|
||||||
|
|
||||||
def init_omni_model_state(
|
def init_omni_model_state(
|
||||||
vllm_config: VllmConfig,
|
vllm_config: VllmConfig,
|
||||||
@@ -28,8 +32,8 @@ def init_omni_model_state(
|
|||||||
"""Create the appropriate ``ModelState`` for *model*.
|
"""Create the appropriate ``ModelState`` for *model*.
|
||||||
|
|
||||||
Returns an ``OmniModelState`` when the model declares Omni lifecycle
|
Returns an ``OmniModelState`` when the model declares Omni lifecycle
|
||||||
capabilities; otherwise
|
capabilities. Otherwise delegates to the upstream v2 factory, wrapping
|
||||||
delegates to the upstream v2 factory.
|
its default state for auxiliary Omni stages.
|
||||||
"""
|
"""
|
||||||
factory = getattr(model, "create_omni_model_state", None)
|
factory = getattr(model, "create_omni_model_state", None)
|
||||||
if factory is not None:
|
if factory is not None:
|
||||||
@@ -49,4 +53,12 @@ def init_omni_model_state(
|
|||||||
)
|
)
|
||||||
|
|
||||||
return OmniModelState(vllm_config, model, encoder_cache, device)
|
return OmniModelState(vllm_config, model, encoder_cache, device)
|
||||||
return _upstream_init_model_state(vllm_config, model, encoder_cache, device)
|
state = _upstream_init_model_state(vllm_config, model, encoder_cache, device)
|
||||||
|
if type(state) is DefaultModelState and isinstance(getattr(vllm_config, "model_config", None), OmniModelConfig):
|
||||||
|
# Auxiliary stages such as the forced aligner have no model-owned
|
||||||
|
# hooks, but Omni's runner still needs its request/output lifecycle.
|
||||||
|
# Keep specialized upstream states and plain vLLM configs intact.
|
||||||
|
from vllm_omni.worker_v2.model_states.omni_model_state import OmniModelState
|
||||||
|
|
||||||
|
return OmniModelState(vllm_config, model, encoder_cache, device)
|
||||||
|
return state
|
||||||
Reference in new issue
Block a user