Compare commits

..
Author SHA1 Message Date
tzhouam fc0a1dc1e4 [Test] Initialize tensor-parallel groups in native SANA accuracy fixtures
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
2026-10-04 03:27:05 +08:00
5 changed files with 63 additions and 47 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
+26
View File
@@ -0,0 +1,26 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
import pytest
@pytest.fixture
def sana_model_parallel(tmp_path):
"""Construct native parallel layers with real single-rank GPU groups."""
from vllm_omni.diffusion.distributed.parallel_state import (
destroy_distributed_env,
init_distributed_environment,
initialize_model_parallel,
)
init_distributed_environment(
world_size=1,
rank=0,
local_rank=0,
distributed_init_method=f"file://{tmp_path / 'sana-distributed'}",
)
try:
initialize_model_parallel()
yield
finally:
destroy_distributed_env()
@@ -202,6 +202,7 @@ def test_sana_video_transformer_matches_frozen_golden(
variant: str, variant: str,
_hardware, _hardware,
accuracy_artifact_root: Path, accuracy_artifact_root: Path,
sana_model_parallel,
) -> None: ) -> None:
del _hardware del _hardware
from vllm.config import LoadConfig from vllm.config import LoadConfig
@@ -189,6 +189,7 @@ def _cfg_prediction(
def test_sana_video_quick_pipeline_stages_match_diffusers( def test_sana_video_quick_pipeline_stages_match_diffusers(
variant: str, variant: str,
_hardware, _hardware,
sana_model_parallel,
) -> None: ) -> None:
"""Compare the first denoising step without replacing native attention. """Compare the first denoising step without replacing native attention.