[CI/Build] Diff-aware source_file_dependencies for CUDA/NPU pipelines

[yenuo26]

PLEASE FILL IN THE PR DESCRIPTION HERE.

Purpose

Make CUDA L2–L5 and NPU L4 jobs diff-aware on PR labels, so a nightly-test / weekly-test / ready / merge-test push no longer uploads the full pipeline. Scheduled main uploads (NIGHTLY=1, WEEKLY=1 / NON_CRITICAL=1, post-merge L3) still run every job.

Fixes #6507.

source_file_dependencies is a preset key (or a list of keys) from .buildkite/common/ci_source_file_dependencies.yml. upload_pipeline.py expands the key, keeps the step only when a changed file matches a prefix, then strips the field before Buildkite sees the YAML. Jobs with no key are always kept. --all / --e2e and BUILDKITE_BRANCH=main disable filtering.

Registry layout:

• Feature keys stay cross-cutting (diffusion_distributed_attention, diffusion_tiny_model, …).
• YAML anchors hold model business-code paths. Job keys (*_function / _perf / _accuracy / _reliability / _doc / _cov) alias those anchors and add the scripts that job actually runs. Pytest paths are listed on the key; they are not inferred from commands.
• Cross-model invalid-param reliability jobs use reliability_invalid_param_h100 / reliability_invalid_param_l4 (model anchors + tests/dfx/reliability/invalid_param_test/), not *_function.

Job wiring that follows from that:

• L2/L3 E2E, L4 CUDA/NPU, and weekly Reliability/Perf jobs point at the matching kind key. Weekly E2E (NON_CRITICAL=1 only) has no source filter.
• Shared H100 diffusion sweeps that timed out are split by model (list below). LingBot single-GPU coverage moved from slow to full_model and is one nightly job, Diffusion X2V · LingBot Function Test (diffusion_lingbot_function). World still skips unless VLLM_OMNI_LINGBOT_WORLD_V2_IMAGE_PATH and VLLM_OMNI_LINGBOT_WORLD_V2_ACTION_DIR are set; CI does not set them.
• X2V Doc Test uses diffusion_image_to_video_doc for the image-to-video README snippets that job owns: Wan2.2 TI2V/I2V, HunyuanVideo-1.5 I2V, LTX-2, SANA-Video, plus LingBot-Video (README command is skipped on /path/to/, but its model code still selects the job). Python API, Prerequisites, Advanced Features, FAQ, and Wan2.1 VACE stay out of this key.

Unused PR labels omni-test, tts-test, diffusion-x2iat-test, and diffusion-x2v-test no longer trigger CI. nightly-test remains the L4 label.

Test Plan

UT

Uploader unit tests construct synthetic pipelines instead of asserting live job labels. They cover key expansion, unrelated-file drops, jobs with no source key, composed doc paths, and stripping source_file_dependencies / expanding mirror_hardwares.

python -m pytest -sv tests/buildkite/test_upload_pipeline.py"

Registry keys referenced by CUDA/NPU YAML are checked by test_pipeline_source_file_dependency_keys_are_registered in the same file.

Weekly H100 diffusion jobs split from the old shared sweeps

The old Diffusion · H100 · Single-GPU and Diffusion · H100 · 2-GPU sweeps were split by model so one pytest process no longer collects every slow+cards_N file. Each job below must collect at least one test (--collect-only should print node ids, not no tests collected). Collection does not need GPU weights; it does need the repo test imports to succeed.

LingBot is not in this weekly list: those cases moved to nightly Diffusion X2V · LingBot Function Test.

Single-GPU (-m "slow and diffusion and H100 and cards_1" --run-level full_model):

pytest --collect-only -q tests/e2e/online_serving/test_qwen_image_edit.py tests/e2e/online_serving/test_qwen_image_edit_expansion.py tests/e2e/online_serving/test_qwen_image_layered.py tests/e2e/online_serving/test_qwen_image_layered_expansion.py tests/e2e/accuracy/test_qwen_image_edit.py tests/e2e/accuracy/test_qwen_image_layered.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_wan_2_1_vace_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_longcat_image_expansion.py tests/e2e/online_serving/test_longcat_image_edit_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_bagel.py tests/e2e/online_serving/test_bagel_expansion.py tests/e2e/offline_inference/test_bagel.py tests/e2e/offline_inference/test_bagel_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_boogu_image.py tests/e2e/online_serving/test_boogu_image_edit.py tests/e2e/online_serving/test_boogu_image_expansion.py tests/e2e/online_serving/test_boogu_image_edit_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_flux_2_dev_expansion.py tests/e2e/offline_inference/test_flux1_schnell_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_ltx.py tests/e2e/online_serving/test_ltx25.py tests/e2e/accuracy/ltx/test_ltx_official_similarity.py tests/e2e/accuracy/ltx/test_ltx25_official_similarity.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_krea2_expansion.py tests/e2e/offline_inference/test_krea2_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_hidream_i1_full_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_sana_video_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/offline_inference/test_sensenova_u1_expansion.py tests/e2e/offline_inference/test_sensenova_u1_text2img_expansion.py tests/e2e/offline_inference/test_sensenova_u1_img2img_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/offline_inference/test_mammoth_moda2_expansion.py -m "slow and diffusion and H100 and cards_1" --run-level full_model
pytest --collect-only -q tests/e2e/accuracy/test_video_streaming_output_similarity.py -m "slow and diffusion and H100 and cards_1" --run-level full_model

2-GPU (-m "slow and diffusion and H100 and cards_2" --run-level full_model):

pytest --collect-only -q tests/e2e/online_serving/test_qwen_image_edit_expansion.py tests/e2e/online_serving/test_qwen_image_layered_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_wan_2_1_vace_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_longcat_image_expansion.py tests/e2e/online_serving/test_longcat_image_edit_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_bagel_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_boogu_image_expansion.py tests/e2e/online_serving/test_boogu_image_edit_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_flux_2_dev_expansion.py tests/e2e/offline_inference/test_flux1_kontext_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_glm_image_expansion.py tests/e2e/offline_inference/test_glm_image_autoround_w4a16_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_ltx.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/offline_inference/test_krea2_expansion.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_dreamzero_expansion.py tests/e2e/accuracy/test_dreamzero.py -m "slow and diffusion and H100 and cards_2" --run-level full_model
pytest --collect-only -q tests/e2e/accuracy/test_hidream_o1_image.py -m "slow and diffusion and cuda and H100 and cards_2" --run-level full_model

3&4-GPU (-m "slow and diffusion and H100 and cards_2" --run-level full_model):

pytest --collect-only -q tests/e2e/online_serving/test_boogu_image_edit_expansion.py -k "test_double_guidance_cfg_parallel and cfg_parallel_3_full" -m "slow and diffusion and cuda and H100 and cards_3" --run-level full_model
pytest --collect-only -q tests/e2e/online_serving/test_magi2.py -m "slow and diffusion and cuda and H100 and cards_4" --run-level full_model

LingBot nightly command (H100, 1 GPU; World skips without the two asset env vars):

pytest -s -v tests/e2e/online_serving/test_lingbot_video.py tests/e2e/online_serving/test_lingbot_video_moe.py tests/e2e/offline_inference/test_lingbot_world_v2.py -m "full_model and diffusion and H100 and cards_1" --run-level "full_model"

Full ready/merge/nightly/weekly execution is expected from Buildkite after this PR is labeled. This change is CI YAML and the uploader only.

vLLM Version: N/A (no runtime / vLLM pin change)

vLLM-Omni Commit: 99ede1b

Test Result

• UT

• Weekly H100 diffusion jobs split from the old shared sweeps
  single GPU

2-GPU

3&4-GPU

• LingBot nightly

• Trigger
  when i modify lingbot and minimax testcase + nightly-test label
  GPU

NPU

when i modify lingbot and minimax testcase + NIGHTLY=1
GPU

NPU

BEFORE SUBMITTING: read CONTRIBUTING.md and run the precheck-pr skill with the code agent for a self-check against project conventions.
(anything written below this line will be removed by GitHub Actions)

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR was classified as CI work.

CI owner: @yenuo26

@yenuo26, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Resolved as of 0d6baf288dcf: the high-priority or low-quality signal noted on an earlier commit no longer applies.

[Gaohan123]

[P2] Keep affected jobs when the pipeline configuration changes

If a PR only changes .buildkite/npu/test-npu-nightly.yml and is triggered with nightly-test, the YAML path matches none of the registered source dependencies, so every test job is filtered out. The equivalent CUDA change leaves only the email aggregation step. This prevents changes to test commands, environments, or hardware settings from being exercised before merge. Please bypass source filtering for the affected pipeline when its YAML changes, and handle shared uploader/registry changes similarly.

[yenuo26]

fixed, Added fallback logic for common modules. If no source code matches are found, it will additionally check whether the changes involve common uploaders, helpers functions, or Buildkite configuration files. If any of these are matched, source filtering will be skipped.

[Gaohan123]

[P2] Include the example scripts executed by the text-to-image doc tests

This preset includes the test wrappers but omits the example scripts they execute. For example, changing only examples/online_serving/text_to_image/openai_chat_client.py or examples/offline_inference/text_to_image/text_to_image.py does not select the Doc Test job, even though those changes can break its tests. Please include examples/online_serving/text_to_image/ and examples/offline_inference/text_to_image/ in this preset so example-only changes receive the corresponding validation.

[yenuo26]

fixed

[Gaohan123]

[P2] Preserve the duplex audio fixtures in the dependency preset

The previous ready/merge dependency lists included tests/assets/minicpmo_4_5/response_required_16k.wav and tests/assets/minicpmo_4_5/soft_interrupt_16k.wav, but this preset drops both. The current duplex helpers still read these files and validate their SHA256 and audio format, so changing either fixture alone can break the tests without selecting the corresponding duplex jobs. Please restore both paths, or include their containing asset directory.

This remains reproducible at 0d6baf288dcf, despite the related earlier discussion being resolved.

[yenuo26]

When deleting, the consideration was that media files are generally not modified in terms of content, but rather replaced along with test files. However, it might be more appropriate to include tests/assets/minicpmo_4_5/, and this has now been added.

[Gaohan123]

[P2] Apply the shared-path bypass even when a model dependency matches

The fallback fixes configuration-only changes, but mixed changes still lose required coverage. At 63114228f2a05, changing only .buildkite/npu/test-npu-nightly.yml retains all 15 jobs; adding a Qwen3-Omni model change to the same diff reduces that to its two performance jobs. Other jobs whose commands, environments, or hardware settings may have changed in that YAML are filtered out. The same reduction occurs when combining tests/helpers/runtime.py with a Qwen3-Omni model change.

Please check the pipeline/shared-path bypass independently of _any_source_dependency_match, so an additional model change cannot suppress validation required by the shared change. The mixed-change test should expect all affected jobs to remain selected.

[yenuo26]

On this point, I lean toward keeping the current design where only the corresponding jobs are triggered based on changes to business code. My reasoning is as follows:
For example, when a user modifies the relevant business code and needs to make changes to the test cases for that model, they would modify the Buildkite configuration to add or adjust the corresponding job, or modify the relevant code in the helper functions. But if we go down the path of mixed-change judgment, then all jobs would be triggered in that case. And I think this kind of situation is likely the majority, because users typically only want to trigger nightly tests in a PR when the nightly test cases have changed. This seems to run counter to our goal of reducing the number of jobs triggered in a PR.

[Gaohan123]

LGTM. Thanks