[Bugfix] Select the AR-Diffusion engine in the DreamZero CFG-parallel deploy config

[tzhouam]

Purpose

Fixes #5046.

DreamZeroPipeline requires the AR-Diffusion engine and raises at construction when engine_backend resolves to anything else. #4534 added engine_backend to vllm_omni/deploy/dreamzero.yaml but not to vllm_omni/deploy/dreamzero_tp1_cfg2.yaml, and nothing else injects it (OmniDiffusionConfig.engine_backend defaults to "default"). Every launch through the CFG-parallel config therefore died during worker startup:

ValueError: DreamZeroPipeline requires the AR-Diffusion engine; set engine_backend:
vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine in the deploy config
(got engine_backend='default').

Both workers died, the executor got EOFError on reader.recv(), and stage init tore the server down — the rest of the traceback in the issue is that cascade.

This is not CI-only: examples/online_serving/dreamzero/run_server.sh defaults to the same config, so the documented multi-GPU serving path was broken too.

Two changes:

1. Add the missing engine_backend to dreamzero_tp1_cfg2.yaml, matching dreamzero.yaml. No other change is needed — CFG parallelism is already handled by the capability, which collapses the negative branch onto local_index=0 when cfg_world >= 2.
2. Add tests/dreamzero/test_deploy_configs.py, a CPU test asserting every shipped dreamzero*.yaml stage selects an AR-Diffusion engine. It resolves the import path through DiffusionEngine.resolve_engine_class rather than string-matching, so a stale or typo'd path fails too. Previously only an H100x2 slow-marked e2e test covered this, which is why it surfaced weekly rather than in PR CI.

Test Plan

• pytest tests/dreamzero/test_deploy_configs.py
• Negative control: temporarily remove the new engine_backend line and confirm only the dreamzero_tp1_cfg2.yaml case fails.
• pre-commit run --files vllm_omni/deploy/dreamzero_tp1_cfg2.yaml tests/dreamzero/test_deploy_configs.py

vLLM Version: 0.25.0

vLLM-Omni Commit: 84bbd57

Test Result

tests/dreamzero/test_deploy_configs.py::test_dreamzero_deploy_configs_exist PASSED
tests/dreamzero/test_deploy_configs.py::test_dreamzero_stage_selects_ar_diffusion_engine[dreamzero.yaml-0-stage0] PASSED
tests/dreamzero/test_deploy_configs.py::test_dreamzero_stage_selects_ar_diffusion_engine[dreamzero_tp1_cfg2.yaml-0-stage1] PASSED
3 passed in 9.96s

Negative control, with the fix reverted:

FAILED tests/dreamzero/test_deploy_configs.py::test_dreamzero_stage_selects_ar_diffusion_engine[dreamzero_tp1_cfg2.yaml-0-stage1]
1 failed, 2 passed

pre-commit: all hooks passed.

Not run: the weekly test_dreamzero_openpi_online e2e test, which needs 2x H100. The new test covers the config-level regression, not the full serving path — worth confirming the weekly job goes green after this merges.

🤖 Generated with Claude Code

[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.

[yenuo26]

please use get_deploy_config_path

[tzhouam]

Done in 6673f83 — now resolves through tests.helpers.stage_config.get_deploy_config_path, matching tests/test_config_factory.py, so the CI-overlay resolution applies too.

The directory is still derived from the helper (Path(get_deploy_config_path("dreamzero.yaml")).parent) rather than hardcoding a filename list, so a future dreamzero*.yaml is covered without editing the test. Each discovered name is then resolved back through the helper.

Re-verified: 3 passed; and with the engine_backend line removed again, only the dreamzero_tp1_cfg2.yaml case fails.