[Bugfix][Qwen-Image] Read resolved offload strategy

[Ai-Eastern]

Purpose

Fixes #8194. The Qwen-Image pipeline still read enable_cpu_offload directly after the offload policy migration in #7327. This triggers the legacy-reader CI assertion and makes initial text-encoder/VAE placement depend on a compatibility alias. Read the resolved MODEL_LEVEL strategy instead; keep layerwise placement and the loader's DiT device choice intact.

Reproduction and root cause

On the issue's base commit, run:

pytest -q tests/diffusion/offloader/test_legacy_flag_readers.py

The assertion reports diffusion/models/qwen_image/pipeline_qwen_image.py:344. The remaining legacy read is in QwenImagePipeline.__init__, where it chooses the initial text-encoder/VAE device. A new CPU-marked mocked pipeline test covers compact mode="module" and mode="layer" configs independently of the old flag.

Test Plan

pytest -q tests/diffusion/offloader/test_legacy_flag_readers.py tests/diffusion/models/qwen_image/test_qwen_image_pipeline_device.py
pytest -q tests/diffusion/offloader/test_legacy_flag_readers.py tests/diffusion/models/qwen_image/test_qwen_image_pipeline_device.py -m "core_model and cpu" --run-level=core_model

Prerequisites: a supported Python environment with the repository's development dependencies. These tests use CPU and mocked model loading; no model weights are needed.

vLLM Version: 0.30.0 in the issue's CI report; local tests used the 0.30.0+cpu wheel in a Linux container.

vLLM-Omni Commit: 7ab582ae2545476284bef8903bec1c0ddc44a082 (PR base).

Test Result

• Passed: the two repository pytest modules in a local Linux CPU container (python:3.12-slim-trixie, vLLM 0.30.0+cpu, PyTorch 2.13.0+cpu): 5 passed with the plain command and 5 passed with -m "core_model and cpu" --run-level=core_model. These are CPU tests with mocked model loading; no model weights or GPU inference were used.
• Passed: a local stdlib-only probe exercised the production strategy resolver and placement expressions for legacy model offload, compact module offload, and compact layer offload. It also scanned runtime files using the existing legacy-reader pattern with normalized paths. This probe is not part of the PR.
• Passed: the original test_no_runtime_module_reads_the_legacy_offload_flags test body against this branch's source in an isolated Windows pytest runner (1 passed). The runner used a lightweight vllm_omni package stub, UTF-8 mode, and a Windows-path adaptation for the test's POSIX-style compatibility allowlist; this is narrower than a run through the repository test environment.
• Passed: Python 3.12 compileall on both changed files; uvx --python 3.12 ruff check and ruff format --check on both changed files; git diff --check.
• Passed: repository hooks check-mark, check-spdx-header, check-forbidden-imports, and check-torch-cuda-call on the changed files.
• Not run: GPU/model inference and full local pre-commit. GitHub build, DCO, documentation, and pre-commit checks passed; maintainer review is pending.

AI assistance: Codex drafted the source change, regression test, local verification probe, and this PR description. The human author reviewed the diff and remains responsible for the final contribution and review responses.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/offloader.md, docs/design/module/diffusion/diffusion_model_integration.md, docs/design/module/diffusion/index.md.

Module owners: @wtomin @david6666666 @Isotr0py

Routing: @wtomin via module of the changed files, module named in the PR description, CODEOWNERS; @david6666666 via module of the changed files, module named in the PR description; @Isotr0py via module of the changed files, module named in the PR description

@Ai-Eastern, 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

Automated triage of commit 417cccb58a53 produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[Ai-Eastern]

Self-review: I reviewed the Qwen-Image offload change and its regression test. Encoder/VAE placement follows the resolved strategy, and DiT keeps the loader-selected device. The compact module and layer cases are covered. Both targeted pytest commands passed locally in a Linux CPU container: 5 passed each. GPU/model inference was not run.

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Direct under experiment vllm-omni-strict-5050-20260829.

[vllm-omni-review-bot]

Omni ReviewBot review

PR description

Qwen-Image still chose the initial text-encoder and VAE device from the legacy enable_cpu_offload alias after the offload-policy migration. The pipeline now asks the shared strategy resolver and parks those modules on CPU only when the resolved strategy is model-level. Layerwise and other non-model-level paths still place encoder/VAE on the pipeline device, and the DiT still follows the loader device context. A mocked CPU test adds compact mode="module" and mode="layer" cases so placement is no longer tied to the old flag.

Change flow

flowchart TD
    A["[EXISTING] od_config offload fields"]:::existing --> B["[EXISTING] resolve_offload_strategy"]:::existing
    B --> C["[CHANGED] QwenImagePipeline encoder/VAE placement"]:::changed
    C --> D["[EXISTING] MODEL_LEVEL parks encoder/VAE on CPU"]:::existing
    C --> E["[EXISTING] DiT stays on loader device"]:::existing
    C --> F["[NEW] compact module vs layer device test"]:::new
    classDef existing fill:#e5e7eb,stroke:#6b7280,color:#111827
    classDef changed fill:#fef3c7,stroke:#d97706,color:#451a03,stroke-width:2px
    classDef new fill:#dcfce7,stroke:#16a34a,color:#052e16,stroke-width:2px
    classDef removed fill:#fee2e2,stroke:#dc2626,color:#450a0a,stroke-width:2px

Loading

CI at 417cccb58a53 (2026-09-27T20:30:14.040539+00:00): required check(s) blocking: buildkite/vllm-omni (missing), and -5 more.

No actionable findings.

[Ai-Eastern]

Both targeted CPU pytest commands passed locally (5 passed each; details are in the PR description). I don't have access to an H100, so I couldn't run the Qwen-Image L2 job locally. Could a maintainer advise how to proceed with the ready label and Buildkite/H100 validation?

[NumberWan]

LGTM

[NickCao]

Already done in #7695