[RFC]: Qwen-Image、Qwen-Image-Layered、Qwen-Image-Edit-Plus、Wan2.2 Production-grade Feature Monitoring

[david6666666]

Motivation.

Framework：

• Abort Request reafator pipeline stage/step pipeline vllm-project/vllm-omni#1368 [Refactor] Refactor Diffusion Scheduler/Executor Boundaries and Request State Flow vllm-project/vllm-omni#1625 依赖continue batching中的 step scheduler [Feat] Support step-boundary abort in diffusion vllm-project/vllm-omni#1769 实现
• Exit on OOM [RFC]: Exit on OOM vllm-project/vllm-omni#1346 [Enhancement] Patch AsyncOmniEngine try_get_output[_async] hanging issues vllm-project/vllm-omni#2153 [Entrypoint][Refactor] vLLM-Omni Entrypoint Refactoring vllm-project/vllm-omni#1908 [Enhancement] Patch OmniStage.try_collect() with _proc alive checks vllm-project/vllm-omni#1560 [Enhancement] Patch OmniStage.try_collect() with ray alive checks vllm-project/vllm-omni#1561 [Enhancement][pytest] Check for process running during start server vllm-project/vllm-omni#1559
• video api async job [Frontend] Rewrite video API for async job lifecycle vllm-project/vllm-omni#1665
• [Feat][Benchmark] Add synchronous video generation endpoint POST /v1/videos/sync for benchmark test vllm-project/vllm-omni#2049
• Accuracy Bench ： （1）vbench clipscore （2）G Edit bench/G Image bench from stepFun [Feature]: Accuracy Benchmarks for T2I and I2I Generation vllm-project/vllm-omni#1788 [Accuracy Benchmark] feat: add accuracy benchmark integrations for t2i and i2i vllm-project/vllm-omni#1917
• Random Bench：[Benchmark] [Diffusion] [Enhancement] Random dataset vllm-project/vllm-omni#1657
• [Bug]: Server return 500 when user send request with illegal parameters vllm-project/vllm-omni#1952

Wan2.2：

• [Bug]: "MISSING params from checkpoint" warning during initialization and potential accuracy degradation on Ascend NPU (vllm-omni v0.18.0rc1) vllm-project/vllm-omni#2229
• [Bug]: Video generation accuracy/quality degradation for Wan2.2 in vllm-omni 0.18.0rc1 vllm-project/vllm-omni#2247 补精度用例看护
• Wan2.2 Performance L4 Test @bjf-frz
• Wan2.2 Accuracy L4 Test [Accuracy Benchmark]Add VBench video accuracy benchmarks t2v and i2v vllm-project/vllm-omni#2209
• [CI]test: add wan22 i2v video similarity e2e vllm-project/vllm-omni#2262
• Wan2.2 [Performance] diffusion timing vllm-project/vllm-omni#1757
• Wan2.2 L4 Test [Test] L4 complete diffusion feature test for Wan2.2 models vllm-project/vllm-omni#2087

Qwen-Image-Layered：

• [Bug]: Diffusion Model Online Service API Does Not Support Empty Prompt, Preventing Models That Do Not Require a Prompt (e.g., qwen-image-layered) from Being Used vllm-project/vllm-omni#1966 [Bug-Fix]fix bug of empty prompt input vllm-project/vllm-omni#2041
• [Bug]: Image-to-image model's multi-image output in the API response retains only the first image vllm-project/vllm-omni#1969 [Bugfix]Fix bug of online server can not return mutli images vllm-project/vllm-omni#2007
• [Bug]: The online inference results of the Qwen-Image-Layered model are inconsistent with the offline inference results, maybe there are accuracy issues in online inference. vllm-project/vllm-omni#1981 [Bugfix] Fix config misalignment between offline and online diffusion inference (Wan2.2, Qwen-Image series) vllm-project/vllm-omni#1979
• [Bug]: Custom parameters (e.g., response_format / size / user) passed to API v1/images/edits are ignored vllm-project/vllm-omni#2081 [API] Add layers and resolution parameters to /v1/images/edits endpoint vllm-project/vllm-omni#2053
• [Bug]: API v1/images/edits does not support generating multiple images vllm-project/vllm-omni#2082 [API] Add layers and resolution parameters to /v1/images/edits endpoint vllm-project/vllm-omni#2053
• [Bugfix] fix: handle Qwen-Image-Layered layered RGBA output for jpeg edits vllm-project/vllm-omni#2297
• [Bugfix] fix: return 400 for unsupported multi-image edits such as Qwen-Image-Layered vllm-project/vllm-omni#2298

v0.18.0 post1 need to cherry-pick

• [Bugfix] fix: handle Qwen-Image-Layered layered RGBA output for jpeg edits vllm-project/vllm-omni#2297
• [Bugfix] fix: return 400 for unsupported multi-image edits such as Qwen-Image-Layered vllm-project/vllm-omni#2298
• Fix completion_tokens not correct vllm-project/vllm-omni#2313

Proposed Change.

Please provide the detailed design document of the RFC using the template.

Feedback Period.

No response

CC List.

No response

Any Other Things.

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.