[Bugfix] Expose base_model_paths property in _DiffusionServingModels

[RuixiangMa]

Purpose

In pure diffusion mode, _DiffusionServingModels uses getattr to intercept undefined attributes. Accessing base_model_paths returned a _Unsupported object instead of the actual list, causing NotImplementedError when
retrieving model names for video generation endpoints

curl -s -D >(grep -i x-request-id >&2) -o >(jq -r '.data[0].b64_json' | base64 --decode > output.png) -X POST "http://localhost:8004/v1/images/edits" -F "model=black-forest-labs/FLUX.2-klein-4B" -F "image=@./test.jpg" -F "prompt=change into orange." -F "size=1024x1024" -F "output_format=png"

APIServer pid=494089)     model_name = serving_models.base_model_paths[0].name
(APIServer pid=494089)                  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^
(APIServer pid=494089) TypeError: '_Unsupported' object is not subscriptable

Test Plan

Test Result

(APIServer pid=111918) INFO 03-10 11:50:44 [diffusion_engine.py:86] Generation completed successfully.
(APIServer pid=111918) INFO 03-10 11:50:44 [diffusion_engine.py:108] Post-processing completed in 0.1113 seconds
(APIServer pid=111918) INFO 03-10 11:50:44 [diffusion_engine.py:111] DiffusionEngine.step breakdown: preprocess=0.00 ms, add_req_and_wait=25044.09 ms, postprocess=111.33 ms, total=25156.01 ms
(APIServer pid=111918) INFO 03-10 11:50:44 [omni_diffusion.py:128] OmniDiffusion.generate total: 25156.33 ms
(APIServer pid=111918) INFO 03-10 11:50:44 [api_server.py:1395] Successfully generated 1 image(s)
(APIServer pid=111918) INFO: 127.0.0.1:52072 - "POST /v1/images/edits HTTP/1.1" 200 OK

[wtomin]

Is your PR going to fix a CI failures introduced by recent commit 7a56049?

[RuixiangMa]

Is your PR going to fix a CI failures introduced by recent commit 7a56049?

Yes, tested against docs/serving/image_edit_api.md using the specified curl command on the main branch (7a56049 (HEAD -> main, origin/main) Add online serving to Stable Audio Diffusion and introduce v1/audio/generate endpoint (#1255))
curl -s -D >(grep -i x-request-id >&2) -o >(jq -r '.data[0].b64_json' | base64 --decode > output.png) -X POST "http://localhost:8004/v1/images/edits" -F "model=black-forest-labs/FLUX.2-klein-4B" -F "image=@./test.jpg" -F "prompt=change into orange." -F "size=1024x1024" -F "output_format=png

If my command is ok, the current failure suggests a recent regression, as tests passed after merging this PR.

[lishunyang12]

LGTM — fix is correct. Left a minor note about the __getattr__ design.

[lishunyang12]

Nit: the broader issue is that __getattr__ returns _Unsupported instead of raising AttributeError, so hasattr() always returns True for any attribute on this class. This fix is fine for the immediate bug, but any future hasattr() check will silently pass too.

[RuixiangMa]

Since the buggy commit got reverted, is this still needed?

[wtomin]

I think it's still needed, because there is a buggy code here:

vllm-omni/vllm_omni/entrypoints/openai/api_server.py

Lines 1400 to 1403 in 74eedd3

	if serving_models and hasattr(serving_models, "base_model_paths") and serving_models.base_model_paths:
	model_name = serving_models.base_model_paths[0].name
	else:
	model_name = "unknown"

[wtomin]

Please solve this conflicts.

[RuixiangMa]

Please solve this conflicts.

Done