Your current environment

OS                           : Ubuntu 24.04.1 LTS (x86_64)
PyTorch version              : 2.10.0+cu128
Python version               : 3.12.11
CUDA runtime version         : Could not collect
GPU 0-3: NVIDIA B200
Nvidia driver version        : 580.126.09
vLLM Version                 : 0.19.0
vLLM-Omni Version            : 0.19.0rc2.dev32+g408365fa0 (git sha: 408365fa0)

Your code version

vllm==0.19.0

408365fa (main branch, 2026-04-08)

🐛 Describe the bug

During pytest execution with pytest -v -s -m "core_model and cpu", every comfyui integration test logs the following AttributeError from the APIServer subprocess on shutdown:

(APIServer pid=XXXXX) Traceback (most recent call last):
  File ".../tests/comfyui/test_comfyui_integration.py", line 452, in run_server
    cmd.cmd(args)
  File ".../vllm_omni/entrypoints/cli/serve.py", line 94, in cmd
    uvloop.run(omni_run_server(args))
  ...
  File ".../vllm_omni/entrypoints/openai/api_server.py", line 356, in omni_run_server_worker
    app.state.openai_serving_speech.shutdown()
    ^^^^^^^^^
AttributeError: 'FastAPI' object has no attribute 'state'

Observed behavior:

• The error appears on every comfyui test (18/18 tests) consistently
• Tests themselves still PASS — the error occurs in the APIServer child process during shutdown, after the test assertions have completed
• Reproduced on both main branch (408365f) and a feature branch based on it

How to reproduce:

pytest -v -s -m "core_model and cpu" -k "comfyui" 2>&1 | grep "AttributeError"

Output:

(APIServer pid=562371) AttributeError: 'FastAPI' object has no attribute 'state'
(APIServer pid=562377) AttributeError: 'FastAPI' object has no attribute 'state'
... (repeated for all 18 comfyui tests)

Root cause:
In vllm_omni/entrypoints/openai/api_server.py:356, the finally block in omni_run_server_worker calls app.state.openai_serving_speech.shutdown(), but by that point uvicorn has already torn down the FastAPI app and app.state is no longer accessible.

Verified that adding a getattr guard eliminates the error while all 18 tests continue to pass.

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.