vllm-omni backend (cuda13) fails to load: editable install points at build path, and unpinned vllm-omni main no longer matches backend.py

backend/python/vllm-omni/install.sh、backend/python/common/libbackend.sh、backend/python/vllm-omni/backend.py を読み、クローンした vllm-omni のリビジョンを import と Omni 引数と比較します。issue にある gallery model のリクエストを再現し、その後、editable-path または API-version のエラーなしに backend がロードされ、音声を返すことを確認します。

LocalAI version:
quay.io/go-skynet/local-ai:master-gpu-nvidia-cuda-13 (sha256:02e823f5f25ddef3f4edb94a8960e506daa4f71595d7c75b041d52b6eaaf425b, LocalAI bcf196d)
Backend cuda13-vllm-omni from quay.io/go-skynet/local-ai-backends:latest-gpu-nvidia-cuda-13-vllm-omni (sha256:d97f6feee8effc5b71201e5d77eb934f746c4f2197fbf10736860d1df558eee3), installed 2026-09-11. Inside it: vllm 0.27.1, vllm_omni 0.27.0rc2.dev102+gc3f80502e.

Environment, CPU architecture, OS, and Version:
Docker Desktop 29.7.2 (Compose 5.5.1) on Windows 11, WSL2 kernel 6.18.33.2-microsoft-standard-WSL2, x86_64. AMD Ryzen 7 7800X3D, NVIDIA RTX 5070 Ti 16 GB, driver 616.92.

Describe the bug
vllm-omni-qwen3-tts-custom-voice from the gallery can't be loaded with the prebuilt cuda13 vllm-omni backend. There are three separate problems, and each one only shows up after the previous one is worked around:

1. ModuleNotFoundError: No module named 'vllm_omni'. The editable install points at the build path.
   backend/python/vllm-omni/install.sh installs vllm-omni with uv pip install -e .. The generated finder venv/lib/python3.12/site-packages/__editable___vllm_omni_*_finder.py hardcodes the build-time location: MAPPING = {'vllm_omni': '/vllm-omni/vllm-omni/vllm_omni'}, plus several NAMESPACES entries with the same prefix. direct_url.json is file:///vllm-omni/vllm-omni. At runtime the backend lives in /backends/cuda13-vllm-omni/, so the source is at /backends/cuda13-vllm-omni/vllm-omni/vllm_omni and the import resolves nothing. _makeVenvPortable in backend/python/common/libbackend.sh fixes the python symlinks, pyvenv.cfg and shebangs, but not editable finders. Rewriting the paths in the finder makes the import work.
   This is probably the root cause of the No module named 'vllm_omni' part of #9162, reported there for cuda12. backend/python/fish-speech/install.sh also uses pip install -e, so it likely has the same problem.

2. ModuleNotFoundError: No module named 'vllm_omni.utils.platform_utils'.
   backend.py line 43 does from vllm_omni.utils.platform_utils import detect_device_type, is_npu. The vllm-omni source bundled in the image doesn't have that module anymore: vllm_omni/utils/ contains audio.py, custom_voice_io.py, forced_aligner.py, mm_outputs.py, … and there is a new vllm_omni/platforms/ package. Neither function is defined anywhere in the bundled source.

3. ValueError: `stage_configs_path` is no longer supported; use `deploy_config` instead.
   After shimming the two helpers above, loading fails here. backend.py (around line 225) always passes stage_configs_path (default None) to Omni(...) for llm/tts model types.

Problems 2 and 3 have the same underlying cause: install.sh clones https://github.com/vllm-project/vllm-omni.git without pinning a tag or commit. Each backend build therefore picks up whatever main is at that moment, and main has moved past the API that backend.py targets.

To Reproduce

1. Run quay.io/go-skynet/local-ai:master-gpu-nvidia-cuda-13 with an NVIDIA GPU.
2. Install vllm-omni-qwen3-tts-custom-voice from the model gallery. This pulls the cuda13-vllm-omni backend.
3. curl http://localhost:8080/v1/audio/speech -H "Content-Type: application/json" -d '{"model":"vllm-omni-qwen3-tts-custom-voice","input":"Hello"}'

Expected behavior
The model loads and returns audio.

Logs

ERROR Failed to load model modelID="vllm-omni-qwen3-tts-custom-voice" error=failed to load model with internal loader: grpc service not ready: backend process exited with code 1: ModuleNotFoundError: No module named 'vllm_omni' backend="vllm-omni"

after fixing the finder paths:

WARN  Backend process exited unexpectedly id="vllm-omni-qwen3-tts-custom-voice" process="run.sh" exitCode="1" stderr="ModuleNotFoundError: No module named 'vllm_omni.utils.platform_utils'"

after shimming platform_utils:

ERROR Failed to load model modelID="vllm-omni-qwen3-tts-custom-voice" error=failed to load model with internal loader: could not load model (no success): Unexpected err=ValueError('`stage_configs_path` is no longer supported; use `deploy_config` instead.'), type(err)=<class 'ValueError'> backend="vllm-omni"

Additional context
Possible fixes:

• Pin vllm-omni in install.sh to a tag or commit that backend.py is tested against, and bump it deliberately.
• Install it non-editable (uv pip install .), or make _makeVenvPortable rewrite editable finder paths to the runtime backend dir. Check other backend/python/*/install.sh for -e installs (at least fish-speech).
• Update backend.py for the current vllm-omni API (vllm_omni.platforms, deploy_config) if moving forward instead of pinning back.

Related: #9162, #9293