install python based backends in v4.0.0 nvidia-12 docker image not successful.

LocalAI version:
4.0.0

Environment, CPU architecture, OS, and Version:
docker, Xeon, ubuntu, 22.04

Describe the bug
uv not installed in cuda-12 images in the recent localai images. So installation of python based backends seem all not completely finished, and module missing error always occur when using models requiring those backends. I tried install uv, but after successful installation, running the install.sh in backend path still throw out error, compaining about not able to locate venv at /backendname/venv

also in the backend cuda12-qwen-asr install.sh does not have proper +x permission
and model vllm-omni-qwen3-tts-custom-voice complain about no context model.

To Reproduce
docker run local-ai:latest-gpu-nvidia-cuda-12 images, install
vllm-omni-qwen3-tts-custom-voice test in webui to test to audio,
debug log show vllm-omni module not found, it is refering to python module, the backend is installed during model installation, and
Expected behavior
model run successfully

Logs



Mar 29 02:40:59 DEBUG context local model name not found, setting to the first model first model name="vllm-omni-qwen3-tts-custom-voice" caller={caller.file="/build/core/http/middleware/request.go" caller.L=115 }

Mar 29 02:40:59 DEBUG LocalAI TTS Request received model="vllm-omni-qwen3-tts-custom-voice" caller={caller.file="/build/core/http/endpoints/localai/tts.go" caller.L=38 }

Mar 29 02:40:59 INFO [WatchDog] LRU evicting model model="vllm-omni-qwen3-tts-custom-voice" lastUsed=0001-01-01T00:00:00Z busy=false caller={caller.file="/build/pkg/model/watchdog.go" caller.L=321 }

Mar 29 02:40:59 DEBUG Model not found model="vllm-omni-qwen3-tts-custom-voice" caller={caller.file="/build/pkg/model/process.go" caller.L=27 }

Mar 29 02:40:59 ERROR [WatchDog] error shutting down model during LRU eviction error=model not found model="vllm-omni-qwen3-tts-custom-voice" caller={caller.file="/build/pkg/model/watchdog.go" caller.L=333 }

Mar 29 02:40:59 DEBUG [WatchDog] LRU eviction complete model="vllm-omni-qwen3-tts-custom-voice" caller={caller.file="/build/pkg/model/watchdog.go" caller.L=335 }

Mar 29 02:40:59 INFO BackendLoader starting modelID="vllm-omni-qwen3-tts-custom-voice" backend="vllm-omni" model="Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" caller={caller.file="/build/pkg/model/initializers.go" caller.L=157 }

Mar 29 02:40:59 DEBUG Loading Model with gRPC modelID="vllm-omni-qwen3-tts-custom-voice" file="/models/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice" backend="vllm-omni" options={vllm-omni Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice vllm-omni-qwen3-tts-custom-voice {{}} 0x10c6dc5f2c08 map[] 20 2 false} caller={caller.file="/build/pkg/model/initializers.go" caller.L=51 }

Mar 29 02:40:59 DEBUG Loading external backend uri="/backends/cuda12-vllm-omni/run.sh" caller={caller.file="/build/pkg/model/initializers.go" caller.L=75 }

Mar 29 02:40:59 DEBUG Loading GRPC Process process="/backends/cuda12-vllm-omni/run.sh" caller={caller.file="/build/pkg/model/process.go" caller.L=124 }

Mar 29 02:40:59 DEBUG GRPC Service will be running id="vllm-omni-qwen3-tts-custom-voice" address="127.0.0.1:33347" caller={caller.file="/build/pkg/model/process.go" caller.L=126 }

Mar 29 02:40:59 DEBUG GRPC stdout id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line="Initializing libbackend for cuda12-vllm-omni" caller={caller.file="/build/pkg/model/process.go" caller.L=174 }

Mar 29 02:40:59 DEBUG GRPC stdout id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line="Using portable Python" caller={caller.file="/build/pkg/model/process.go" caller.L=174 }

Mar 29 02:40:59 DEBUG GRPC stdout id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line="Added /backends/cuda12-vllm-omni/lib to LD_LIBRARY_PATH for GPU libraries" caller={caller.file="/build/pkg/model/process.go" caller.L=174 }

Mar 29 02:41:01 DEBUG GRPC stderr id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line="Traceback (most recent call last):" caller={caller.file="/build/pkg/model/process.go" caller.L=165 }

Mar 29 02:41:01 DEBUG GRPC stderr id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line=" File "/backends/cuda12-vllm-omni/backend.py", line 31, in " caller={caller.file="/build/pkg/model/process.go" caller.L=165 }

Mar 29 02:41:01 DEBUG GRPC stderr id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line=" from vllm_omni.entrypoints.omni import Omni" caller={caller.file="/build/pkg/model/process.go" caller.L=165 }

Mar 29 02:41:01 DEBUG GRPC stderr id="vllm-omni-qwen3-tts-custom-voice-127.0.0.1:33347" line="ModuleNotFoundError: No module named 'vllm_omni'" caller={caller.file="/build/pkg/model/process.go" caller.L=165 }

root@61b5c4e955d1:/# uv
bash: uv: command not found
root@61b5c4e955d1:/# cd backends/
root@61b5c4e955d1:/backends# ls
cuda12-llama-cpp cuda12-qwen-tts cuda12-vllm-omni qwen-asr transformers
cuda12-qwen-asr cuda12-transformers llama-cpp qwen-tts vllm-omni
root@61b5c4e955d1:/backends# cd cuda12-vllm-omni/
root@61b5c4e955d1:/backends/cuda12-vllm-omni# ./install.sh
Initializing libbackend for cuda12-vllm-omni
Using portable Python
starting requirements install for /backends/cuda12-vllm-omni/requirements.txt
./common/libbackend.sh: line 439: uv: command not found
root@61b5c4e955d1:/backends/cuda12-vllm-omni#

Additional context

Start by reproducing the Docker issue with the CUDA 12 image and inspect common/libbackend.sh, the backend install scripts, and cuda12-vllm-omni/backend.py. Check the reported uv, virtual-environment, executable-permission, and vllm_omni import failures across the affected backends. Done means the backend installation completes in the image and the reported model runs successfully.