fix(local-ai): enable NVIDIA GPU passthrough

[bitsandbots]

Summary

• The local-ai community container only requests /dev/dri, which is the Mesa/AMD/Intel Vulkan device passthrough mechanism.
• On NVIDIA-only hosts (no Mesa-compatible device), Vulkan has nothing to bind to and every backend (llama-cpp, stablediffusion-ggml, whisper) silently falls back to llvmpipe (CPU-only software rendering).
• This makes image generation in particular unusably slow — CPU-based Stable Diffusion routinely took 2+ minutes per image in my testing, exceeding integration_openai's image_request_timeout (240s default) and causing Assistant sticker/image tasks to fail outright.
• plex, jellyfin, and memories already solve this for their own GPU-accelerated use cases via "enable_nvidia_gpu": true, which DockerActionManager turns into Runtime: nvidia + the correct DeviceRequests when the host operator has set NEXTCLOUD_ENABLE_NVIDIA_GPU=true on the mastercontainer. local-ai.json was missing this flag.

Change

• Add "enable_nvidia_gpu": true to community-containers/local-ai/local-ai.json, alongside the existing "devices": ["/dev/dri"] (same pattern as memories.json, so Mesa-only hosts are unaffected and NVIDIA hosts now get real acceleration).

Testing

Reproduced and fixed on an NVIDIA RTX 5070 host:

• Before: vulkaninfo inside the container showed only llvmpipe (PHYSICAL_DEVICE_TYPE_CPU); a 512x512 SD1.5 image generation took 140s+ (cold) and still exceeded the timeout under normal load.
• After (host operator sets NEXTCLOUD_ENABLE_NVIDIA_GPU=true on the mastercontainer + has nvidia-container-toolkit installed/configured): vulkaninfo correctly reports the NVIDIA GeForce RTX 5070 (PHYSICAL_DEVICE_TYPE_DISCRETE_GPU), and the same image generation completed in 4.5s once the backend was warm, with nvidia-smi confirming the stablediffusion-ggml process resident on the GPU.
• Verified local-ai.json still validates against php/containers-schema.json (the enable_nvidia_gpu boolean property already exists in the schema, used by the other GPU-enabled community containers).

Notes

• This is opt-in: nothing changes unless the host operator explicitly sets NEXTCLOUD_ENABLE_NVIDIA_GPU=true on the mastercontainer (per the existing hardware-acceleration docs in the main readme) and has the NVIDIA Container Toolkit installed. Mesa/AMD/Intel-only hosts are unaffected since /dev/dri passthrough is untouched.

🤖 Drafted with Claude Code

[szaimen]

@bitsandbots did you test and verified that this works locally?