Please describe the feature you want

When running Tabby on a CPU-only host, the official Docker image still attempts to launch the CUDA-based llama-server process for embeddings.
This causes continuous errors such as:

error while loading shared libraries: libcuda.so.1: cannot open shared object file: No such file or directory

Currently, CPU users must manually override the container’s entrypoint to /opt/tabby/bin/tabby-cpu, which bypasses the default startup logic but breaks the intended Docker UX.

Feature request:
Please add a way to run Tabby’s Docker image in CPU-only mode without errors — for example:

• Support an environment variable like CPU_ONLY=true that skips launching llama-server, or
• Add a dedicated registry.tabbyml.com/tabbyml/tabby:cpu tag, or
• Automatically detect missing CUDA and gracefully start tabby-cpu instead of tabby.

This would make CPU-only deployments first-class and eliminate the libcuda.so.1 crash loop.

Additional context

• Affected image: registry.tabbyml.com/tabbyml/tabby:latest

• Host environment: Linux VPS, no NVIDIA GPU

• The issue reproduces with default compose and resolved only by forcing the entrypoint:

  entrypoint: ["/opt/tabby/bin/tabby-cpu"]

• Example log snippet:

  /opt/tabby/bin/llama-server: error while loading shared libraries: libcuda.so.1
  

This change would improve accessibility for developers hosting Tabby on small or shared servers without GPUs.

Please reply with a 👍 if you want this feature.