GPU mode fails on RTX 5070 / Blackwell / compute capability 12.0

[vanishrap]

Environment:

OS: Windows 11
Docker Desktop: 29.4.3
Python: 3.14
GPU: NVIDIA GeForce RTX 5070 Laptop GPU
GPU compute capability: 12.0

Reproduction steps:

1. Install File Brain with pip install -U file-brain.
2. Start Docker Desktop.
3. Run file-brain.
4. Let File Brain auto-detect GPU mode.
5. Continue setup until Typesense / collection creation.

Actual result:

File Brain starts the GPU Typesense image:
hamza5/typesense-gpu:29.0-cuda11.8.0-cudnn8-runtime-ubuntu22.04

Typesense starts and loads the embedding model, then fails during CUDA / ONNX execution with:
cudaErrorNoKernelImageForDevice: no kernel image is available for execution on the device

After this, the Typesense health endpoint hangs:
curl.exe http://127.0.0.1:8108/health

File Brain then reports collection/stat errors such as:
Error getting stats: Collection not found
Error getting file type distribution: Collection not found

Expected result:

File Brain should not enable GPU mode if the current CUDA image is incompatible with the detected GPU.

For RTX 50-series / Blackwell / compute capability 12.0, File Brain should either use a compatible CUDA 12.8+ GPU image, automatically fall back to CPU mode, or show a clear compatibility warning.

Workaround:

Force CPU mode with:
setx FILEBRAIN_GPU_MODE force-cpu

Then restart PowerShell, remove old containers/data, and rerun File Brain.

The CPU image works:
typesense/typesense:29.0

Health check returns:
{"ok":true}

Suspected cause:
The GPU image uses CUDA 11.8, but RTX 5070 / Blackwell requires newer CUDA support for compute capability 12.0.