CUDA: Check failed (999 vs. 0) on RTX 4050 - MXNet compatibility with CUDA 12.2?

[yukiofy]

Can I'm trying to run CompreFace with GPU support on my local machine using the compreface-core:1.2.0-arcface-r100-gpu and also tested with 1.2.0-mobilenet-gpu.

My system specs:

GPU: NVIDIA GeForce RTX 4050 Laptop GPU
Driver Version: 535.247.01
CUDA Version: 12.2
Host OS: Debian 12
Docker: version 28.1.1
NVIDIA Container Toolkit: installed and GPU is detected inside container (nvidia-smi works)

In the container logs (docker logs compreface-core), I get the following error:

CUDA: Check failed: e == cudaSuccess (999 vs. 0): unknown error ... RuntimeError: simple_bind error. Arguments: data: (1, 3, 480, 640)

Despite the container detecting the GPU (nvidia-smi runs inside the container), MXNet fails when calling simple_bind, which suggests a CUDA runtime mismatch.

What I've confirmed:
nvidia-smi inside the container works fine.

The runtime: nvidia is properly set in docker-compose.yml.

Both arcface-r100-gpu and mobilenet-gpu produce the same error.

The logs show MXNet trying to use CUDA 11.8.

My suspicion
Since my GPU (RTX 4050) requires CUDA >= 12, and the base images use CUDA 11.8, I believe there's an incompatibility between the MXNet version used in CompreFace and the required CUDA version for my GPU.

Questions
Is there an official version or plan to support CUDA 12.2+ for newer GPUs like RTX 4050?

Would recompiling MXNet with CUDA 12.2 solve this issue?

Do you provide a Dockerfile or guide to build compreface-core with a newer CUDA version?

Thank you for this amazing project, and I’d be happy to contribute back if needed.