Inference fails with 16GB GPU (even on low VRAM mode)

[omar-yasser-ayoub]

Hello! I'm having issues using the model with my GPU.

This is the command I am running: python run.py tree.png --output-dir output/ --low-vram-mode

Using CPU mode, the model runs fine and outputs the expected mesh.

However, using my GPU, I get the following:

Device used:  cuda
Loading in low VRAM mode
Settings -> Mode=base, Device=cuda, Torchscript=disabled
  0%|                                                                            | 0/1 [00:00<?, ?it/s]
Memory access fault by GPU node-1 (Agent handle: 0x2760f520) on address 0x7cab71237000. Reason: Page not present or supervisor privilege.
Aborted (core dumped)

looking at rocm-smi, I see that there is 37% VRAM usage (around 6GB as the documentation states with low VRAM mode), however the usage quickly spikes up to 100% and crashes.

These are my specs:

PyTorch Version: 2.6.0+rocm6.3.3.git88b97641
ROCm Version: 6.3.42134-a9a80e791
GPU: AMD Radeon RX 7800 XT
CPU: AMD Ryzen 7 7800x3d
Memory: 32GB

Worth noting I do have my iGPU disabled from BIOS, so there's only one GPU detected in rocminfo

Not sure if there is any more information I should be giving, but thank you for taking the time to read :)