Memory access fault by GPU node-4 (Agent handle: 0x5618a9f81270) on address 0x7fd1000e5000. Reason: Page not present or supervisor privilege

Question

I am getting this error trying to run pytorch using rocm on a Radeon VII. I am trying to run disco-diffusion

Answer 1

This error I guess is the application using more vram than your gpu have,

I am using radeon 5700xt, and using Tensorflow_rocm, and encounter "Memory access fault by GPU node-1" error,

I just fix this error, by pretending I am using radeon 580, by runing the command “export HSA_OVERRIDE_GFX_VERSION=8.3.0” before launching python. (GFX830 is the offical model name of radeon 580)

If it is not work for you, try another model name. Maybe it helps, good luck!!