rocm-build
rocm-build copied to clipboard
xuhuisheng
rocm-build/navi10 for Stable Diffusion (for example: AUTOMATIC1111/stable-diffusion-webui)
Hello! I don't know much about coding and compilations. I don’t understand if it’s possible to use your instructions to make the rx5700 (navi10) card work in Stable Diffusion. Can you help me figure it out? Or maybe there is an opportunity to make a patch that will add support for rx5700 card in the AUTOMATIC1111 build?
I am afraid there is no guarantee for navi10 run properly on ROCm, as AMD didn't test navi10 in there development environment totally.
My patch is just let ROCm can run mnist with tensoflow/pytorch on navi10. But there is more easier way to do that, we can use an environment variable export HSA_OVERRIDE_GFX_VERSION=10.3.0 to pretend navi10 is a navi21 card, to let tensoflow-rocm/pytorch run on it. It is surprised mnist run sucessfully.
Emm~, rebuild may be more stable, but HSA_OVERRIDE way is more easier.
I'm not familar with stable diffussion, if it can run with tensorflow or pytorch, you can try HSA_OVERRIDE way.
Or, could you provide some instructions for using stable diffusion, that I can have a try when I have time. Now I am a little busy to test gfx803 on blender.
I can only say that I acted according to the instructions from this page. I used "running inside docker" section. (I'm on Ubuntu 20.04)
Running with "--no-half" alias enables Stable Diffusing without using GPU, CPU only mode (which means very slow generation)
Unfortunately, I test navi10 - RX5700xt hanging on ROCm-5.3.
So I guess navi10 may not support stable-diffusion.
Unfortunately, I test navi10 - RX5700xt hanging on ROCm-5.3.
So I guess navi10 may not support stable-diffusion.
with export HSA_OVERRIDE_GFX_VERSION=10.3.0 and other 2 parameters described in the repo 'SD' , the SD could run under GPU mode and its speed is much faster than the one under Windows.
Interesting, Cgmb said the ISA of RDNA1 is different with RDNA2, so RX5700xt cannot use 10.3.0 directly.
Unfortunately, I test navi10 - RX5700xt hanging on ROCm-5.3. So I guess navi10 may not support stable-diffusion.
with export HSA_OVERRIDE_GFX_VERSION=10.3.0 and other 2 parameters described in the repo 'SD' , the SD could run under GPU mode and its speed is much faster than the one under Windows.