And if you just run the 'rocminfo' command, does something appear? According to TechPowerUp, your GPU also uses the GFX803, so it should work normally.
The main problem is my cpu is i7 3770s. Would you tell me please which cpu and motherboard support atomics operation? Or what is your system configuration?
Oh, I know what's wrong. I compiled PyTorch for newer processors (for Intel CPUs, 4th generation or higher), so it attempts to utilize AVX2. However, your CPU does not support this particular instruction set. I will recompile the binaries today to accommodate processors without AVX2. Tomorrow, please check the GitHub link where you will find a new PyTorch build specifically designed for CPUs lacking AVX2 support.
SD 1.5 worked fine, but the XL model gives an error when it reaches 100% I think AUTOMATIC1111 is not good for XL, is it possible to use Forge or fooocus on almalinux?
I've never tried Forge, but Fooocus works. However, it needs some Torch packages that I haven't yet compiled for Polaris GPUs. I'll see what I can do. In a few days, take a look at the GitHub repository.
Hey, have you tried ComfyUI yet? Since you are looking for low VRAM consumption, it is the most optimized GUI for Stable Diffusion. It is a little more complex than the others, but it is worth it. I tested it here on my RX 550 4GB, and it generates images about 30% faster and consumes much less VRAM compared to SD WebUI. The only package you will need to install additionally is torchaudio v2.2.2 (I've already put it on GitHub).
@@Luinux-Tech I tried to install comfyUI, but I couldn't get it to start. I don't know what I'm doing wrong, is installing it on almalinux the same as on ubuntu?
Yes, if you have ROCm and amdgpu-dkms installed, just install Torch, TorchVision, and TorchAudio in your Python venv (Python version 3.9 or 3.12). Git clone ComfyUI and install the ComfyUI dependencies with pip, then just start the main.py file with your venv's python. What version of Ubuntu are you using?
Most XL models do work with 8GB of VRAM, but the recommended amount is 12GB or more. If you are going to use SDXL + LORAs, you will probably need more than 8GB. You can try to reduce VRAM consumption by using Tiled Diffusion & VAE.
It depends on the number of steps. The more steps you use, the more detailed the image will be. A good value is 30 steps. With your GPU and using 30 steps, you should get an image every 1-2 minutes.
thank you so much,it's save my life.
Glad it helped!
Thanks for the tutorial, any luck getting Ollama to work with gfx803 ?
Thanks, I haven't tested Ollama, but koboldcpp-rocm works perfectly.
i installed rocminfo from epel then run the code but it shows nothing. my GPU is rx580 2048sp. what should I do?
And if you just run the 'rocminfo' command, does something appear? According to TechPowerUp, your GPU also uses the GFX803, so it should work normally.
@@Luinux-Tech no it shows only cpu
Hmm, I don't know. Maybe rocminfo needs amdgpu-dkms. Try doing it up to step 4, then restart and test the "rocminfo" command again.
The main problem is my cpu is i7 3770s. Would you tell me please which cpu and motherboard support atomics operation? Or what is your system configuration?
Oh, I know what's wrong. I compiled PyTorch for newer processors (for Intel CPUs, 4th generation or higher), so it attempts to utilize AVX2. However, your CPU does not support this particular instruction set. I will recompile the binaries today to accommodate processors without AVX2. Tomorrow, please check the GitHub link where you will find a new PyTorch build specifically designed for CPUs lacking AVX2 support.
SD 1.5 worked fine, but the XL model gives an error when it reaches 100% I think AUTOMATIC1111 is not good for XL, is it possible to use Forge or fooocus on almalinux?
I've never tried Forge, but Fooocus works. However, it needs some Torch packages that I haven't yet compiled for Polaris GPUs. I'll see what I can do. In a few days, take a look at the GitHub repository.
Hey, have you tried ComfyUI yet? Since you are looking for low VRAM consumption, it is the most optimized GUI for Stable Diffusion. It is a little more complex than the others, but it is worth it. I tested it here on my RX 550 4GB, and it generates images about 30% faster and consumes much less VRAM compared to SD WebUI. The only package you will need to install additionally is torchaudio v2.2.2 (I've already put it on GitHub).
@@Luinux-Tech I tried to install comfyUI, but I couldn't get it to start. I don't know what I'm doing wrong, is installing it on almalinux the same as on ubuntu?
Yes, if you have ROCm and amdgpu-dkms installed, just install Torch, TorchVision, and TorchAudio in your Python venv (Python version 3.9 or 3.12). Git clone ComfyUI and install the ComfyUI dependencies with pip, then just start the main.py file with your venv's python. What version of Ubuntu are you using?
now i'm using almalinux 9!
I have an RX 580 8GB, will I be able to use XL models?
Most XL models do work with 8GB of VRAM, but the recommended amount is 12GB or more. If you are going to use SDXL + LORAs, you will probably need more than 8GB. You can try to reduce VRAM consumption by using Tiled Diffusion & VAE.
Rx 570 4gb ?
Yes, it works. Remember to use the parameters for GPUs with low VRAM.
Can I use windos
Try this: ua-cam.com/video/W75iBfnFmnU/v-deo.html
@@Luinux-Tech how long it takes to generate 4 picture with my gpu?
It depends on the number of steps. The more steps you use, the more detailed the image will be. A good value is 30 steps. With your GPU and using 30 steps, you should get an image every 1-2 minutes.