Need help with the installation (AMD Radeon RX 580 (8GB))

[snake-py]

SO I have been trying for sometime now to run this on my machine and always I run into

HIP out of memory. Tried to allocate 576.00 MiB (GPU 0; 7.98 GiB total capacity; 6.72 GiB already allocated; 394.00 MiB free; 7.51 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_HIP_ALLOC_CONF

SO I am wondering if my gpu is too weak. I read that the minimum requirement would be 4gb.

The odd thing is that even without generating any images my gpu memory skyroeckets only when using the webui

 Device 0 [AMD Radeon RX Vega] PCIe GEN 3@16x RX: 7.824 GiB/s TX: 73.14 GiB/s
 GPU 862MHz  MEM 800MHz  TEMP  51°C FAN   0% POW  20 / 185 W
 GPU[                               1%] MEM[|||||||||||||||||7.314Gi/7.984Gi]

I am trying to run this on Ubuntu. I tried the following installations:

• automatic
• natively
• docker

I thought I need to install romc (which I did). Then I was hopinmg the docker container will solve this for me. BTW the image is super huge! I have right now only this image downloaded and only the container build for SD webui

TYPE            TOTAL     ACTIVE    SIZE      RECLAIMABLE
Images          61        1         183.7GB   161.5GB (87%)
Containers      1         1         209.7MB   0B (0%)
Local Volumes   16        0         3.239GB   3.239GB (100%)
Build Cache     437       0         21.56GB   21.56GB

version: '3'
services:
    rocm-pytorch:
        command: tail -F anything # used so the container keeps running and i can debug inside
        # command: python launch.py --precision full --no-half
        network_mode: host
        build:
            context: .
            dockerfile: Dockerfile
        devices:
            - /dev/kfd
            - /dev/dri
        group_add:
            - video
        ipc: host
        cap_add:
            - SYS_PTRACE
        security_opt:
            - seccomp=unconfined
        volumes:
            - .:/dockerx/stable-diffusion-webui

FROM rocm/pytorch:latest

WORKDIR /dockerx/stable-diffusion-webui
COPY . .
RUN pip install -r requirements.txt

[dilectiogames]

yes, that is an "out of Vram" error. but your card specs allows you to use SD without issues.

it never worked or you get this error in certain situations?

if it never worked i suggest deleting the installation and trying again to install it then don't install any extension and hit the generate button without any prompt or change in options.

you can backup the models, loras, stuff like that, no need to download them again.

[snake-py]

okay so played around with it a little more and sometimes it works. It seems to only work with the default model and with a width and height of 512. I wanted to test out this model here https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 but as soon as I add this one my ram usage exploads. Only loading it pushes my gpu to 7.5gb used.

To me it seems with the default model I cannot create decent results. It is probably normal that some models work and some doesn't?

[dilectiogames]

add this tag in your webui-user.bat

set COMMANDLINE_ARGS=--lowvram

this should fix the out of memory errors when using the SDXL model. This will double the generation time so remove that lowvram option if you want to use normal SD 1.2 and 2 models

for reference I have these options

PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.8,max_split_size_mb:256
set COMMANDLINE_ARGS=--xformers --lowvram --theme dark

themedark will make automatic be the dark version to stop bleaching your eyes

the garbage collection and max split size, it may work without but if you still get errors use them. those are required if you want to push the image size to the limits. it will hit the performance. the max split size especially. lower it goes longer ti takes

for example if the image needs 384Mb and you have 394Mb free, if I remember correctly, automatic will try to use 512 block by default. 512 can't fit in 394, you get the error

if is set to 256 then it will try to use

256 + 256 and you get an out of error again

but if max_split_size_mb is set to 128 ( you needed the number to be multiple of 16, so 32, 64, 128) (32 and 64 are too extreme)

then you have 128 + 128 + 128 you get 384 and can fit in 394. success

but speed gets a hit.

the lowvram will reduce the space used by SDXL model so you get more vram to fit the images in

also add the refiner, it will look bad without it, do you have this file: sd_xl_refiner_1.0.safetensors?

[AlexeyJersey]

try this settings (im on RX5700XT)

set COMMANDLINE_ARGS= --opt-sub-quad-attention --medvram --disable-nan-check --precision full --no-half --theme dark

-OR-

set COMMANDLINE_ARGS= --medvram --listen --share --no-gradio-queue --opt-sub-quad-attention --sub-quad-q-chunk-size 256 --sub-quad-kv-chunk-size 256 --sub-quad-chunk-threshold 70 --theme dark

1st is my stable setup, which works fine, but sometimes getting 'memory issue' on hires settings or some models (like juggernautXL)
2nd - test set up, but seems like will be my stable setup as it show good results with less 'memory issues'