49RpK5dY

Try adding HSA_OVERRIDE_GFX_VERSION=10.3.0 in launch command. Work on RX 5700 when launching with `TORCH_COMMAND='pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/rocm5.1.1' HSA_OVERRIDE_GFX_VERSION=10.3.0 python launch.py --precision full --no-half --medvram` `export HSA_OVERRIDE_GFX_VERSION=10.3.0` also works

@Veyrdite Have you tried deleting `venv` dir and reinstalling with `HSA_OVERRIDE_GFX_VERSION=10.3.0`? Also, running it inside docker might solve the issue with No HIP GPUs are available. I managed to get...

> But tried to follow [#1122 (comment)](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/1122#issuecomment-1261334452) but also get an error in the last step for amdgpu-install That seems like a very convoluted way of solving a problem that...

> Is there a chance I need to install more on my host system? It works without any ROCm packets on my system. I assume it should work on yours...

> It still required me to use COMMANDLINE_ARGS='--skip-torch-cuda-test' Try launching/reinstalling with `TORCH_COMMAND='pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/rocm5.1.1' HSA_OVERRIDE_GFX_VERSION=10.3.0 python launch.py --precision full --no-half --medvram`. It no longer requires passing `REQS_FILE='requirements.txt'`,...

@capitalistspz Try deleting and recreating venv up until `python -m pip install --upgrade pip wheel` and after that launching it with a script ``` #!/bin/bash cd $HOME/stable-diffusion-webui source venv/bin/activate export...

I noticed that black bars appear if the input image dimensions are not divisible by 128. I assume this happens because it is first downsampled to half of the original...

Changing `down_sample_rate = target_scale / 4` to `down_sample_rate = target_scale / 2` in [ldsr_model_arch.py](https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/modules/ldsr_model_arch.py) worked as it no longer produces a black bar on 512x832 input image with 2x upscale...

I retried `rocm/pytorch` again and decided to wait. It's been over an hour already since it started generation and still stuck at `0%| | 0/20 [00:00

Well I guess it's just me. No need to keep the issue open.