>>109178183
This is how I launch comfyui. Your exact command will be different. Like will you be using conda instead of venv? idk how things work on Windows, but I think this stuff is in the bat file.
source ~/comfy-env/bin/activate
TORCH_BLAS_PREFER_HIPBLASLT=0 PYTORCH_ALLOC_CONF=expandable_segments:True TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 MIOPEN_FIND_MODE=FAST CUDA_LAUNCH_BLOCKING=1 HSA_OVERRIDE_GFX_VERSION=10.3.0 python ~/ComfyUI/main.py --lowvram --preview-method latent2rgb --disable-dynamic-vram
You definitely won't use hsa override. I need it because the 6950xt isn't officially supported. I think on --lowvram and --disable-dynamic-vram are relevant to you pretty much, but showing you how it looks on Linux, in case you go that route.
Like the --preview-method latent2rgb is a fast preview method, but it doesn't look as good as taesd, which requires that you install the relevant file for it.
comfyui usage that is nice needs to set the output folder and file based on date. Otherwise, it's a mess.
Always remember that Google search ai is pretty good and free and knows everything about your gpu and comfyui.