Add xformers for ROCm support

[Looong01]

Description

Add xformers for ROCm support

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[w-e-w]

I'm not familiar with rocm or AMD GPUs so this comment is based out of knowledge of the code base and not about specifics

---

similer to TORCH_COMMAND

stable-diffusion-webui/webui.sh

Lines 130 to 175 in 0120768

	gpu_info=$(lspci 2>/dev/null | grep -E "VGA|Display")
	case "$gpu_info" in
	*"Navi 1"*)
	export HSA_OVERRIDE_GFX_VERSION=10.3.0
	if [[ -z "${TORCH_COMMAND}" ]]
	then
	pyv="$(${python_cmd} -c 'import sys; print(f"{sys.version_info[0]}.{sys.version_info[1]:02d}")')"
	# Using an old nightly compiled against rocm 5.2 for Navi1, see https://github.com/pytorch/pytorch/issues/106728#issuecomment-1749511711
	if [[ $pyv == "3.8" ]]
	then
	export TORCH_COMMAND="pip install https://download.pytorch.org/whl/nightly/rocm5.2/torch-2.0.0.dev20230209%2Brocm5.2-cp38-cp38-linux_x86_64.whl https://download.pytorch.org/whl/nightly/rocm5.2/torchvision-0.15.0.dev20230209%2Brocm5.2-cp38-cp38-linux_x86_64.whl"
	elif [[ $pyv == "3.9" ]]
	then
	export TORCH_COMMAND="pip install https://download.pytorch.org/whl/nightly/rocm5.2/torch-2.0.0.dev20230209%2Brocm5.2-cp39-cp39-linux_x86_64.whl https://download.pytorch.org/whl/nightly/rocm5.2/torchvision-0.15.0.dev20230209%2Brocm5.2-cp39-cp39-linux_x86_64.whl"
	elif [[ $pyv == "3.10" ]]
	then
	export TORCH_COMMAND="pip install https://download.pytorch.org/whl/nightly/rocm5.2/torch-2.0.0.dev20230209%2Brocm5.2-cp310-cp310-linux_x86_64.whl https://download.pytorch.org/whl/nightly/rocm5.2/torchvision-0.15.0.dev20230209%2Brocm5.2-cp310-cp310-linux_x86_64.whl"
	else
	printf "\e[1m\e[31mERROR: RX 5000 series GPUs python version must be between 3.8 and 3.10, aborting...\e[0m"
	exit 1
	fi
	fi
	;;
	*"Navi 2"*) export HSA_OVERRIDE_GFX_VERSION=10.3.0
	;;
	*"Navi 3"*) [[ -z "${TORCH_COMMAND}" ]] && \
	export TORCH_COMMAND="pip install torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm5.7"
	;;
	*"Renoir"*) export HSA_OVERRIDE_GFX_VERSION=9.0.0
	printf "\n%s\n" "${delimiter}"
	printf "Experimental support for Renoir: make sure to have at least 4GB of VRAM and 10GB of RAM or enable cpu mode: --use-cpu all --no-half"
	printf "\n%s\n" "${delimiter}"
	;;
	*)
	;;
	esac
	if ! echo "$gpu_info" | grep -q "NVIDIA";
	then
	if echo "$gpu_info" | grep -q "AMD" && [[ -z "${TORCH_COMMAND}" ]]
	then
	export TORCH_COMMAND="pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.7"
	elif npu-smi info 2>/dev/null
	then
	export TORCH_COMMAND="pip install torch==2.1.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu; pip install torch_npu==2.1.0"
	fi
	fi

feels like this should be done in webui.sh via XFORMERS_PACKAGE environment variable when not configured and trying to run rocminfo

potential issues with the current way you do things is that it might cause people that has multi gpus break
people could have an Nvidia and AMD GPU at the same time
they could have configure it wherever you are invite environment variables to use a certain GPU which may not be the one you expect, and so the test for rocminfo could pass but they're running on Nvidia
also if people have already configured XFORMERS_PACKAGE your code would ignore what they have configured if they rocminfo check passed

[Soulreaver90]

So I went ahead and tried it and see xformers loaded, but I don't see any speed improvements at all. Is there anything else that needs to be done or is rocm support still being worked on? I am on the latest ROCM and PyTorch.

Edit: I downgraded my PyTorch to the version with ROCM 6.1 since that is what xformers 0.0.28 was built against. Realized that no matter what I did, even adding --xformers, my cross attention would still default to doggettx. I added "--force-enable-xformers" and I now see xformers in the cross attention dropdown. However when selecting it and generating anything, I get errors thrown out left and right.
Has anyone confirmed xformers for ROCM works with A1111? Is it a xformers issue or a A1111?