Add support for `amd-smi` to perform device isolation #87

	name: Pytorch CUDA Tests

	on:
	workflow_dispatch:
	pull_request:
	types: [opened, reopened, synchronize]

	concurrency:
	group: ${{ github.workflow }}-${{ github.event.pull_request.number \|\| github.ref }}
	cancel-in-progress: true

	jobs:
	build_image_and_run_gpu_tests:
	runs-on: hf-dgx-01
	steps:
	- name: Checkout
	uses: actions/checkout@v2

	- name: Build image
	run: docker build
	--file docker/cuda.dockerfile
	--build-arg USER_ID=$(id -u)
	--build-arg GROUP_ID=$(id -g)
	--build-arg CUDA_VERSION=12.1.1
	--build-arg TORCH_CUDA=cu121
	--tag opt-bench-cuda:12.1.1
	.

	- name: Run tests
	run: docker run
	--rm
	--net host
	--pid host
	--shm-size 64G
	--env USE_CUDA="1"
	--entrypoint /bin/bash
	--volume $(pwd):/workspace/optimum-benchmark
	--workdir /workspace/optimum-benchmark
	--gpus '"device=0,1"'
	opt-bench-cuda:12.1.1
	-c "pip install -e .[test,peft,diffusers] && pytest -k 'cuda and pytorch' -x"

Provide feedback