Add support for `amd-smi` to perform device isolation #85

	name: OnnxRuntime CUDA Training 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=11.8.0
	--build-arg TORCH_CUDA=cu118
	--tag opt-bench-cuda:11.8.0
	.

	- name: Run tests
	run: docker run
	--rm
	--gpus '"device=0,1"'
	--entrypoint /bin/bash
	--volume $(pwd):/workspace/optimum-benchmark
	--workdir /workspace/optimum-benchmark
	--env TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0 7.5 8.0 8.6+PTX"
	--env USE_CUDA="1"
	opt-bench-cuda:11.8.0
	-c "pip install -e .[test,onnxruntime-training,peft] && python -m torch_ort.configure && pytest -k 'cuda and onnxruntime and training' -x"

Provide feedback