vllm/.buildkite/run-amd-test.sh

46 lines
1.0 KiB
Bash

# This script runs test inside the corresponding ROCm docker container.
set -ex
# Print ROCm version
echo "--- ROCm info"
rocminfo
echo "--- Resetting GPUs"
echo "reset" > /opt/amdgpu/etc/gpu_state
while true; do
sleep 3
if grep -q clean /opt/amdgpu/etc/gpu_state; then
echo "GPUs state is \"clean\""
break
fi
done
echo "--- Building container"
sha=$(git rev-parse --short HEAD)
image_name=rocm_${sha}
container_name=rocm_${sha}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)
docker build \
-t ${image_name} \
-f Dockerfile.rocm \
--progress plain \
.
remove_docker_container() {
docker rm -f ${container_name} || docker image rm -f ${image_name} || true
}
trap remove_docker_container EXIT
echo "--- Running container"
docker run \
--device /dev/kfd --device /dev/dri \
--network host \
--rm \
-e HF_TOKEN \
--name ${container_name} \
${image_name} \
/bin/bash -c "${@}"