Ray支持在GPU上运行任务，这对于需要大量计算资源的应用（如深度学习）来说非常有用。本文是在火山引擎容器服务VKE上使用Ray运行GPU任务的步骤。

1 前提条件

• VKE集群运行正常。

• VKE集群中至少有一个节点配置了 NVIDIA GPU 并安装了相应的驱动程序。

• 部署EMR on VKE产品中的Ray服务

2 选择Ray镜像

在EMR on VKE产品中部署RayCluster组件时，采用的镜像不包含CUDA依赖的。

EMR中提供含CUDA依赖的Ray镜像，里面也包括TensorFlow 、PyTorch依赖，参考镜像仓库获取相关镜像。

3 Pod中配置GPU资源

在EMR管控平台中提供有RayCluster的安装部署，可以通过Yaml文件方式指定Ray镜像和GPU个数。对于RayJob方式，也可以通过Yaml文件方式。Yaml文件如下示例：

groupName: gpu-group
replicas: 0
minReplicas: 0
maxReplicas: 5
...
template:
    spec:
     ...
     containers:
      - name: ray-node
        image: emr-vke-public-cn-beijing.cr.volces.com/emr/ray:2.9.3-cu11.8.0-py3.9-ubuntu20.04-1.2.0
        ...
        resources:
         nvidia.com/gpu: 1 # Optional, included just for documentation.
         cpu: 3
         memory: 50Gi
        limits:
         nvidia.com/gpu: 1 # Required to use GPU.
         cpu: 3
         memory: 50Gi
         ...

示例中nvidia.com/gpu表明可以使用 Nvidia GPUs ，可以配置RayCluster的Yaml中headGroupSpec和workerGroupSpecs片段中。同时Ray Autoscaler特性也使用对GPU的弹性。

4 Ray作业中使用GPU资源

一旦部署了可以访问 GPU 的 Ray pod，就能够执行带有 GPU 请求的task和 actor。例如，通过@ray.remote(num_gpus=1) 注释 1 个 GPU 的 actor：

import ray

ray.init()

@ray.remote(num_gpus=1)
class GPUActor:
    def say_hello(self):
        print("I live in a pod with GPU access.")

# Request actor placement.
gpu_actors = [GPUActor.remote() for _ in range(2)]
# The following command will block until two Ray pods with GPU access are scaled
# up and the actors are placed.
ray.get([actor.say_hello.remote() for actor in gpu_actors])

# scale up GPU resources.
ray.autoscaler.sdk.request_resources(bundles=[{"GPU": 1}] * 2)

ray.autoscaler.sdk.request_resources(bundles=[])

示例中@ray.remote(num_gpus=1)表明actor需要一个GPU。

5 使用示例

"""Example of a custom gym environment and model. Run this for a demo.

This example shows:
  - using a custom environment
  - using a custom model
  - using Tune for grid search

You can visualize experiment results in ~/ray_results using TensorBoard.
"""
import argparse

import ray
from ray import air, tune
from ray.rllib.examples.env.gpu_requiring_env import GPURequiringEnv
from ray.rllib.utils.framework import try_import_tf, try_import_torch
from ray.rllib.utils.test_utils import check_learning_achieved
from ray.tune.registry import get_trainable_cls
from pyarrow import fs

tf1, tf, tfv = try_import_tf()
torch, nn = try_import_torch()

parser = argparse.ArgumentParser()
parser.add_argument(
    "--run", type=str, default="PPO", help="The RLlib-registered algorithm to use."
)
parser.add_argument(
    "--framework",
    choices=["tf", "tf2", "torch"],
    default="torch",
    help="The DL framework specifier.",
)
parser.add_argument("--num-gpus", type=float, default=0.5)
parser.add_argument("--num-workers", type=int, default=1)
parser.add_argument("--num-gpus-per-worker", type=float, default=0.0)
parser.add_argument("--num-envs-per-worker", type=int, default=1)
parser.add_argument(
    "--as-test",
    action="store_true",
    help="Whether this script should be run as a test: --stop-reward must "
    "be achieved within --stop-timesteps AND --stop-iters.",
)
parser.add_argument(
    "--stop-iters", type=int, default=50, help="Number of iterations to train."
)
parser.add_argument(
    "--stop-timesteps", type=int, default=100000, help="Number of timesteps to train."
)
parser.add_argument(
    "--stop-reward", type=float, default=180.0, help="Reward at which we stop training."
)

if __name__ == "__main__":
    args = parser.parse_args()
    ray.init()

    # These configs have been tested on a p2.8xlarge machine (8 GPUs, 16 CPUs),
    # where ray was started using only one of these GPUs:
    # $ ray start --num-gpus=1 --head

    # Note: A strange error could occur when using tf:
    #       "NotImplementedError: Cannot convert a symbolic Tensor
    #       (default_policy/cond/strided_slice:0) to a numpy array."
    #       In rllib/utils/exploration/random.py.
    #       Fix: Install numpy version 1.19.5.

    # Tested arg combinations (4 tune trials will be setup; see
    # tune.grid_search over 4 learning rates below):
    # - num_gpus=0.5 (2 tune trials should run in parallel).
    # - num_gpus=0.3 (3 tune trials should run in parallel).
    # - num_gpus=0.25 (4 tune trials should run in parallel)
    # - num_gpus=0.2 + num_gpus_per_worker=0.1 (1 worker) -> 0.3
    #   -> 3 tune trials should run in parallel.
    # - num_gpus=0.2 + num_gpus_per_worker=0.1 (2 workers) -> 0.4
    #   -> 2 tune trials should run in parallel.
    # - num_gpus=0.4 + num_gpus_per_worker=0.1 (2 workers) -> 0.6
    #   -> 1 tune trial should run in parallel.

    config = (
        get_trainable_cls(args.run)
        .get_default_config()
        # Setup the test env as one that requires a GPU, iff
        # num_gpus_per_worker > 0.
        .environment(
            GPURequiringEnv if args.num_gpus_per_worker > 0.0 else "CartPole-v1"
        )
        .framework(args.framework)
        .resources(
            # How many GPUs does the local worker (driver) need? For most algos,
            # this is where the learning updates happen.
            # Set this to > 1 for multi-GPU learning.
            num_gpus=args.num_gpus,
            # How many GPUs does each RolloutWorker (`num_workers`) need?
            num_gpus_per_worker=args.num_gpus_per_worker,
        )
        # How many RolloutWorkers (each with n environment copies:
        # `num_envs_per_worker`)?
        .rollouts(
            num_rollout_workers=args.num_workers,
            # This setting should not really matter as it does not affect the
            # number of GPUs reserved for each worker.
            num_envs_per_worker=args.num_envs_per_worker,
        )
        # 4 tune trials altogether.
        .training(lr=tune.grid_search([0.005, 0.003, 0.001, 0.0001]))
    )

    stop = {
        "training_iteration": args.stop_iters,
        "timesteps_total": args.stop_timesteps,
        "episode_reward_mean": args.stop_reward,
    }

    # Note: The above GPU settings should also work in case you are not
    # running via ``Tuner.fit()``, but instead do:

    # >> from ray.rllib.algorithms.ppo import PPO
    # >> algo = PPO(config=config)
    # >> for _ in range(10):
    # >>     results = algo.train()
    # >>     print(results)

    s3 = fs.S3FileSystem(access_key='xxx', secret_key='xxx',endpoint_override='xxx', force_virtual_addressing=True)

    results = tune.Tuner(
        args.run, param_space=config.to_dict(), run_config=air.RunConfig(stop=stop, storage_path="ceshi-dts/ray-store", storage_filesystem = s3)
    ).fit()

    if args.as_test:
        check_learning_achieved(results, args.stop_reward)
    ray.shutdown()

示例中是采用GPU执行torch模型训练，作业执行过程中的中间数据存储到对象存储TOS中，所以需要填写实际的access_key、secret_key和endpoint_override参数。

可以在Ray dashboard中可以GPU资源的使用情况。