如何在Snowflake UI外（VS Code）通过PyTorch容器服务运行Python代码？

你可以通过以下两种方式，在Snowflake Notebook UI外调用同一GPU计算池运行PyTorch代码：

方法1：Snowpark Python API提交作业 #

直接用Snowpark的Container Services API在VS Code里编写代码提交GPU任务：

from snowflake.snowpark import Session
from snowflake.snowpark.container import ContainerService

# 初始化Snowflake会话
session = Session.builder.configs({
    "account": "你的账户标识",
    "user": "用户名",
    "password": "密码",
    "role": "有权限的角色",
    "warehouse": "关联仓库",
    "database": "目标数据库",
    "schema": "目标模式"
}).create()

# 定义GPU容器作业配置
container_job = ContainerService(
    session=session,
    name="PYTORCH_GPU_TASK",
    compute_pool="你的自定义GPU计算池名称",
    # 用Snowflake官方PyTorch镜像或自定义镜像
    image="snowflake.azurecr.io/snowflake/ml-python-pytorch:latest",
    spec={
        "containers": [
            {
                "name": "pytorch-runner",
                "command": ["python", "@你的代码Stage路径/script.py"],
                "resources": {
                    "requests": {"cpu": "4", "memory": "16Gi", "nvidia.com/gpu": "1"},
                    "limits": {"cpu": "4", "memory": "16Gi", "nvidia.com/gpu": "1"}
                }
            }
        ]
    }
)

# 启动作业并查看日志
container_job.create()
container_job.start()
print(container_job.get_logs())

session.close()

注意：提前将你的PyTorch脚本上传到Snowflake Stage中。

方法2：Snowflake CLI命令行提交 #

在VS Code终端用Snowflake CLI直接提交任务：

1. 安装CLI：pip install snowflake-cli-labs
2. 配置连接：运行snow configure按提示输入账户信息
3. 创建作业配置文件pytorch_gpu_job.yml：

name: PYTORCH_GPU_TASK
compute_pool: YOUR_GPU_COMPUTE_POOL_NAME
spec:
  containers:
    - name: pytorch-runner
      image: snowflake.azurecr.io/snowflake/ml-python-pytorch:latest
      command: ["python", "@code_stage/your_script.py"]
      resources:
        requests:
          cpu: "4"
          memory: "16Gi"
          nvidia.com/gpu: "1"
        limits:
          cpu: "4"
          memory: "16Gi"
          nvidia.com/gpu: "1"
stage_path: @code_stage

4. 提交作业：snow container services create --file pytorch_gpu_job.yml
5. 查看状态和日志：

   snow container services status --name PYTORCH_GPU_TASK
   snow container services logs --name PYTORCH_GPU_TASK
   

核心注意事项 #

• 确保你的计算池是GPU类型（创建时选择了GPU实例规格）
• 自定义镜像需上传到Snowflake私有镜像仓库
• 操作账户需拥有CREATE CONTAINER SERVICE、计算池USAGE等权限

内容的提问来源于stack exchange，提问作者B_Miner