yr.InvokeOptions.debug

yr.InvokeOptions.debug#

InvokeOptions.debug: DebugConfig#

Configure the stateful/stateless function runtime environment with conda, pip, working_dir, and env_vars.

  • conda provides different Python runtime environments for stateful function.

    • Specify an existing conda environment (the environment exists on all nodes) runtime_env = {"conda":"pytorch_p39"}

    • Create and use conda environments through configuration. runtime_env["conda"] = {"name":"myenv","channels": ["conda-forge"], "dependencies": ["python=3.9", "msgpack-python=1.0.5", "protobuf", "libgcc-ng", "cloudpickle=2.0.0", "cython=3.0.10", "pyyaml=6.0.2"]}

    • Create and use a conda environment through a YAML file (the YAML file meets the conda requirements). runtime_env = {"conda":"/home/env.yaml"}

  • pip installs dependencies for Python runtime environment.

  • working_dir configure the code path of the job.

  • env_vars configure process-level environment variables. runtime_env = {"env_vars":{"OMP_NUM_THREADS": "32", "TF_WARNINGS": "none"}}

  • shared_dir supports configuring a shared directory for some instance, with yr managing the lifecycle of this shared directory. shared_dir supports two fields: name and TTL. The name field only allows numbers, letters, “-”, and “_”. The TTL supports integers greater than 0 and less than INTMAX. runtime_env = {"shared_dir":{"name": "user_define", "TTL": 5}}

  • Constraints of runtime_env:

    • The keys supported by runtime_env are conda, env_vars, pip, working_dir. Other keys will not take effect and will not cause errors.

    • Run the yr function with conda. The environment needs to have yr and its third-party dependencies. It is recommended that users first create a conda environment and then specify it with runtime_env, for example: runtime_env = {"conda":"pytorch_p39"}

    • runtime_env supports creating and switching conda environments using configurations. The configuration needs to install third-party dependencies for yr, for example: runtime_env["conda"] = {"name":"myenv","channels": ["conda-forge"], "dependencies": ["python=3.9", "msgpack-python=1.0.5", "protobuf", "libgcc-ng", "cloudpickle=2.0.0", "cython=3.0.10", "pyyaml=6.0.2"]}

    • The environment created using conda in runtime_env needs to be cleaned up by the user.

    • In runtime_env, conda can use pip to install dependencies, which are managed directly by conda. runtime_env = {"conda":{'name': 'my_project_env', 'channels': ['defaults', 'conda-forge'], 'dependencies': ['python=3.9', {'pip': ['requests==2.25.1']}]}}

    • Currently, Python 3.9 and Python 3.11 SDKs are available. The Python version of conda needs to be consistent with the SDK version.

    • If both InvokeOptions.env_vars and InvokeOptions.runtime_env[“env_vars”] are configured with the same key, the configuration in InvokeOptions.env_vars will be used.

    • If InvokeOptions.runtime_env[“working_dir”] is configured, use this configuration, otherwise, use YR.Config.working_dir and finally use the configuration in InvokeOptions.env_vars.

    • If you use conda, you need to specify the environment variable YR_CONDA_HOME to point to installation path.

    • shared_dir has the following constraints: 1. It is not recommended to configure different TTL for the same shared directory. 2. The minimum cleanup interval for shared directories is 5 seconds. 3. When multiple yr Agents are deployed on the same node, each Agent must be configured with different root directory to prevent conflicts in shared directory management.