Skip to main content

uniflow.plugins.ray.run_config

Ray Train RunConfig helper defaulted to UniFlow-managed storage.

Any Ray-based workflow task (trainer, and future tasks with their own Ray Train steps) can call create_run_config instead of hand-rolling its own storage_path/storage_filesystem defaulting from a task-specific storage_backend parameter. Centralizing this here keeps "where does Ray Train write checkpoints" in one shared place as the task catalog grows, rather than each task re-deriving it independently.

create_run_config​

def create_run_config(**kwargs) -> ray.train.RunConfig

Build a ray.train.RunConfig defaulted to UniFlow-managed storage.

Resolves storage_path/storage_filesystem from the same UF_STORAGE_URL environment variable that DatasetVariable and ModelVariable already use for their own storage location, via the same filesystem-resolution logic RayDatasetIO uses (native PyArrow S3, or fsspec when UF_PLUGIN_RAY_USE_FSSPEC=1). Falls back to a local temp directory when UF_STORAGE_URL is unset, so local/sandbox runs keep working without extra configuration.

Arguments:

  • **kwargs - Any ray.train.RunConfig keyword argument. Explicitly passing storage_path and/or storage_filesystem overrides the UF_STORAGE_URL-derived default for that field.

Returns:

A ray.train.RunConfig with storage_path/storage_filesystem defaulted as described above.