v0.2.5

Summary

Adds cache_scope to @huggingface_hub to control where Hugging Face caching mechanism, enabling sharing across runs, users, and flows while reducing redundant downloads and avoiding throttling. The cache key scheme for non-checkpoint scopes now considers revision, allow_patterns, and ignore_patterns to prevent collisions. See commit 467b430.

What’s new

• New decorator option to @huggingface_hub: cache_scope
  • Values: checkpoint (default), flow, global
  • Controls the base path used for HF caches written via @huggingface_hub
  • flow will cache all repos under the same flow regardless of namespace
  • global will cache all repos under a fixed path regardless of flow/namespace.
  • For checkpoint, cache key format is unchanged for backward compatibility
  • For flow and global, cache keys also include material parameters: revision, allow_patterns, ignore_patterns

Usage

from metaflow import huggingface_hub

# Default (per-step scope): unchanged behavior
@huggingface_hub
@step
def default_scope(self):
    from metaflow import current
    ref = current.huggingface_hub.snapshot_download(
        repo_id="mistralai/Mistral-7B-Instruct-v0.1",
        allow_patterns=["*.safetensors", "*.json"],
    )

# Share cache within a flow even across namespaces (even @project/branch)
@huggingface_hub(cache_scope="flow")
@step
def flow_scope(self):
    from metaflow import current
    ref = current.huggingface_hub.snapshot_download(
        repo_id="mistralai/Mistral-7B-Instruct-v0.1",
        allow_patterns=["*.safetensors", "*.json"],
    )

# Model cached accessed across flows (Accessible to any/all flow executions within metaflow environment)
@huggingface_hub(load=["mistralai/Mistral-7B-Instruct-v0.1"], cache_scope="global")
@step
def load_into_path(self):
    from metaflow import current
    model_path = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]