flyte.prefetch

Prefetch utilities for Flyte.

This module provides functionality to prefetch various artifacts from remote registries, such as HuggingFace models.

Class	Description
`HuggingFaceModelInfo`	Information about a HuggingFace model to store.
`ShardConfig`	Configuration for model sharding.
`StoredModelInfo`	Information about a stored model.
`VLLMShardArgs`	Arguments for sharding a model using vLLM.

Method	Description
`hf_model()`	Store a HuggingFace model to remote storage.

Methods

hf_model()

        
    
def hf_model(
    repo: str,
    raw_data_path: str | None = None,
    artifact_name: str | None = None,
    architecture: str | None = None,
    task: str = 'auto',
    modality: tuple[str, ...] = ('text',),
    serial_format: str | None = None,
    model_type: str | None = None,
    short_description: str | None = None,
    shard_config: ShardConfig | None = None,
    hf_token_key: str | None = 'HF_TOKEN',
    resources: Resources = Resources(cpu='2', memory='8Gi', gpu=None, disk='50Gi', shm=None),
    force: int = 0,
) -> Run

Store a HuggingFace model to remote storage.

This function downloads a model from the HuggingFace Hub and prefetches it to remote storage. It supports optional sharding using vLLM for large models.

The prefetch behavior follows this priority:

If the model isn’t being sharded, stream files directly to remote storage.
If streaming fails, fall back to downloading a snapshot and uploading.
If sharding is configured, download locally, shard with vLLM, then upload.

On success the platform records a model artifact for the stored Dir: the artifact name is artifact_name (default: the repo name), the version is the HuggingFace commit id, the searchable metadata carries the model facts (framework/architecture/task/modality/serial_format plus the source repo and commit), and the repo’s README is attached as the model card. Retrieve it later with flyte.remote.Artifact.get(artifact_name).

Example usage:

        
    
import flyte

flyte.init(endpoint="my-flyte-endpoint")

# Store a model without sharding
run = flyte.prefetch.hf_model(
    repo="meta-llama/Llama-2-7b-hf",
    hf_token_key="HF_TOKEN",
)
run.wait()

# Prefetch and shard a model
from flyte.prefetch import ShardConfig, VLLMShardArgs

run = flyte.prefetch.hf_model(
    repo="meta-llama/Llama-2-70b-hf",
    shard_config=ShardConfig(
        engine="vllm",
        args=VLLMShardArgs(tensor_parallel_size=8),
    ),
    accelerator="A100:8",
    hf_token_key="HF_TOKEN",
)
run.wait()

Parameter	Type	Description
`repo`	`str`	The HuggingFace repository ID (e.g., ‘meta-llama/Llama-2-7b-hf’).
`raw_data_path`	`str \| None`
`artifact_name`	`str \| None`	Optional name for the stored artifact. If not provided, the repo name will be used (with ‘.’ replaced by ‘-’).
`architecture`	`str \| None`	Model architecture from HuggingFace config.json.
`task`	`str`	Model task (e.g., ‘generate’, ‘classify’, ’embed’). Default: ‘auto’.
`modality`	`tuple[str, ...]`	Modalities supported by the model. Default: (’text’,).
`serial_format`	`str \| None`	Model serialization format (e.g., ‘safetensors’, ‘onnx’).
`model_type`	`str \| None`	Model type (e.g., ’transformer’, ‘custom’).
`short_description`	`str \| None`	Short description of the model.
`shard_config`	`ShardConfig \| None`	Optional configuration for model sharding with vLLM.
`hf_token_key`	`str \| None`	Name of the secret containing the HuggingFace token. Default: ‘HF_TOKEN’. Pass None to prefetch public models anonymously (no secret required).
`resources`	`Resources`
`force`	`int`	Force re-prefetch. Increment to force a new prefetch. Default: 0.

Returns: A Run object representing the prefetch task execution.

flyte.prefetch

Directory

Classes

Methods

Methods

hf_model()