[Jobs] Draft for 'run functions as jobs' API

[Wauplin]

API design

from huggingface_hub import as_job

@as_job(flavor="a10g-small", dependencies=["torch"])
def complex_stuff(*args, **kwargs) -> ...:
    import torch
    ...

result = complex_stuff(...)

• functions and args are serialized using cloudpickle
• job is run using uv
• script is sent as base64 in env variable
• return value is sent in logs as base64

^ clearly not a robust future-proof design but should be good enough to make some first tests.

For bigger input/output values, user must use a repo on the Hub as intermediary step. Or a bucket if they have one 😏

Example

import os
import platform
from huggingface_hub import as_job

@as_job(flavor="cpu-basic")
def cpu_runtime_info(name: str) -> dict:
    """Collect runtime information from a CPU job."""
    return {
        "name": name,  # dummy value received as input
        "platform": platform.platform(),
        "python_version": platform.python_version(),
        "cpu_count": os.cpu_count(),
    }

@as_job(flavor="a10g-small", dependencies=["torch"])
def gpu_runtime_info(name: str) -> dict:
    """Collect runtime information from a GPU job."""
    import torch
    return {
        "name": name,  # dummy value received as input
        "platform": platform.platform(),
        "python_version": platform.python_version(),
        "cpu_count": os.cpu_count(),
        "torch_version": torch.__version__,
    }

if __name__ == "__main__":
    info1 = cpu_runtime_info("my-cpu-job")
    info2 = gpu_runtime_info("my-gpu-job")

    print("CPU job info:", info1)
    print("GPU job info:", info2)

Job 69404b931a39f67af5a41f1c is running remotely. Fetching logs:
https://huggingface.co/jobs/Wauplin/69404b931a39f67af5a41f1c
-----
Installed 1 package in 1ms
Job 69404b9a1a39f67af5a41f1d is running remotely. Fetching logs:
https://huggingface.co/jobs/Wauplin/69404b9a1a39f67af5a41f1d
-----
(...)

CPU job info: {'name': 'my-cpu-job', 'platform': 'Linux-6.12.55-74.119.amzn2023.x86_64-x86_64-with-glibc2.36', 'python_version': '3.11.14', 'cpu_count': 16}
GPU job info: {'name': 'my-gpu-job', 'platform': 'Linux-6.12.58-82.121.amzn2023.x86_64-x86_64-with-glibc2.36', 'python_version': '3.11.14', 'cpu_count': 8, 'torch_version': '2.9.1+cu128'}

TODO

• resolve params spec available only in Python 3.10+

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.