uv-func 0.1.1

A package to run functions within dedicated UV Python environments.

uv-func

Run Python functions in isolated UV virtual environments with their own dependencies – like uv run for individual functions.

Why uv-func?

Problem: Need different library versions in the same program? Want to use heavy dependencies (PyTorch, TensorFlow) in just one function? Dealing with dependency conflicts?

Solution: Decorate functions to run in isolated environments with specific dependencies, automatic caching, and real-time logging.

Quick Start

Installation

pip install uv-func

Usage

import uv_func

@uv_func.run(dependencies=["torch==2.8.0", "numpy==1.26.4"])
def torch_sum(x: list[int], y: list[int]) -> list[float]:
    import torch
    x_tensor = torch.tensor(x).float()
    y_tensor = torch.tensor(y).float()
    return (x_tensor + y_tensor).tolist()  # Return serializable list

result = torch_sum([1,2,3], [4,5,6])  # [5.0, 7.0, 9.0]

Examples

Different Library Versions

@uv_func.run(dependencies=["numpy==1.21.0"])
def legacy_compute(data: list) -> float:
    import numpy as np
    return float(np.array(data).sum())

@uv_func.run(dependencies=["numpy==1.26.4"])
def modern_compute(data: list) -> float:
    import numpy as np
    return float(np.array(data, dtype=np.float32).mean())

Data Processing Pipeline

@uv_func.run(dependencies=["polars==0.20.0"])
def process_data(file_path: str) -> dict:
    import polars as pl
    df = pl.read_csv(file_path).filter(pl.col("value") > 100)
    return {"data": df.to_dicts(), "shape": df.shape}

@uv_func.run(dependencies=["matplotlib==3.8.0", "seaborn==0.13.0"])
def visualize(data: list[float], filename: str = "plot.png") -> dict:
    import matplotlib.pyplot as plt
    import seaborn as sns
    plt.figure(figsize=(10, 6))
    sns.histplot(data)
    plt.savefig(filename)
    plt.close()
    return {"filename": filename, "points": len(data)}

How It Works

1. Hash dependencies → Check cache for existing environment
2. Create/reuse UV virtual environment with exact dependencies
3. Serialize function + arguments with cloudpickle
4. Execute in subprocess with isolated Python interpreter
5. Return serialized results to parent process

⚠️ Serialization Constraints

uv-func uses cloudpickle for IPC. Keep inputs/outputs simple!

✅ Use: primitives (int, float, str, bool), collections (list, dict, tuple), simple dataclasses
❌ Avoid: complex objects, file handles, database connections, threading locks, raw library objects

# ✅ GOOD: Return dict
@uv_func.run(dependencies=["requests"])
def fetch(url: str) -> dict:
    import requests
    resp = requests.get(url)
    return {"status": resp.status_code, "data": resp.json()}

# ❌ BAD: Return complex object
@uv_func.run(dependencies=["requests"])
def fetch_bad(url: str):  # Don't do this
    import requests
    return requests.get(url)  # Response object won't serialize

API Reference

@uv_func.run(dependencies, verbose=False)

• dependencies (list[str]): Pip-style dependencies (e.g., ["torch==2.0.1", "numpy>=1.20"])
• verbose (bool): Enable detailed environment setup logs
• Returns: Callable executing function in isolation

Note: Arguments and returns must be cloudpickle-serializable.

Features

• 🚀 Fast: Leverages UV's speed for environment management
• 📦 Cached: Reuses environments based on dependency hash
• 🔄 Streaming: Real-time stdout/stderr from isolated functions
• 🛡️ Safe: Full exception propagation with tracebacks
• 🧵 Thread-safe: Works in multi-threaded applications

Advanced Usage

# Verbose mode for debugging
@uv_func.run(dependencies=["tensorflow==2.13.0"], verbose=True)
def train():
    import tensorflow as tf
    return tf.__version__

# Error handling
try:
    result = some_isolated_func()
except ChildProcessError as e:
    print(f"Function failed: {e}")

Requirements

• Python 3.12+
• UV package manager

Contributing

Contributions welcome! Please submit a Pull Request.

License

MIT License - see LICENSE file for details.

Related Projects

• uv - Fast Python package installer and resolver