Environment:Sktime Pytorch forecasting Core Python Dependencies

Knowledge Sources	pytorch-forecasting PyPI
Domains	Infrastructure, Deep_Learning, Time_Series
Last Updated	2026-02-08 08:00 GMT

Overview

Python 3.10+ environment with PyTorch >= 2.0, Lightning >= 2.0, pandas, scikit-learn, scipy, and numpy as mandatory runtime dependencies.

Description

This environment defines the core dependency stack required for all pytorch-forecasting functionality. It is built on PyTorch for tensor operations, Lightning for training orchestration, pandas for data handling, scikit-learn for preprocessing utilities, and scipy for statistical operations. The package excludes PyTorch 2.0.1 due to known bugs and pins Lightning below 2.7.0 for compatibility. A runtime adaptation detects Lightning >= 2.6 and adjusts checkpoint loading behavior (weights_only parameter) accordingly.

Usage

This environment is required for all pytorch-forecasting workflows. Every Implementation page depends on these core packages being installed. Without this environment, no models, datasets, or metrics can be used.

System Requirements

Category	Requirement	Notes
OS	Linux, macOS, or Windows	macOS requires `brew install libomp` for PyTorch OpenMP support
Hardware	CPU minimum; NVIDIA GPU optional	CUDA GPU accelerates training but is not required; MPS (Apple Silicon) partially supported
Python	>= 3.10, < 3.15	Tested on 3.10, 3.11, 3.12, 3.13, 3.14
Disk	~2GB	For core dependencies installation

Dependencies

System Packages

`libomp` (macOS only, via Homebrew) — required by PyTorch for OpenMP parallelism

Python Packages

`torch` >= 2.0.0, != 2.0.1, < 3.0.0
`lightning` >= 2.0.0, < 2.7.0
`numpy` <= 3.0.0
`scipy` >= 1.8, < 2.0
`pandas` >= 1.3.0, < 3.0.0
`scikit-learn` >= 1.2, < 2.0
`scikit-base` < 0.14.0

Credentials

No credentials are required for the core dependency stack.

Quick Install

# Install pytorch-forecasting with all core dependencies
pip install pytorch-forecasting

# On macOS, also install libomp for PyTorch
brew install libomp

Code Evidence

Core dependency declarations from `pyproject.toml:9,29-37`:

requires-python = ">=3.10,<3.15"

dependencies = [
  "numpy<=3.0.0",
  "torch >=2.0.0,!=2.0.1,<3.0.0",
  "lightning >=2.0.0,<2.7.0",
  "scipy >=1.8,<2.0",
  "pandas >=1.3.0,<3.0.0",
  "scikit-learn >=1.2,<2.0",
  "scikit-base <0.14.0",
]

Lightning >= 2.6 adaptation from `models/base/_base_model.py:820-826`:

from skbase.utils.dependencies import _check_soft_dependencies

if not _check_soft_dependencies("lightning<2.6", severity="none"):
    if "weights_only" not in kwargs:
        kwargs["weights_only"] = False
else:
    kwargs.pop("weights_only", None)

MPS device handling with graceful fallback from `utils/_utils.py:463-471`:

if isinstance(device, str):
    if device == "mps":
        if hasattr(torch.backends, device):
            if torch.backends.mps.is_available() and torch.backends.mps.is_built():
                device = torch.device("mps")
            else:
                device = torch.device("cpu")
    else:
        device = torch.device(device)

Common Errors

Error Message	Cause	Solution
`ModuleNotFoundError: No module named 'lightning'`	Lightning not installed	`pip install lightning>=2.0.0`
`torch 2.0.1 is excluded`	Known-buggy PyTorch release	Install any other torch version: `pip install torch>=2.0.0,!=2.0.1`
`MPS accelerator has a bug`	MultivariateNormalDistributionLoss used on Apple MPS	Use CPU or CUDA instead of MPS for multivariate distribution losses
`weights_only` TypeError on checkpoint load	Lightning < 2.6 does not accept this parameter	The library handles this automatically; ensure pytorch-forecasting is up to date

Compatibility Notes

macOS: Requires `brew install libomp` for PyTorch OpenMP threading. MPS (Apple Silicon) is partially supported but has known bugs with multivariate normal distributions (`pytorch/pytorch#98074`). MPS is globally disabled in the test suite.
Windows: Core functionality works but DDP (distributed data parallel) tests are skipped due to known issues (`#1623`). Hyperparameter tuning tests are also skipped on Windows.
CUDA: Fully supported but not required. Tests run on CPU by default. If CUDA is available, it is auto-detected via `accelerator="auto"` in Lightning Trainer.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment