Environment:Shiyu coder Kronos PyTorch CUDA Environment

Knowledge Sources	Kronos PyTorch
Domains	Infrastructure, Deep_Learning
Last Updated	2026-02-09 13:47 GMT

Overview

Python 3.x environment with PyTorch >= 2.0.0, optional CUDA GPU acceleration, and core scientific computing libraries for financial time series forecasting.

Description

This environment provides the base runtime for all Kronos model operations including inference, finetuning, and backtesting. It is built around PyTorch >= 2.0.0 with automatic device detection supporting NVIDIA CUDA GPUs, Apple Silicon MPS, and CPU fallback. The environment includes numerical computing libraries (NumPy, Pandas), tensor operation utilities (einops), and HuggingFace Hub integration for model downloading. It supports both single-GPU inference and multi-GPU distributed training via the DDP environment extension.

Usage

Use this environment for all Kronos workflows: single-series prediction, batch prediction, Qlib finetuning, CSV finetuning, and web UI inference. This is the mandatory base prerequisite for every Implementation in the repository.

System Requirements

Category	Requirement	Notes
OS	Linux (Ubuntu 20.04+), macOS, or Windows	Linux recommended for GPU training
Hardware	NVIDIA GPU (optional)	CUDA-capable GPU for acceleration; MPS on Apple Silicon; CPU fallback supported
VRAM	4GB+ recommended	Kronos-mini (4.1M params) runs on limited VRAM; Kronos-base (102.3M params) needs more
Python	3.8+	Compatible with standard CPython
Disk	2GB minimum	For model weights and dependencies

Dependencies

System Packages

CUDA Toolkit (optional, for GPU acceleration)
`git` (for cloning repository)

Python Packages

`torch` >= 2.0.0
`numpy`
`pandas` == 2.2.2
`einops` == 0.8.1
`huggingface_hub` == 0.33.1
`matplotlib` == 3.9.3
`tqdm` == 4.67.1
`safetensors` == 0.6.2

Credentials

No credentials required for the base environment. Model weights are publicly available on HuggingFace Hub without authentication:

`NeoQuasar/Kronos-mini` (4.1M parameters)
`NeoQuasar/Kronos-small` (24.7M parameters)
`NeoQuasar/Kronos-base` (102.3M parameters)
`NeoQuasar/Kronos-Tokenizer-2k`
`NeoQuasar/Kronos-Tokenizer-base`

Quick Install

# Install all core dependencies
pip install torch>=2.0.0 numpy pandas==2.2.2 einops==0.8.1 huggingface_hub==0.33.1 matplotlib==3.9.3 tqdm==4.67.1 safetensors==0.6.2

Code Evidence

Device auto-detection from `model/kronos.py:494-501`:

# Auto-detect device if not specified
if device is None:
    if torch.cuda.is_available():
        device = "cuda:0"
    elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
        device = "mps"
    else:
        device = "cpu"

HuggingFace Hub integration from `model/kronos.py:13`:

class KronosTokenizer(nn.Module, PyTorchModelHubMixin):

Einops dependency from `model/module.py:2`:

from einops import rearrange, reduce

Common Errors

Error Message	Cause	Solution
`ModuleNotFoundError: No module named 'einops'`	einops not installed	`pip install einops==0.8.1`
`RuntimeError: CUDA out of memory`	Insufficient GPU VRAM	Use smaller model (Kronos-mini) or reduce batch size
`ModuleNotFoundError: No module named 'model'`	Incorrect Python path	Ensure repository root is in PYTHONPATH or run from repo root

Compatibility Notes

NVIDIA CUDA GPUs: Full support with automatic detection. Recommended for training.
Apple Silicon (MPS): Supported for inference via `torch.backends.mps`. Training not tested.
CPU: Fully supported but significantly slower. Suitable for small-scale inference only.
PyTorch >= 2.0: Required. Uses `F.scaled_dot_product_attention` which leverages FlashAttention when available on compatible hardware.

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment