Skip to main content

​
TimeGPT on Ray

Ray is an open-source unified compute framework that helps scale Python workloads for distributed computing. In this tutorial, you will learn how to distribute TimeGPT forecasting jobs on top of Ray.
Open In Colab
This guide uses Fugue to easily run code across various distributed computing frameworks, including Ray.

​
Overview

Ray

A framework for scaling Python beyond a single machine.

Fugue

A unified interface to execute Python code on several backends (including Ray).

TimeGPT

A Nixtla service for scalable time-series forecasting.
By combining Ray, Fugue, and TimeGPT, you can quickly scale your time-series forecasting tasks in distributed environments.
Below is an outline of what we’ll cover:
  1. Installation
  2. Load Your Data
  3. Initialize Ray
  4. Use TimeGPT on Ray
  5. Shutdown Ray
1

1. Installation

Install Ray using Fugue. Fugue provides an easy-to-use interface for distributed computation. It lets you run Python code on several distributed computing frameworks, including Ray.
Fugue Ray Installation
pip install fugue[ray]
When executing on a distributed Ray cluster, ensure the nixtla library is installed on all workers.
2

2. Load Your Data

Load your dataset into a pandas DataFrame. This tutorial uses hourly electricity prices from various markets:
Load Dataset Example
import pandas as pd

df = pd.read_csv(
    'https://raw.githubusercontent.com/Nixtla/transfer-learning-time-series/main/datasets/electricity-short.csv',
    parse_dates=['ds'],
)
df.head()
unique_iddsy
0BE2016-10-22 00:00:0070.00
1BE2016-10-22 01:00:0037.10
2BE2016-10-22 02:00:0037.10
3BE2016-10-22 03:00:0044.75
4BE2016-10-22 04:00:0037.10

Preview of the first few rows of data

3

3. Initialize Ray

Here, we’re spinning up a Ray cluster locally by creating a head node. You can scale this to multiple machines in a real cluster environment.
Ray Cluster Initialization
import ray
from ray.cluster_utils import Cluster

ray_cluster = Cluster(
    initialize_head=True,
    head_node_args={"num_cpus": 2}
)

ray.init(address=ray_cluster.address, ignore_reinit_error=True)

# Convert your DataFrame to Ray format:
ray_df = ray.data.from_pandas(df)
ray_df
Ray Initialization Logs
INFO:ray._private.services:Ray runtime started.
INFO:ray.dashboard:Dashboard is available at http://127.0.0.1:8265
INFO:ray.cluster_utils:Cluster address: 127.0.0.1:6379
4

4. Use TimeGPT on Ray

With Ray, you can run TimeGPT similar to a standard (non-distributed) local environment. Operations such as forecast still apply directly to Ray Dataset objects.
Begin by creating a NixtlaClient. Replace my_api_key_provided_by_nixtla with your own API key.
NixtlaClient Initialization
from nixtla import NixtlaClient

nixtla_client = NixtlaClient(
  api_key='my_api_key_provided_by_nixtla'
)
If you prefer using an Azure AI endpoint, specify the base_url and api_key for Azure.
NixtlaClient Azure Setup
nixtla_client = NixtlaClient(
  base_url="your azure ai endpoint",
  api_key="your api_key"
)
TimeGPT Forecasting on Ray
%%capture
fcst_df = nixtla_client.forecast(ray_df, h=12)
Public API models supported include timegpt-1 (default) and timegpt-1-long-horizon.
Inspect the forecast results by converting to a pandas DataFrame:
Inspect Forecast Results
fcst_df.to_pandas().tail()
You can also perform cross-validation on Ray. The following sample code performs a cross-validation procedure using rolling windows:
TimeGPT Cross-validation on Ray
%%capture
cv_df = nixtla_client.cross_validation(
ray_df,
h=12,
freq='H',
n_windows=5,
step_size=2
)
After computation, convert cv_df to pandas to view the results:
Inspect Cross-validation Results
cv_df.to_pandas().tail()
For adding extra predictors or exogenous variables, refer to the Exogenous Variables Tutorial. While running on Ray, simply use your ray_df in place of a pandas DataFrame.
5

5. Shutdown Ray

Always shut down Ray after you finish your tasks to free up resources.
Shutdown Ray Example
ray.shutdown()
Congratulations! You’ve successfully used TimeGPT on Ray for distributed forecasting.
⌘I