# Say Goodbye to the GPU Waiting Game — SageMaker Now Supports Instance Preference Lists!

Hi everyone, it's Shii-chan!

Today I found an AWS update that machine learning engineers are going to love. If you've ever been stuck waiting for a scarce GPU instance, this one's for you.

## What was announced?

According to AWS's What's New, Amazon SageMaker AI training and processing jobs now support prioritized "instance preference lists." Many AI training, fine-tuning, and data processing workloads run comparably well on several different instance types or sizes, but until now, you had to name exactly one instance type when submitting a job and then wait for SageMaker to find that specific instance for you.

## The story so far

High-demand GPUs can have unpredictable wait times depending on when you submit. That pushed some users toward building complex retry logic on their own, or submitting several jobs at once with different instance types just to grab whichever one became available first.

## What changes

You can now provide a prioritized list of the instance types your workload can accept when you submit a job. SageMaker works through the list in order and automatically launches your job on the first configuration where capacity is available.

For example, your preference list could look like this:

- 1st choice: 2 instances of ml.g6.48xlarge
- 2nd choice: 4 instances of ml.g5.48xlarge

Thanks to this, your training or processing job should be more likely to start sooner. You can also mix on-demand capacity with reserved capacity from SageMaker Flexible Training Plans within a single job submission.

## Dive Deep

You can configure this through the SageMaker console, API, SDK, or CLI. Just specify your instance type and count preferences in priority order when submitting the job, and SageMaker takes care of finding an available configuration.

It's available today in all AWS Regions where SageMaker AI is available. For more details, check the official docs on [instance preference lists](https://docs.aws.amazon.com/sagemaker/latest/dg/train-instance-preferences.html) or the AWS Machine Learning blog post covering the launch.

The announcement doesn't mention specific pricing details.

## Wrap-up

- SageMaker AI training and processing jobs now support prioritized instance preference lists
- Previously, you could only specify a single instance type, which meant waiting through GPU capacity crunches
- SageMaker now automatically runs your job on the first available configuration from your list
- You can combine on-demand capacity with reserved SageMaker Flexible Training Plans capacity
- Available via console, API, SDK, and CLI, in every Region where SageMaker AI is supported, starting today

This is a great fit if you've been fighting capacity waits for in-demand GPU instances.