Notes on Using MPS (Metal Performance Shaders) with DJL's PyTorch Backend

⚠️ This article was automatically translated by OpenAI (gpt-4o). It may be edited eventually, but please be aware that it may contain incorrect information at this time.

DJL (Deep Java Library) has supported MPS since version 0.20.0. I couldn't find any samples, so I decided to try it out and take some notes.

The sample code is here. I tested it on an Apple M2 Pro, with 32 GB of memory, running macOS 13.5.2.

It seems you can create a Device instance with Device.of("mps", 0) as shown below:

import ai.djl.Device;
import ai.djl.ndarray.NDArray;
import ai.djl.ndarray.NDManager;
import ai.djl.ndarray.types.Shape;

public class Main {
	public static void main(String[] args) {
		int dimension = 1024;
		Device device = Device.of("mps", 0);
		//Device device = Device.cpu();
		System.out.println(device.isGpu()); // false
		try (NDManager manager = NDManager.newBaseManager(device)) {
			NDArray array1 = manager.randomUniform(0, 1, new Shape(dimension, dimension));
			NDArray array2 = manager.randomUniform(0, 1, new Shape(dimension, dimension));
			NDArray result = array1.add(array2).mul(10).matMul(array1.transpose()).div(5);
			System.out.println(result);
		}
	}
}

Although MPS itself is not a GPU, using the MPS API allows the GPU to be utilized. Therefore, when you run code that uses MPS, the % GPU number in the Activity Monitor will be greater than 0.