How to Use MPS for Mac GPU with PyTorch

To use mps for Mac GPU acceleration in PyTorch, set your device with torch.device('mps') and move your tensors and models to this device using .to(device). This enables PyTorch to run computations on Apple Silicon GPUs via Metal Performance Shaders.

Syntax

Use torch.device('mps') to specify the Mac GPU device. Then move your tensors and models to this device with .to(device). This tells PyTorch to run operations on the Mac GPU instead of CPU.

• device = torch.device('mps'): sets the device to Mac GPU
• tensor.to(device): moves tensor to GPU
• model.to(device): moves model to GPU

python

import torch

device = torch.device('mps')  # Set device to Mac GPU

tensor = torch.randn(3, 3).to(device)  # Move tensor to GPU

# Example model move
# model = MyModel()
# model.to(device)

Example

This example shows how to create a tensor on the Mac GPU using MPS and perform a simple operation. It demonstrates device setup, tensor creation, and moving data to the GPU.

python

import torch

# Check if MPS is available
if torch.backends.mps.is_available():
    device = torch.device('mps')
else:
    device = torch.device('cpu')

# Create tensor on CPU
x_cpu = torch.tensor([1.0, 2.0, 3.0])

# Move tensor to MPS device
x_mps = x_cpu.to(device)

# Perform operation on MPS
y_mps = x_mps * 2

# Move result back to CPU
y_cpu = y_mps.to('cpu')

print('Original tensor:', x_cpu)
print('Tensor on MPS device:', x_mps)
print('Result after multiplication:', y_cpu)

Output

Original tensor: tensor([1., 2., 3.])
Tensor on MPS device: tensor([1., 2., 3.], device='mps:0')
Result after multiplication: tensor([2., 4., 6.])

Common Pitfalls

Common mistakes when using MPS with PyTorch on Mac include:

• Not checking if MPS is available before using it, which causes errors on unsupported Macs.
• Forgetting to move both model and data to the mps device, leading to device mismatch errors.
• Using operations or PyTorch features not yet supported on MPS, which can cause runtime errors.

Always check torch.backends.mps.is_available() before using MPS and move all tensors and models to the same device.

python

import torch

# Wrong way: no availability check
# device = torch.device('mps')  # May error if MPS unsupported

# Right way:
if torch.backends.mps.is_available():
    device = torch.device('mps')
else:
    device = torch.device('cpu')

# Also remember to move model and data:
# model.to(device)
# data.to(device)

Quick Reference

Summary tips for using MPS on Mac with PyTorch:

• Check availability with torch.backends.mps.is_available().
• Set device with torch.device('mps').
• Move all tensors and models to the MPS device.
• Use CPU fallback if MPS is not available.
• Test your code for unsupported operations on MPS.

Key Takeaways

Always check if MPS is available before using it to avoid errors on unsupported Macs.

Set device with torch.device('mps') and move all tensors and models to this device.

Use CPU fallback when MPS is not available to keep code portable.

Not moving model and data to the same device causes runtime errors.

Some PyTorch features may not be fully supported on MPS yet.