PyTorch DirectML

0.2.5.dev240914 verified Fri May 01 auth: no python

Provides a DirectML backend for hardware acceleration in PyTorch, enabling GPU acceleration on Windows devices with DirectX 12 compatible GPUs (including AMD, Intel, and NVIDIA). Current version is 0.2.5.dev240914, with preview releases following a monthly cadence.

pip install torch-directml

Common errors

error ModuleNotFoundError: No module named 'torch_directml' ↓

cause torch-directml is not installed or installed in wrong environment.

fix

Run pip install torch-directml and ensure you are using the correct Python environment.

error RuntimeError: DirectML device not found ↓

cause No compatible DirectX 12 GPU is available or drivers are outdated.

fix

Ensure Windows 10/11, DirectX 12 feature level 12_0, and latest GPU drivers are installed.

error AttributeError: module 'torch_directml' has no attribute 'device_count' ↓

cause Using an older version of torch-directml that does not have device_count.

fix

Upgrade to torch-directml >=0.2.4: pip install --upgrade torch-directml.

Warnings

gotcha torch-directml does not support CUDA APIs; use torch_directml.device() instead of torch.device('cuda'). ↓

fix Use `torch_directml.device()` or `torch_directml.device(torch_directml.device_count() - 1)` for specific devices.

deprecated The `dml` module alias is deprecated; always import `torch_directml`. ↓

fix Use `import torch_directml` instead of `import dml`.

gotcha Only Windows is supported; Linux/WSL is not supported for torch-directml (separate package tensorflow-directml may work for TF). ↓

fix Run on Windows 10/11 with DirectX 12 capable GPU.

Imports

torch_directml
wrong
```
import directml
```
correct
```
import torch_directml
```
The correct import is torch_directml, not directml.

Quickstart

Basic setup: import torch_directml, get the default DirectML device, move a tensor to it.

import torch
import torch_directml
device = torch_directml.device()
tensor = torch.tensor([1.0, 2.0, 3.0]).to(device)
print(tensor.device)