torchruntime is a lightweight package for automatically installing the appropriate variant of PyTorch on a user's computer, based on their OS, and GPU manufacturer and GPU model.
This package is used by Easy Diffusion, but you're welcome to use it as well. It's useful for developers who make PyTorch-based apps that target users with NVIDIA, AMD and Intel graphics cards (as well as CPU-only usage), on Windows, Mac and Linux.
It lets you treat PyTorch as a single dependency (like it should be), and lets you assume that each user will get the most-performant variant of PyTorch suitable for their computer's OS and hardware.
It deals with the complexity of the variety of torch builds and configurations required for CUDA, AMD (ROCm, DirectML), Intel (xpu/DirectML/ipex), and CPU-only.
Compatibility table: Click here to see the supported graphics cards and operating systems.
Supports Windows, Linux, and Mac.
pip install torchruntime
This command should be run on the user's computer, or while creating platform-specific builds:
python -m torchruntime install
This will install torch, torchvision, and torchaudio, and will decide the variant based on the user's OS, GPU manufacturer and GPU model number. See customizing packages for more options.
This should be run inside your program, to initialize the required environment variables (if any) for the variant of torch being used.
import torchruntime
torchruntime.init_torch()By default, python -m torchruntime install will install the latest available torch, torchvision and torchaudio suitable on the user's platform.
You can customize the packages to install by including their names:
- For e.g. to install only
torchandtorchvision, you can runpython -m torchruntime install torch torchvision - To install specific versions (in pip format), you can run
python -m torchruntime install "torch>2.0" "torchvision==0.20"
Supported torch packages: torch, torchvision, torchaudio, torchao.
Note: If you specify package versions, please keep in mind that the version may not be available to all the users on all the torch platforms. For e.g. a user with Python 3.8 would not be able to install torch 2.5 (or higher), because torch 2.5 dropped support for Python 3.8.
So in general, it's better to avoid specifying a version unless it really matters to you (or you know what you're doing). Instead, please allow torchruntime to pick the latest-possible version for the user.
The list of platforms on which torchruntime can install a working variant of PyTorch.
Note: This list is based on user feedback (since I don't have all the cards). Please let me know if your card is supported (or not) by opening a pull request or issue or messaging on Discord (with supporting logs).
CPU-only:
| OS | Supported? | Notes |
|---|---|---|
| Windows | ✅ Yes | x86_64 |
| Linux | ✅ Yes | x86_64 and aarch64 |
| Mac (M1/M2/M3/M4) | ✅ Yes | arm64. mps backend |
| Mac (Intel) | ✅ Yes | x86_64. Stopped after torch 2.2.2 |
NVIDIA:
| Series | Supported? | OS | Notes |
|---|---|---|---|
| 40xx | ✅ Yes | Win/Linux | Uses CUDA 124 |
| 30xx | ✅ Yes | Win/Linux | Uses CUDA 124 |
| 20xx | ✅ Yes | Win/Linux | Uses CUDA 124 |
| 10xx/16xx | ✅ Yes | Win/Linux | Uses CUDA 124. Full-precision required on 16xx series |
AMD:
| Series | Supported? | OS | Notes |
|---|---|---|---|
| 7xxx | ✅ Yes | Win/Linux | Navi3/RDNA3 (gfx110x). ROCm 6.2 on Linux. DirectML on Windows |
| 6xxx | ✅ Yes | Win/Linux | Navi2/RDNA2 (gfx103x). ROCm 6.2 on Linux. DirectML on Windows |
| 6xxx on Intel Mac | ✅ Yes | Intel Mac | gfx103x. 'mps' backend |
| 5xxx | ✅ Yes | Win/Linux | Navi1/RDNA1 (gfx101x). Full-precision required. DirectML on Windows. Linux only supports upto ROCm 5.2. Waiting for this for ROCm 6.2 support. |
| 5xxx on Intel Mac | ❓ Untested (WIP) | Intel Mac | gfx101x. Implemented but need testers, please message on Discord |
| 4xxxG/Radeon VII | ✅ Yes | Win/Linux | Vega 20 gfx906. Need testers for Windows, please message on Discord |
| 2xxxG/Radeon RX Vega 56 | ✅ Yes | Win/Linux | Vega 10 gfx900. ROCm 5.2 on Linux. DirectML on Windows |
| 4xx/5xx/Polaris | Win | gfx80x. Works with DirectML on Windows (notes, 4GB bug). Did not work with ROCm5.7 with custom-compiled PyTorch 1.13 on Linux (notes). |
Apple:
| Series | Supported? | Notes |
|---|---|---|
| M1/M2/M3/M4 | ✅ Yes | 'mps' backend |
| AMD 6xxx on Intel Mac | ✅ Yes | Intel Mac |
| AMD 5xxx on Intel Mac | ❓ Untested (WIP) | Intel Mac |
Intel:
| Series | Supported? | OS | Notes |
|---|---|---|---|
| Arc | ❓ Untested (WIP) | Win/Linux | Implemented but need testers, please message on Discord. Backends: 'xpu' or DirectML or ipex |
See API for a complete list of module functions.
pip install torch installs the CPU-only version of torch, so it won't utilize your GPU's capabilities.
Different models of AMD cards require different LLVM targets, and sometimes different ROCm versions. And ROCm currently doesn't work on Windows, so AMD on Windows is best served (currently) with DirectML.
And plenty of AMD cards work with ROCm (even when they aren't in the official list of supported cards). Information about these cards (for e.g. the LLVM target to use) is pretty scattered.
torchruntime deals with this complexity for your convenience.
📢 I'm looking for contributions in these specific areas:
- More testing on consumer AMD GPUs.
- More support for older AMD GPUs. Explore: Compile and host PyTorch wheels and rocm (on GitHub) for older AMD gpus (e.g. 580/590/Polaris) with the required patches.
- Intel GPUs.
- Testing on professional AMD GPUs (e.g. the Instinct series).
- An easy-to-run benchmark script (that people can run to check the level of compatibility on their platform).
- Improve the logic for supporting multiple AMD GPUs with different ROCm compatibility. At present, it just picks the latest GPU, which means it doesn't support running workloads on multiple AMD GPUs in parallel.
Please message on the Discord community if you have AMD or Intel GPUs, and would like to help with testing or adding support for them! Thanks!
- Code contributors on Easy Diffusion.
- Users on Easy Diffusion's Discord who've helped with testing on various GPUs.
- PCI Database automatically generated from the PCI ID Database at http://pci-ids.ucw.cz