torch.compiler — PyTorch 2.7 documentation (original) (raw)
torch.compiler is a namespace through which some of the internal compiler methods are surfaced for user consumption. The main function and the feature in this namespace is torch.compile.
torch.compile is a PyTorch function introduced in PyTorch 2.x that aims to solve the problem of accurate graph capturing in PyTorch and ultimately enable software engineers to run their PyTorch programs faster. torch.compile is written in Python and it marks the transition of PyTorch from C++ to Python.
torch.compile leverages the following underlying technologies:
- TorchDynamo (torch._dynamo) is an internal API that uses a CPython feature called the Frame Evaluation API to safely capture PyTorch graphs. Methods that are available externally for PyTorch users are surfaced through the
torch.compilernamespace. - TorchInductor is the default
torch.compiledeep learning compiler that generates fast code for multiple accelerators and backends. You need to use a backend compiler to make speedups throughtorch.compilepossible. For NVIDIA, AMD and Intel GPUs, it leverages OpenAI Triton as the key building block. - AOT Autograd captures not only the user-level code, but also backpropagation, which results in capturing the backwards pass “ahead-of-time”. This enables acceleration of both forwards and backwards pass using TorchInductor.
Note
In some cases, the terms torch.compile, TorchDynamo, torch.compilermight be used interchangeably in this documentation.
As mentioned above, to run your workflows faster, torch.compile through TorchDynamo requires a backend that converts the captured graphs into a fast machine code. Different backends can result in various optimization gains. The default backend is called TorchInductor, also known as inductor, TorchDynamo has a list of supported backends developed by our partners, which can be see by running torch.compiler.list_backends() each of which with its optional dependencies.
Some of the most commonly used backends include:
Training & inference backends
| Backend | Description |
|---|---|
| torch.compile(m, backend="inductor") | Uses the TorchInductor backend. Read more |
| torch.compile(m, backend="cudagraphs") | CUDA graphs with AOT Autograd. Read more |
| torch.compile(m, backend="ipex") | Uses IPEX on CPU. Read more |
| torch.compile(m, backend="onnxrt") | Uses ONNX Runtime for training on CPU/GPU. Read more |
Inference-only backends
| Backend | Description |
|---|---|
| torch.compile(m, backend="tensorrt") | Uses Torch-TensorRT for inference optimizations. Requires import torch_tensorrt in the calling script to register backend. Read more |
| torch.compile(m, backend="ipex") | Uses IPEX for inference on CPU. Read more |
| torch.compile(m, backend="tvm") | Uses Apache TVM for inference optimizations. Read more |
| torch.compile(m, backend="openvino") | Uses OpenVINO for inference optimizations. Read more |