Awesome

PyTorch/XLA

<b>Current CI status:</b>

PyTorch/XLA is a Python package that uses the XLA deep learning compiler to connect the PyTorch deep learning framework and Cloud TPUs. You can try it right now, for free, on a single Cloud TPU VM with Kaggle!

Take a look at one of our Kaggle notebooks to get started:

• Stable Diffusion with PyTorch/XLA 2.0
• Distributed PyTorch/XLA Basics

Installation

TPU

To install PyTorch/XLA stable build in a new TPU VM:

pip install torch~=2.5.0 torch_xla[tpu]~=2.5.0 -f https://storage.googleapis.com/libtpu-releases/index.html -f https://storage.googleapis.com/libtpu-wheels/index.html

To install PyTorch/XLA nightly build in a new TPU VM:

pip3 install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cpu
pip install 'torch_xla[tpu] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.7.0.dev-cp310-cp310-linux_x86_64.whl' -f https://storage.googleapis.com/libtpu-releases/index.html -f https://storage.googleapis.com/libtpu-wheels/index.html

GPU Plugin

PyTorch/XLA now provides GPU support through a plugin package similar to libtpu:

pip install torch~=2.5.0 torch_xla~=2.5.0 https://storage.googleapis.com/pytorch-xla-releases/wheels/cuda/12.1/torch_xla_cuda_plugin-2.5.0-py3-none-any.whl

Getting Started

To update your existing training loop, make the following changes:

-import torch.multiprocessing as mp
+import torch_xla as xla
+import torch_xla.core.xla_model as xm

 def _mp_fn(index):
   ...

+  # Move the model paramters to your XLA device
+  model.to(xla.device())

   for inputs, labels in train_loader:
+    with xla.step():
+      # Transfer data to the XLA device. This happens asynchronously.
+      inputs, labels = inputs.to(xla.device()), labels.to(xla.device())
       optimizer.zero_grad()
       outputs = model(inputs)
       loss = loss_fn(outputs, labels)
       loss.backward()
-      optimizer.step()
+      # `xm.optimizer_step` combines gradients across replicas
+      xm.optimizer_step(optimizer)

 if __name__ == '__main__':
-  mp.spawn(_mp_fn, args=(), nprocs=world_size)
+  # xla.launch automatically selects the correct world size
+  xla.launch(_mp_fn, args=())

If you're using DistributedDataParallel, make the following changes:

 import torch.distributed as dist
-import torch.multiprocessing as mp
+import torch_xla as xla
+import torch_xla.distributed.xla_backend

 def _mp_fn(rank):
   ...

-  os.environ['MASTER_ADDR'] = 'localhost'
-  os.environ['MASTER_PORT'] = '12355'
-  dist.init_process_group("gloo", rank=rank, world_size=world_size)
+  # Rank and world size are inferred from the XLA device runtime
+  dist.init_process_group("xla", init_method='xla://')
+
+  model.to(xm.xla_device())
+  ddp_model = DDP(model)

-  model = model.to(rank)
-  ddp_model = DDP(model, device_ids=[rank])

   for inputs, labels in train_loader:
+    with xla.step():
+      inputs, labels = inputs.to(xla.device()), labels.to(xla.device())
       optimizer.zero_grad()
       outputs = ddp_model(inputs)
       loss = loss_fn(outputs, labels)
       loss.backward()
       optimizer.step()

 if __name__ == '__main__':
-  mp.spawn(_mp_fn, args=(), nprocs=world_size)
+  xla.launch(_mp_fn, args=())

Additional information on PyTorch/XLA, including a description of its semantics and functions, is available at PyTorch.org. See the API Guide for best practices when writing networks that run on XLA devices (TPU, CUDA, CPU and...).

Our comprehensive user guides are available at:

Documentation for the latest release

Documentation for master branch

PyTorch/XLA tutorials

• Cloud TPU VM quickstart
• Cloud TPU Pod slice quickstart
• Profiling on TPU VM
• GPU guide

Reference implementations

The AI-Hypercomputer/tpu-recipes repo. contains examples for training and serving many LLM and diffusion models.

Available docker images and wheels

Python packages

PyTorch/XLA releases starting with version r2.1 will be available on PyPI. You can now install the main build with pip install torch_xla. To also install the Cloud TPU plugin corresponding to your installed torch_xla, install the optional tpu dependencies after installing the main build with

pip install torch_xla[tpu] \
  -f https://storage.googleapis.com/libtpu-wheels/index.html \
  -f https://storage.googleapis.com/libtpu-releases/index.html

GPU and nightly builds are available in our public GCS bucket.

pip3 install torch==2.6.0.dev20240925+cpu --index-url https://download.pytorch.org/whl/nightly/cpu
pip3 install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-nightly%2B20240925-cp310-cp310-linux_x86_64.whl

The torch wheel version 2.6.0.dev20240925+cpu can be found at https://download.pytorch.org/whl/nightly/torch/.

Use nightly build after 08/20/2024

You can also add yyyymmdd after torch_xla-2.6.0.dev (or the latest dev version) to get the nightly wheel of a specified date. Here is an example:

pip3 install torch==2.5.0.dev20240820+cpu --index-url https://download.pytorch.org/whl/nightly/cpu
pip3 install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.5.0.dev20240820-cp310-cp310-linux_x86_64.whl

The torch wheel version 2.6.0.dev20240925+cpu can be found at https://download.pytorch.org/whl/nightly/torch/.

Use nightly build with C++11 ABI after 10/28/2024

By default, torch is built with pre-C++11 version of ABI (see https://github.com/pytorch/pytorch/issues/51039). torch_xla follows that and ships pre-C++11 builds by default. However, the lazy tensor tracing performance can be improved by building the code with C++11 ABI. As a result, we provide C++11 ABI builds for interested users to try, especially if you find your model performance bottlenecked in Python lazy tensor tracing.

You can add .cxx11 after yyyymmdd to get the C++11 ABI variant of a specific nightly wheel. Here is an example to install nightly builds from 10/28/2024:

pip3 install torch==2.6.0.dev20241028+cpu.cxx11.abi --index-url https://download.pytorch.org/whl/nightly
pip3 install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241028.cxx11-cp310-cp310-linux_x86_64.whl

As of 12/11/2024, the torch_xla C++11 ABI wheel is named differently and can be installed as follows:

pip3 install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0.dev20241211+cxx11-cp310-cp310-linux_x86_64.whl

The torch wheel version 2.6.0.dev20241028+cpu.cxx11.abi can be found at https://download.pytorch.org/whl/nightly/torch/.

Docker

To use the above dockers, please pass --privileged --net host --shm-size=16G along. Here is an example:

docker run --privileged --net host --shm-size=16G -it us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:nightly_3.10_tpuvm /bin/bash

To run on compute instances with GPUs.

Troubleshooting

If PyTorch/XLA isn't performing as expected, see the troubleshooting guide, which has suggestions for debugging and optimizing your network(s).

Providing Feedback

The PyTorch/XLA team is always happy to hear from users and OSS contributors! The best way to reach out is by filing an issue on this Github. Questions, bug reports, feature requests, build issues, etc. are all welcome!

Contributing

See the contribution guide.

Disclaimer

This repository is jointly operated and maintained by Google, Meta and a number of individual contributors listed in the CONTRIBUTORS file. For questions directed at Meta, please send an email to opensource@fb.com. For questions directed at Google, please send an email to pytorch-xla@googlegroups.com. For all other questions, please open up an issue in this repository here.

Additional Reads

You can find additional useful reading materials in

• Performance debugging on Cloud TPU VM
• Lazy tensor intro
• Scaling deep learning workloads with PyTorch / XLA and Cloud TPU VM
• Scaling PyTorch models on Cloud TPUs with FSDP

Related Projects

• OpenXLA
• HuggingFace
• JetStream