GitHub - pytorch/xla: Enabling PyTorch on XLA Devices (e.g. Google TPU) (original) (raw)

PyTorch/XLA

Current CI status:

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 withKaggle!

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==2.7.0 'torch_xla[tpu]==2.7.0'

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

pip 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-2.8.0.dev-cp310-cp310-linux_x86_64.whl'

C++11 ABI builds

As of 03/18/2025 and starting from Pytorch/XLA 2.7 release, C++11 ABI builds are the default and we no longer provide wheels built with pre-C++11 ABI.

In Pytorch/XLA 2.6, we'll provide wheels and docker images built with two C++ ABI flavors: C++11 and pre-C++11. Pre-C++11 is the default to align with PyTorch upstream, but C++11 ABI wheels and docker images have better lazy tensor tracing performance.

To install C++11 ABI flavored 2.6 wheels (Python 3.10 example):

pip install torch==2.6.0+cpu.cxx11.abi
https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0%2Bcxx11-cp310-cp310-manylinux_2_28_x86_64.whl
'torch_xla[tpu]'
-f https://storage.googleapis.com/libtpu-releases/index.html
-f https://storage.googleapis.com/libtpu-wheels/index.html
-f https://download.pytorch.org/whl/torch

The above command works for Python 3.10. We additionally have Python 3.9 and 3.11 wheels:

• 3.9: https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0%2Bcxx11-cp39-cp39-manylinux_2_28_x86_64.whl
• 3.10: https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0%2Bcxx11-cp310-cp310-manylinux_2_28_x86_64.whl
• 3.11: https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.6.0%2Bcxx11-cp311-cp311-manylinux_2_28_x86_64.whl

To access C++11 ABI flavored docker image:

us-central1-docker.pkg.dev/tpu-pytorch-releases/docker/xla:r2.6.0_3.10_tpuvm_cxx11

If your model is tracing bound (e.g. you see that the host CPU is busy tracing the model while TPUs are idle), switching to the C++11 ABI wheels/docker images can improve performance. Mixtral 8x7B benchmarking results on v5p-256, global batch size 1024:

• Pre-C++11 ABI MFU: 33%
• C++ ABI MFU: 39%

Github Doc Map

Our github contains many useful docs on working with different aspects of PyTorch XLA, here is a list of useful docs spread around our repository:

• docs/source/learn: docs for learning concepts associated with XLA, troubleshooting, pjrt, eager mode, and dynamic shape.
• docs/source/accelerators: references to GPU and TPU accelerator documents.
• docs/source/perf: documentation about performance specific aspects of PyTorch/XLA such as: AMP, DDP, Dynamo, Fori loop, FSDP, quantization, recompilation, and SPMD
• docs/source/features: documentation on distributed torch, pallas, scan, stable hlo, and triton.
• docs/source/contribute: documents on setting up PyTorch for development, and guides for lowering operations.
• PJRT plugins:
  • CPU
  • CUDA
• torchax/docs: torchax documents
  • torchax/examples: torchax examples

Getting Started

Following here are guides for two modes:

• Single process: one Python interpreter controlling a single GPU/TPU at a time
• Multi process: N Python interpreters are launched, corresponding to N GPU/TPUs found on the system

Another mode is SPMD, where one Python interpreter controls all N GPU/TPUs found on the system. Multi processing is more complex, and is not compatible with SPMD. This tutorial does not dive into SPMD. For more on that, check ourSPMD guide.

Simple single process

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

+import torch_xla

def train(model, training_data, ...): ... for inputs, labels in train_loader:

• with torch_xla.step(): inputs, labels = training_data[i]

•  inputs, labels = inputs.to('xla'), labels.to('xla')
   optimizer.zero_grad()
   outputs = model(inputs)
   loss = loss_fn(outputs, labels)
   loss.backward()
   optimizer.step()

• torch_xla.sync() ...

if name == 'main': ...

• Move the model paramters to your XLA device

• model.to('xla') train(model, training_data, ...) ...

The changes above should get your model to train on the TPU.

Multi processing

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

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

def _mp_fn(index): ...

• Move the model paramters to your XLA device

• model.to(torch_xla.device())

  for inputs, labels in train_loader:

• with torch_xla.step():

•  # Transfer data to the XLA device. This happens asynchronously.

•  inputs, labels = inputs.to(torch_xla.device()), labels.to(torch_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)

• torch_xla.launch automatically selects the correct world size

• torch_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, gradient_as_bucket_view=True)

• 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 theAPI 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-recipesrepo. 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]

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

Use nightly build

You can also add yyyymmdd like torch_xla-2.8.0.devyyyymmdd (or the latest dev version) to get the nightly wheel of a specified date. Here is an example:

pip3 install torch==2.8.0.dev20250423+cpu --index-url https://download.pytorch.org/whl/nightly/cpu
pip3 install https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250423-cp310-cp310-linux_x86_64.whl

The torch wheel version 2.8.0.dev20250423+cpu can be found at https://download.pytorch.org/whl/nightly/torch/.

older versions

Docker

NOTE: Since PyTorch/XLA 2.7, all builds will use the C++11 ABI by default

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 theCONTRIBUTORS file. For questions directed at Meta, please send an email to opensource@fb.com. For questions directed at Google, please send an email topytorch-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