PyTorch on XLA Devices¶

PyTorch runs on XLA devices, like TPUs, with the torch_xla package. This document describes how to run your models on these devices.

Creating an XLA Tensor¶

PyTorch/XLA adds a new xla device type to PyTorch. This device type works just like other PyTorch device types. For example, here’s how to create and print an XLA tensor:

import torch
import torch_xla
import torch_xla.core.xla_model as xm

t = torch.randn(2, 2, device=xm.xla_device())
print(t.device)
print(t)

This code should look familiar. PyTorch/XLA uses the same interface as regular PyTorch with a few additions. Importing torch_xla initializes PyTorch/XLA, and xm.xla_device() returns the current XLA device. This may be a CPU or TPU depending on your environment.

XLA Tensors are PyTorch Tensors¶

PyTorch operations can be performed on XLA tensors just like CPU or CUDA tensors.

For example, XLA tensors can be added together:

t0 = torch.randn(2, 2, device=xm.xla_device())
t1 = torch.randn(2, 2, device=xm.xla_device())
print(t0 + t1)

Or matrix multiplied:

print(t0.mm(t1))

Or used with neural network modules:

l_in = torch.randn(10, device=xm.xla_device())
linear = torch.nn.Linear(10, 20).to(xm.xla_device())
l_out = linear(l_in)
print(l_out)

Like other device types, XLA tensors only work with other XLA tensors on the same device. So code like

l_in = torch.randn(10, device=xm.xla_device())
linear = torch.nn.Linear(10, 20)
l_out = linear(l_in)
print(l_out)
# Input tensor is not an XLA tensor: torch.FloatTensor

will throw an error since the torch.nn.Linear module is on the CPU.

Running Models on XLA Devices¶

Building a new PyTorch network or converting an existing one to run on XLA devices requires only a few lines of XLA-specific code. The following snippets highlight these lines when running on a single device and multiple devices with XLA multi-processing.

Running on a Single XLA Device¶

The following snippet shows a network training on a single XLA device:

import torch_xla.core.xla_model as xm

device = xm.xla_device()
model = MNIST().train().to(device)
loss_fn = nn.NLLLoss()
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)

for data, target in train_loader:
  optimizer.zero_grad()
  data = data.to(device)
  target = target.to(device)
  output = model(data)
  loss = loss_fn(output, target)
  loss.backward()

  optimizer.step()
  xm.mark_step()

This snippet highlights how easy it is to switch your model to run on XLA. The model definition, dataloader, optimizer and training loop can work on any device. The only XLA-specific code is a couple lines that acquire the XLA device and mark the step. Calling xm.mark_step() at the end of each training iteration causes XLA to execute its current graph and update the model’s parameters. See XLA Tensor Deep Dive for more on how XLA creates graphs and runs operations.

Running on Multiple XLA Devices with Multi-processing¶

PyTorch/XLA makes it easy to accelerate training by running on multiple XLA devices. The following snippet shows how:

import torch_xla.core.xla_model as xm
import torch_xla.distributed.parallel_loader as pl
import torch_xla.distributed.xla_multiprocessing as xmp

def _mp_fn(index):
  device = xm.xla_device()
  mp_device_loader = pl.MpDeviceLoader(train_loader, device)

  model = MNIST().train().to(device)
  loss_fn = nn.NLLLoss()
  optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)

  for data, target in mp_device_loader:
    optimizer.zero_grad()
    output = model(data)
    loss = loss_fn(output, target)
    loss.backward()
    xm.optimizer_step(optimizer)

if __name__ == '__main__':
  xmp.spawn(_mp_fn, args=())

There are three differences between this multi-device snippet and the previous single device snippet. Let’s go over then one by one.

xmp.spawn()
- Creates the processes that each run an XLA device.
- Each process will only be able to access the device assigned to the current process. For example on a TPU v4-8, there will be 4 processes being spawn up and each process will own a TPU device.
- Note that if you print the xm.xla_device() on each process you will see xla:0 on all devices. This is because each process can only see one device. This does not mean multi-process is not functioning. The only execution is with PJRT runtime on TPU v2 and TPU v3 since there will be #devices/2 processes and each process will have 2 threads(check this doc for more details).
MpDeviceLoader
- Loads the training data onto each device.
- MpDeviceLoader can wrap on a torch dataloader. It can preload the data to the device and overlap the dataloading with device execution to improve the performance.
- MpDeviceLoader also call xm.mark_step for you every batches_per_execution(default to 1) batch being yield.
xm.optimizer_step(optimizer)
- Consolidates the gradients between devices and issues the XLA device step computation.
- It is pretty much a all_reduce_gradients + optimizer.step() + mark_step and returns the loss being reduced.

The model definition, optimizer definition and training loop remain the same.

NOTE: It is important to note that, when using multi-processing, the user can start retrieving and accessing XLA devices only from within the target function of xmp.spawn() (or any function which has xmp.spawn() as parent in the call stack).

See the full multiprocessing example for more on training a network on multiple XLA devices with multi-processing.

Running on TPU Pods¶

Multi-host setup for different accelerators can be very different. This doc will talk about the device independent bits of multi-host training and will use the TPU + PJRT runtime(currently available on 1.13 and 2.x releases) as an example.

Before you being, please take a look at our user guide at here which will explain some Google Cloud basis like how to use gcloud command and how to setup your project. You can also check here for all Cloud TPU Howto. This doc will focus on the PyTorch/XLA perspective of the Setup.

Let’s assume you have the above mnist example from above section in a train_mnist_xla.py. If it is a single host multi device training, you would ssh to the TPUVM and run command like

PJRT_DEVICE=TPU python3 train_mnist_xla.py

Now in order to run the same models on a TPU v4-16 (which has 2 host, each with 4 TPU devices), you will need to

Make sure each host can access the training script and training data. This is usually done by using the gcloud scp command or gcloud ssh command to copy the training scripts to all hosts.
Run the same training command on all hosts at the same time.

gcloud alpha compute tpus tpu-vm ssh $USER-pjrt --zone=$ZONE --project=$PROJECT --worker=all --command="PJRT_DEVICE=TPU python3 train_mnist_xla.py"

Above gcloud ssh command will ssh to all hosts in TPUVM Pod and run the same command at the same time..

NOTE: You need to run run above gcloud command outside of the TPUVM vm.

The model code and training scirpt is the same for the multi-process training and the multi-host training. PyTorch/XLA and the underlying infrastructure will make sure each device is aware of the global topology and each device’s local and global ordinal. Cross-device communication will happen across all devices instead of local devices.

For more details regarding PJRT runtime and how to run it on pod, please refer to this doc. For more information about PyTorch/XLA and TPU pod and a complete guide to run a resnet50 with fakedata on TPU pod, please refer to this guide.

XLA Tensor Deep Dive¶

Using XLA tensors and devices requires changing only a few lines of code. But even though XLA tensors act a lot like CPU and CUDA tensors, their internals are different. This section describes what makes XLA tensors unique.

XLA Tensors are Lazy¶

CPU and CUDA tensors launch operations immediately or eagerly. XLA tensors, on the other hand, are lazy. They record operations in a graph until the results are needed. Deferring execution like this lets XLA optimize it. A graph of multiple separate operations might be fused into a single optimized operation, for example.

Lazy execution is generally invisible to the caller. PyTorch/XLA automatically constructs the graphs, sends them to XLA devices, and synchronizes when copying data between an XLA device and the CPU. Inserting a barrier when taking an optimizer step explicitly synchronizes the CPU and the XLA device. For more information about our lazy tensor design, you can read this paper.

XLA Tensors and bFloat16¶

PyTorch/XLA can use the bfloat16 datatype when running on TPUs. In fact, PyTorch/XLA handles float types (torch.float and torch.double) differently on TPUs. This behavior is controlled by the XLA_USE_BF16 and XLA_DOWNCAST_BF16 environment variable:

By default both torch.float and torch.double are torch.float on TPUs.
If XLA_USE_BF16 is set, then torch.float and torch.double are both bfloat16 on TPUs.
If XLA_DOWNCAST_BF16 is set, then torch.float is bfloat16 on TPUs and torch.double is float32 on TPUs.
If a PyTorch tensor has torch.bfloat16 data type, this will be directly mapped to the TPU bfloat16 (XLA BF16 primitive type).

Developers should note that XLA tensors on TPUs will always report their PyTorch datatype regardless of the actual datatype they’re using. This conversion is automatic and opaque. If an XLA tensor on a TPU is moved back to the CPU it will be converted from its actual datatype to its PyTorch datatype. Depending on how your code operates, this conversion triggered by the type of processing unit can be important.

Memory Layout¶

The internal data representation of XLA tensors is opaque to the user. They do not expose their storage and they always appear to be contiguous, unlike CPU and CUDA tensors. This allows XLA to adjust a tensor’s memory layout for better performance.

Moving XLA Tensors to and from the CPU¶

XLA tensors can be moved from the CPU to an XLA device and from an XLA device to the CPU. If a view is moved then the data its viewing is also copied to the other device and the view relationship is not preserved. Put another way, once data is copied to another device it has no relationship with its previous device or any tensors on it. Again, depending on how your code operates, appreciating and accommodating this transition can be important.

Saving and Loading XLA Tensors¶

XLA tensors should be moved to the CPU before saving, as in the following snippet:

import torch
import torch_xla
import torch_xla.core.xla_model as xm

device = xm.xla_device()

t0 = torch.randn(2, 2, device=device)
t1 = torch.randn(2, 2, device=device)

tensors = (t0.cpu(), t1.cpu())

torch.save(tensors, 'tensors.pt')

tensors = torch.load('tensors.pt')

t0 = tensors[0].to(device)
t1 = tensors[1].to(device)

This lets you put the loaded tensors on any available device, not just the one on which they were initialized.

Per the above note on moving XLA tensors to the CPU, care must be taken when working with views. Instead of saving views it is recommended that you recreate them after the tensors have been loaded and moved to their destination device(s).

A utility API is provided to save data by taking care of previously moving it to CPU:

import torch
import torch_xla
import torch_xla.core.xla_model as xm

xm.save(model.state_dict(), path)

In case of multiple devices, the above API will only save the data for the master device ordinal (0).

In case where memory is limited compared to the size of the model parameters, an API is provided that reduces the memory footprint on the host:

import torch_xla.utils.serialization as xser

xser.save(model.state_dict(), path)

This API streams XLA tensors to CPU one at a time, reducing the amount of host memory used, but it requires a matching load API to restore:

import torch_xla.utils.serialization as xser

state_dict = xser.load(path)
model.load_state_dict(state_dict)

Directly saving XLA tensors is possible but not recommended. XLA tensors are always loaded back to the device they were saved from, and if that device is unavailable the load will fail. PyTorch/XLA, like all of PyTorch, is under active development and this behavior may change in the future.

PyTorch/XLA API¶

xla_model¶

torch_xla.core.xla_model.xla_device(n=None, devkind=None)[source]¶

Returns a given instance of an XLA device.

Parameters

n (python:int, optional) – The specific instance (ordinal) to be returned. If specified, the specific XLA device instance will be returned. Otherwise the first device of devkind will be returned.
devkind (string..., optional) – If specified, one of TPU, GPU or CPU.

Returns

A torch.device with the requested instance.

torch_xla.core.xla_model.get_xla_supported_devices(devkind=None, max_devices=None)[source]¶

Returns a list of supported devices of a given kind.

Parameters

devkind (string..., optional) – If specified, one of TPU, GPU or CPU (the ‘GPU’ XLA device is currently not implemented).
max_devices (python:int, optional) – The maximum number of devices to be returned of that kind.

Returns

The list of device strings.

torch_xla.core.xla_model.xla_device_hw(device)[source]¶

Returns the hardware type of the given device.

Parameters: device (string or torch.device) – The xla device that will be mapped to the real device.
Returns: A string representation of the hardware type (CPU, TPU, GPU) of the given device.

torch_xla.core.xla_model.get_ordinal(defval=0)[source]¶

Retrieves the replication ordinal of the current thread.

The ordinals range from 0 to xrt_world_size() minus 1.

Parameters: defval (python:int, optional) – The default value to be returned in case there is no replication information available. Ignored for PjRt. Default: 0
Returns: The replication ordinal of the current thread.

torch_xla.core.xla_model.get_local_ordinal(defval=0)[source]¶

Retrieves the replication local ordinal of the current thread.

The local ordinals range from 0 to the number of local devices minus 1.

Parameters: defval (python:int, optional) – The default value to be returned in case there is no replication information available. Ignored for PjRt. Default: 0
Returns: The replication local ordinal of the current thread.

torch_xla.core.xla_model.is_master_ordinal(local=True)[source]¶

Checks whether the current process is the master ordinal (0).

Parameters: local (bool) – Whether the local or global master ordinal should be checked. In case of multi-host replication, there is only one global master ordinal (host 0, device 0), while there are NUM_HOSTS local master ordinals. Default: True
Returns: A boolean indicating whether the current process is the master ordinal.

torch_xla.core.xla_model.xrt_world_size(defval=1)[source]¶

Retrieves the number of devices which is taking part of the replication.

Parameters: defval (python:int, optional) – The default value to be returned in case there is no replication information available. Default: 1
Returns: The number of devices which is taking part of the replication.

torch_xla.core.xla_model.all_reduce(reduce_type, inputs, scale=1.0, groups=None, cctx=None, pin_layout=True)[source]¶

Performs an inplace reduce operation on the input tensor(s).

Parameters

reduce_type (string) – One of xm.REDUCE_SUM, xm.REDUCE_MUL, xm.REDUCE_AND, xm.REDUCE_OR, xm.REDUCE_MIN and xm.REDUCE_MAX.
inputs – Either a single torch.Tensor or a list of torch.Tensor to perform the all reduce op to.
scale (python:float) – A default scaling value to be applied after the reduce. Default: 1.0
groups (list, optional) –
A list of list, representing the replica groups for the all_reduce() operation. Example: [[0, 1, 2, 3], [4, 5, 6, 7]]

defines two groups, one with the [0, 1, 2, 3] replicas and one with the [4, 5, 6, 7] replicas. If None there will be only one group with all the replicas in it.
pin_layout (bool, optional) – whether to pin the layout for this communication op. Layout pining can prevent potential data corruption when each process that participate in the communication has slightly different program, but it might cause some xla compiation to fail. Unpin the layout when you see error message like “HloModule has a mix of layout constrained”.

Returns

If a single torch.Tensor is passed, the return value is a torch.Tensor holding the reduced value (across the replicas). If a list/tuple is passed, this function performs an inplace all-reduce op on the input tensors, and returns the list/tuple itself.

torch_xla.core.xla_model.all_gather(value, dim=0, groups=None, output=None, pin_layout=True)[source]¶

Performs an all-gather operation along a given dimension.

Parameters

value (torch.Tensor) – The input tensor.
dim (python:int) – The gather dimension. Default: 0
groups (list, optional) –
A list of list, representing the replica groups for the all_gather() operation. Example: [[0, 1, 2, 3], [4, 5, 6, 7]]

defines two groups, one with the [0, 1, 2, 3] replicas and one with the [4, 5, 6, 7] replicas. If None there will be only one group with all the replicas in it.
output (torch.Tensor) – Optional output tensor.
pin_layout (bool, optional) – whether to pin the layout for this communication op. Layout pining can prevent potential data corruption when each process that participate in the communication has slightly different program, but it might cause some xla compiation to fail. Unpin the layout when you see error message like “HloModule has a mix of layout constrained”.

Returns

A tensor which has, in the dim dimension, all the values from the participating replicas.

torch_xla.core.xla_model.all_to_all(value, split_dimension, concat_dimension, split_count, groups=None, pin_layout=True)[source]¶

Performs an XLA AllToAll() operation on the input tensor.

See: https://www.tensorflow.org/xla/operation_semantics#alltoall

Parameters

value (torch.Tensor) – The input tensor.
split_dimension (python:int) – The dimension upon which the split should happen.
concat_dimension (python:int) – The dimension upon which the concat should happen.
split_count (python:int) – The split count.
groups (list, optional) –
A list of list, representing the replica groups for the all_reduce() operation. Example: [[0, 1, 2, 3], [4, 5, 6, 7]]

defines two groups, one with the [0, 1, 2, 3] replicas and one with the [4, 5, 6, 7] replicas. If None there will be only one group with all the replicas in it.
pin_layout (bool, optional) – whether to pin the layout for this communication op. Layout pining can prevent potential data corruption when each process that participate in the communication has slightly different program, but it might cause some xla compiation to fail. Unpin the layout when you see error message like “HloModule has a mix of layout constrained”.

Returns

The result torch.Tensor of the all_to_all() operation.

torch_xla.core.xla_model.add_step_closure(closure, args=(), run_async=False)[source]¶

Adds a closure to the list of the ones to be run at the end of the step.

Many times during model training there is the need to print/report (print to console, post to tensorboard, etc…) information which require the content of intermediary tensors to be inspected. Inspecting different tensors content in different points of the model code requires many executions and typically causes performance issues. Adding a step closure will ensure that it will be run after the barrier, when all the live tensors will be already materialized to device data. Live tensors which will include the ones captured by the closure arguments. So using add_step_closure() will ensure a single execution will be performed, even when multiple closures are queued, requiring multiple tensors to be inspected. Step closures will be run sequentially in the order they have been queued. Note that even though using this API the execution will be optimized, it is advised to throttle the printing/reporting events once every N steps.

Parameters

closure (callable) – The function to be called.
args (tuple) – The arguments to be passed to the closure.
run_async – If True, run the closure asynchronously.

torch_xla.core.xla_model.wait_device_ops(devices=[])[source]¶

Waits for all the async operations on the given devices to complete.

Parameters: devices (string..., optional) – The devices whose async ops need to be waited for. If empty, all the local devices will be waited for.

torch_xla.core.xla_model.optimizer_step(optimizer, barrier=False, optimizer_args={}, groups=None, pin_layout=True)[source]¶

Run the provided optimizer step and issue the XLA device step computation.

Parameters

optimizer (torch.Optimizer) – The torch.Optimizer instance whose step() function needs to be called. The step() function will be called with the optimizer_args named arguments.
barrier (bool, optional) – Whether the XLA tensor barrier should be issued in this API. If using the PyTorch XLA ParallelLoader or DataParallel support, this is not necessary as the barrier will be issued by the XLA data loader iterator next() call. Default: False
optimizer_args (dict, optional) – Named arguments dictionary for the optimizer.step() call.
groups (list, optional) –
A list of list, representing the replica groups for the all_reduce() operation. Example: [[0, 1, 2, 3], [4, 5, 6, 7]]

defines two groups, one with the [0, 1, 2, 3] replicas and one with the [4, 5, 6, 7] replicas. If None there will be only one group with all the replicas in it.
pin_layout (bool, optional) – whether to pin the layout when reducing gradients. See xm.all_reduce for details.

Returns

The same value returned by the optimizer.step() call.

torch_xla.core.xla_model.save(data, file_or_path, master_only=True, global_master=False)[source]¶

Saves the input data into a file.

The saved data is transferred to PyTorch CPU device before being saved, so a following torch.load() will load CPU data. Care must be taken when working with views. Instead of saving views it’s recommended that you recreate them after the tensors have been loaded and moved to their destination device(s).

Parameters

data – The input data to be saved. Any nested combination of Python objects (list, tuples, sets, dicts, …).
file_or_path – The destination for the data saving operation. Either a file path or a Python file object. If master_only is False the path or file objects must point to different destinations as otherwise all the writes from the same host will override each other.
master_only (bool, optional) – Whether only the master device should save the data. If False, the file_or_path argument should be a different file or path for each of the ordinals taking part to the replication, otherwise all the replicas on the same host will be writing to the same location. Default: True
global_master (bool, optional) – When master_only is True this flag controls whether every host’s master (if global_master is False) saves the content, or only the global master (ordinal 0). Default: False

torch_xla.core.xla_model.rendezvous(tag, payload=b'', replicas=[])[source]¶

Waits for all the mesh clients to reach the named rendezvous.

Parameters

tag (string) – The name of the rendezvous to join.
payload (bytes, optional) – The payload to be sent to the rendezvous.
replicas (list, python:int) – The replica ordinals taking part of the rendezvous. Empty means all replicas in the mesh. Default: []

Returns

The payloads exchanged by all the other cores, with the payload of core ordinal i at position i in the returned tuple.

torch_xla.core.xla_model.do_on_ordinals(target, data=(), ordinals=(0, ))[source]¶

Runs a function only on a given set of ordinals.

Parameters

target (callable) – The function to be run on ordinals.
data – Any input data for the target function which contains tensors. All the XLA tensors used by the target function must be passed in this argument. Every other data used by the function can be captured by the Python interpreter as usual. Default: ()
ordinals (list, python:int) – The list/set of ordinals where the target function should run. Default: (0,)

Returns

In the ordinals that ran the target function, the function return value, otherwise None.

torch_xla.core.xla_model.mesh_reduce(tag, data, reduce_fn)[source]¶

Performs an out-of-graph client mesh reduction.

Parameters

tag (string) – The name of the rendezvous to join.
data – The data to be reduced. The reduce_fn callable will receive a list with the copies of the same data coming from all the mesh client processes (one per core).
reduce_fn (callable) – A function which receives a list of data-like objects and returns the reduced result.

Returns

The reduced value.

torch_xla.core.xla_model.set_rng_state(seed, device=None)[source]¶

Sets the random number generator state.

Parameters

seed (python:integer) – The state to be set.
device (string, optional) – The device where the RNG state needs to be set. If missing the default device seed will be set.

torch_xla.core.xla_model.get_rng_state(device=None)[source]¶

Gets the current running random number generator state.

Parameters: device (string, optional) – The device whose RNG state needs to be retrieved. If missing the default device seed will be set.
Returns: The RNG state, as integer.

torch_xla.core.xla_model.get_memory_info(device)[source]¶

Retrieves the device memory information.

Parameters: device (string) – The device whose memory information are requested.
Returns: A dictionary with kb_free (free memory in KB) and kb_total (total memory in KB) keys.

torch_xla.core.functions.all_reduce(reduce_type, value, scale=1.0, groups=None)[source]¶

Performs an inplace reduce operation on the input tensor.

This is the same as xm.all_reduce() but supports autograd differentiation.

Parameters

reduce_type (string) – One of REDUCE_SUM, REDUCE_MUL, REDUCE_AND, REDUCE_OR, REDUCE_MIN and REDUCE_MAX.
value (torch.Tensor) – The to perform the all reduce op to.
scale (python:float) – A default scaling value to be applied after the reduce. Default: 1.0
groups (list, optional) –
A list of list, representing the replica groups for the all_reduce() operation. Example: [[0, 1, 2, 3], [4, 5, 6, 7]]

defines two groups, one with the [0, 1, 2, 3] replicas and one with the [4, 5, 6, 7] replicas. If None there will be only one group with all the replicas in it.

Returns

The reduced value across the selected replicas.

torch_xla.core.functions.all_gather(value, dim=0)[source]¶

Performs an all-gather operation along a given dimension.

This is the same as xm.all_gather() but supports autograd differentiation.

Parameters

value (torch.Tensor) – The input tensor.
dim (python:int) – The gather dimension. Default: 0

Returns

A tensor which has, in the dim dimension, all the values from the participating replicas.

torch_xla.core.functions.nms(boxes, scores, score_threshold, iou_threshold, output_size)[source]¶

Performs a Non Maximal Suppression operation.

Parameters

boxes (torch.Tensor) – A torch.Tensor of shape [N, 4] listing the boxes coordinates in (y0, x0, y1, x1) form.
scores (torch.Tensor) – A torch.Tensor of shape [N] listing the scores of each box.
score_threshold (torch.Tensor) – The minimum score for a box to qualify as valid.
iou_threshold (torch.Tensor) – The minimum IOU (Intersection Over Union) score to trigger overlap logic.
output_size (python:int) – The maximum number of returned indices (must be lower or equal to N).

Returns

A tuple of torch.Tensor with the first element being the selected box indices, and the second element being the number of valid boxes.

distributed¶

class torch_xla.distributed.parallel_loader.ParallelLoader(loader, devices, batchdim=0, batches_per_execution=1, loader_prefetch_size=8, device_prefetch_size=4, host_to_device_transfer_threads=1)[source]¶

Wraps an existing PyTorch DataLoader with background data upload.

Parameters

loader (torch.utils.data.DataLoader) – The PyTorch DataLoader to be wrapped.
devices (torch.device…) – The list of devices where the data has to be sent. The i-th sample returned by the loader will be sent to devices[i % len(devices)].
batchdim (python:int, optional) – The dimension which is holding the batch size. Default: 0
loader_prefetch_size (python:int, optional) – The max capacity of the queue used by the thread which is reading samples from the loader, to be processed by the worker threads which upload data to the devices. Default: 8
device_prefetch_size (python:int, optional) – The max size of the per-device queues, where the worker threads deposit tensors which have already been sent to devices. Default: 4
host_to_device_transfer_threads (python:int, optional) – The number of threads that work in parallel to transfer data from loader queue to device queue. Default: 1

per_device_loader(device)[source]¶

Retrieves the loader iterator object for the given device.

Parameters: device (torch.device) – The device whole loader is being requested.
Returns: The loader iterator object for the device. This is not a torch.utils.data.DataLoader interface, but a Python iterator which returns the same tensor data structure as returned by the wrapped torch.utils.data.DataLoader, but residing on XLA devices.

torch_xla.distributed.xla_multiprocessing.spawn(fn, args=(), nprocs=None, join=True, daemon=False, start_method='spawn')[source]¶

Enables multi processing based replication.

Parameters

fn (callable) – The function to be called for each device which takes part of the replication. The function will be called with a first argument being the global index of the process within the replication, followed by the arguments passed in args.
args (tuple) – The arguments for fn. Default: Empty tuple
nprocs (python:int) – The number of processes/devices for the replication. At the moment, if specified, can be either 1 or the maximum number of devices.
join (bool) – Whether the call should block waiting for the completion of the processes which have being spawned. Default: True
daemon (bool) – Whether the processes being spawned should have the daemon flag set (see Python multi-processing API). Default: False
start_method (string) – The Python multiprocessing process creation method. Default: spawn

Returns

The same object returned by the torch.multiprocessing.spawn API. If nprocs is 1 the fn function will be called directly, and the API will return None.

class torch_xla.distributed.xla_multiprocessing.MpModelWrapper(model)[source]¶

Wraps a model to minimize host memory usage when fork method is used.

This class should be used together with the spawn(…, start_method=’fork’) API to minimize the use of host memory. Instead of creating models on each multiprocessing process, hence replicating the model’s initial host memory, the model is created once at global scope, and then moved into each device inside the spawn() target function. Example:

WRAPPED_MODEL = xmp.MpModelWrapper(MyNetwork())

def _mp_fn(index, ...):
  device = xm.xla_device()
  model = WRAPPED_MODEL.to(device)
  ...

xmp.spawn(_mp_fn, ..., start_method='fork')

This method has two advantages. First it uses only one copy of the memory pages to host the original model weights, and second it serializes the move of the wrapped model into each device, by lowering the load onto the system memory during the process.

to(device)[source]¶

Retrieves the model moved onto the specified device.

Parameters: device (torch.device) – The device where the model should be moved onto.
Returns: The model on the specified device.

class torch_xla.distributed.xla_multiprocessing.MpSerialExecutor[source]¶

Utility to run a function in a serialized fashion among multi-core processes.

Example:

# At global scope.
SERIAL_EXEC = xmp.MpSerialExecutor()

def load_dataset(path):
  return maybe_download_and_load(path)

def _mp_fn(index, ...):
  # Avoid all cores downloading the same data with the serial executor.
  dataset = SERIAL_EXEC.run(lambda: load_dataset('/tmp/mnist-data'))
  ...

xmp.spawn(_mp_fn, ...)

run(fn)[source]¶

Runs the provided function serialized WRT each per-core process.

Parameters: fn (callable) – The function to run in a serialized fashion.
Returns: The fn return value.

utils¶

class torch_xla.utils.tf_record_reader.TfRecordReader(path, compression='', buffer_size=16777216, transforms=None)[source]¶

Reads TfRecords or TfExamples.

Parameters

path (string) – The path to the file containing TfRecords.
compression (string, optional) – The compression type. The empty string for no compression, otherwise ZLIB or GZIP. Default: No compression.
buffer_size (python:int, optional) – The size of the buffer to be used to read TfRecords. Default: 16 * 1024 * 1024
transforms (dict, optional) – A dictionary with the key matching the TfExample label name, and value which is either a callable which will be called to tranform the matching tensor data, or STR for string conversion.

class torch_xla.utils.utils.SampleGenerator(data, sample_count)[source]¶

Iterator which returns multiple samples of a given input data.

Can be used in place of a PyTorch DataLoader to generate synthetic data.

Parameters

data – The data which should be returned at each iterator step.
sample_count – The maximum number of data samples to be returned.

class torch_xla.utils.utils.DataWrapper[source]¶: Utility class to wrap data structures to be sent to device.

torch_xla.utils.serialization.save(data, path, master_only=True, global_master=False)[source]¶

Saves the input data into a file.

The saved data is transferred to PyTorch CPU device before being saved, so a following torch.load() will load CPU data. Care must be taken when working with views. Instead of saving views it’s recommended that you recreate them after the tensors have been loaded and moved to their destination device(s).

Parameters

data – The input data to be saved. Any nested combination of Python objects (list, tuples, sets, dicts, …).
path – The destination file for the data saving operation. If master_only is False the path must point to different destinations as otherwise all the writes from the same host will override each other.
master_only (bool, optional) – Whether only the master device should save the data. If False, the path argument should be a different path for each of the ordinals taking part to the replication, otherwise all the replicas on the same host will be writing to the same location. Default: True
global_master (bool, optional) – When master_only is True this flag controls whether every host’s master (if global_master is False) saves the content, or only the global master (ordinal 0). Default: False

torch_xla.utils.serialization.load(path)[source]¶

Loads data previously saved with the save() API.

Parameters: path (str) – The path passed to the save() API.
Returns: The loaded data.

torch_xla.utils.gcsfs.open(path, mode='r', encoding=None)[source]¶

Opens a Google Cloud Storage (GCS) file for reading or writing.

Parameters

path (string) – The GCS path of the file. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS bucket, and PATH is a / delimited path.
mode (string, optional) – The open mode, similar to the open() API. Default: ‘r’
encoding (string, optional) – The character encoding to be used to decode bytes into strings when opening in text mode. Default: None

Returns

The GCS file object.

torch_xla.utils.gcsfs.list(path)[source]¶

Lists the content of a GCS bucket.

Parameters: path (string) – The GCS path of the file. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS bucket, and PATH is a / delimited path.
Returns: A list of GcsBlob objects.

torch_xla.utils.gcsfs.stat(path)[source]¶

Fetches the information of a GCS file.

Parameters: path (string) – The GCS path of the file. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS bucket, and PATH is a / delimited path.
Returns: A GcsBlob object.

torch_xla.utils.gcsfs.remove(path)[source]¶

Removes a GCS blob.

Parameters: path (string) – The GCS path of the file. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS bucket, and PATH is a / delimited path.

torch_xla.utils.gcsfs.rmtree(path)[source]¶

Removes all the GCS blobs within a given path.

Parameters

path (string) –

The GCS path of the file pattern or folder. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS

bucket, and PATH is a / delimited path.

torch_xla.utils.gcsfs.read(path)[source]¶

Reads the whole content of a GCS blob.

Parameters: path (string) – The GCS path of the file. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS bucket, and PATH is a / delimited path.
Returns: The bytes stored within the GCS blob.

torch_xla.utils.gcsfs.write(path, content)[source]¶

Write a string/bytes or file into a GCS blob.

Parameters

path (string) – The GCS path of the file. Must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS bucket, and PATH is a / delimited path.
content (string, bytes or file object) – The content to be written into path.

torch_xla.utils.gcsfs.generic_open(path, mode='r', encoding=None)[source]¶

Opens a file (GCS or not) for reding or writing.

Parameters

path (string) –
The path of the file to be opened. If a GCS path, it must be “gs://BUCKET_NAME/PATH” where BUCKET_NAME is the name of the GCS

bucket, and PATH is a / delimited path.
mode (string, optional) – The open mode, similar to the open() API. Default: ‘r’
encoding (string, optional) – The character encoding to be used to decode bytes into strings when opening in text mode. Default: None

Returns

The opened file object.

torch_xla.utils.gcsfs.generic_read(path)[source]¶

Reads the whole content of the provided location.

Parameters: path (string) – The GCS path or local path to be read.
Returns: The bytes stored within the GCS blob or local file.

torch_xla.utils.gcsfs.generic_write(output_string, path, makedirs=False)[source]¶

Write a string/bytes or file into a GCS blob or local disk.

Depending on the path passed in, this API can write to local or GCS file. Checks if the path starts with the ‘gs://’ prefix, and uses open otherwise.

Parameters

output_string (string) – The string to be written to the output.
path (string) – The GCS path or local path of the output.
makedirs (bool) – Whether the path parent folders should be created if missing. Default: False

torch_xla.utils.gcsfs.is_gcs_path(path)[source]¶

Checks whether a path is a GCS path.

Parameters: path (string) – The path to be checked.
Returns: Whether path is a GCS path.

class torch_xla.utils.cached_dataset.CachedDataset(data_set, path, max_files_per_folder=1000, compress=True)[source]¶

Wraps an existing torch.utils.data.Dataset by providing file caching.

The CachedDataset can be used to trade the CPU/RAM resources required to process a raw dataset, with storage/network resources. Example:

train_dataset = datasets.MNIST(
    FLAGS.datadir,
    train=True,
    download=True,
    transform=transforms.Compose(
        [transforms.ToTensor(),
         transforms.Normalize((0.1307,), (0.3081,))]))
train_dataset = CachedDataset(train_dataset, FLAGS.dscache_dir)

The CachedDataset will transparently cache the original Dataset samples, so that every run after the first, will not trigger any more CPU/RAM usage related to the raw samples processing. Once a CachedDataset is fully cached, it can be exported (ie, tar.gz) and used in different machines. Just unpack the tar.gz and pass None as original Dataset: Example:

train_dataset = CachedDataset(None, FLAGS.dscache_dir)

To fully cache CachedDataset just run the warmup() API. A CachedDataset saved on GCS has the advantage to be able to be used from different machines without explicit exporting.

Parameters

data_set (torch.utils.data.Dataset) – The raw torch.utils.data.Dataset to be cached. It can be set to None in case all the input samples are stored within the path folder.
path (string) – The path where the dataset samples should be stored/loaded. The path needs to be writeable, unless all the samples are already stored. The path can be a GCS path (prefixed with gs://).
max_files_per_folder (python:int) – The maximum amount of files to be stored within a single folder. If data_set is None this value is ignored and taken from the cached metadata. Default: 1000
compress (bool) – Whether the saved samples should be compressed. Compression saves space at the expense of CPU required to compress/decompress. If data_set is None this value is ignored and taken from the cached metadata. Default: True

test¶

Troubleshooting¶

Note that the information in this section is subject to be removed in future releases of the PyTorch/XLA software, since many of them are peculiar to a given internal implementation which might change.

To diagnose issues, we can use the execution metrics and counters provided by PyTorch/XLA The first thing to check when model is slow is to generate a metrics report.

Metrics report is extremely helpful in diagnosing issues. Please try to include it in your bug report sent to us if you have it.

Perform A Auto-Metrics Analysis¶

We provide ways to automatically analyze the metrics report and provide a summary. Simply run your workload with PT_XLA_DEBUG=1. Some example output would be

pt-xla-profiler: CompileTime too frequent: 21 counts during 11 steps
pt-xla-profiler: TransferFromServerTime too frequent: 11 counts during 11 steps
pt-xla-profiler: Op(s) not lowered: aten::_ctc_loss, aten::_ctc_loss_backward,  Please open a GitHub issue with the above op lowering requests.
pt-xla-profiler: CompileTime too frequent: 23 counts during 12 steps
pt-xla-profiler: TransferFromServerTime too frequent: 12 counts during 12 steps

Following section will explain how to get and understand a more detial metrics report.

Get A Metrics Report¶

Put the following line in your program to generate a report:

import torch_xla.debug.metrics as met

# For short report that only contains a few key metrics.
print(met.short_metrics_report())
# For full report that includes all metrics.
print(met.short_metrics_report())

Understand The Metrics Report¶

The report includes things like:

how many time we issue XLA compilations and time spent on issuing.
how many times we execute and time spent on execution
how many device data handles we create/destroy etc.

This information is reported in terms of percentiles of the samples. An example is:

Metric: CompileTime
  TotalSamples: 202
  Counter: 06m09s401ms746.001us
  ValueRate: 778ms572.062us / second
  Rate: 0.425201 / second
  Percentiles: 1%=001ms32.778us; 5%=001ms61.283us; 10%=001ms79.236us; 20%=001ms110.973us; 50%=001ms228.773us; 80%=001ms339.183us; 90%=001ms434.305us; 95%=002ms921.063us; 99%=21s102ms853.173us

We also provide counters, which are named integer variables which track internal software status. For example:

Counter: CachedSyncTensors
  Value: 395

In this report, any counter that starts with aten:: indicates a context switch between the XLA device and CPU, which can be a potential performance optimization area in the model code.

Counters are useful to understand which operations are routed back to the CPU engine of PyTorch. They are fully qualified with their C++ namespace:

Counter: aten::nonzero
  Value: 33

If you see aten:: ops other than nonzero and _local_scalar_dense, that usually means a missing lowering in PyTorch/XLA. Feel free to open a feature request for it on GitHub issues.

Clar The Metrics Report¶

If you want to clear the metrics between steps/epoches, you can use

import torch_xla.debug.metrics as met

met.clear_all()

Performance Profiling¶

To profile your workload in depth to undertand bottlenecks please check the following resources:

Known Performance Caveats¶

PyTorch/XLA behaves semantically like regular PyTorch and XLA tensors share the full tensor interface with CPU & GPU tensors. However, constraints in XLA/hardware and the lazy evaluation model suggest certain patterns might result in bad performance.

If your model shows bad performance, keep in mind the following caveats:

XLA/TPU yield degraded performance with too many recompilations.

XLA compilation is expensive. PyTorch/XLA automatically recompiles the graph every time new shapes are encountered. Usually models should stabilize within a few steps and you can see huge speedup for the rest of training.

In order to avoid recompilations, not only must shapes be constant, but computations across XLA devices in all hosts should also be constant.

Possible sources:
- Direct or indirect uses of nonzero introduce dynamic shapes; for example, masked indexing base[index] where index is a mask tensor.
- Loops with a different number of iterations between steps can result in different execution graphs, thus require recompilations.
Solution:
- Tensor shapes should be the same between iterations, or a low number of shape variations should be used.
- Pad tensors to fixed sizes when possible.
Certain operations don’t have native translations to XLA.

For these operations PyTorch/XLA automatically transfers to the CPU memory, evaluates on CPU, and transfers the result back to the XLA device. Doing too many such operations during the training step can lead to significant slowdowns.

Possible sources:
- The item() operation explicitly asks to evaluate the result. Don’t use it unless it’s necessary.
Solution:
- For most ops we can lower them to XLA to fix it. Checkout metrics report section to find out the missing ops and open a feature request on GitHub.
- Even when a PyTorch tensor is known as a scalar, avoid using tensor.item(). Keep it as a tensor and use tensor operations on it.
- Use torch.where to substitute control flow when applicable. E.g. The control flow with item() used in clip_grad*norm* is problematic and impacts performance, so we have patched clip_grad_norm_ by calling torch.where instead, which gives us a dramatic performance improvement. .. code-block:: python
  
  … else:
  
  device = parameters[0].device total_norm = torch.zeros([], device=device if parameters else None) for p in parameters:
  
  param_norm = p.grad.data.norm(norm_type) ** norm_type total_norm.add_(param_norm)
  
  total_norm = (total_norm ** (1. / norm_type))
  
  clip_coef = torch.tensor(max_norm, device=device) / (total_norm + 1e-6) for p in parameters:
  
  p.grad.data.mul_(torch.where(clip_coef < 1, clip_coef, torch.tensor(1., device=device)))
Iterators in ``torch_xla.distributed.data_parallel`` may drop the last few batches in the input iterator.

This is to make sure we do the same amount of work on all XLA devices.

Solution:
- When dataset is small, and there are too few steps, this may result in a no-op epoch. Therefore, it is better to use small batch sizes in those cases.

XLA Tensor Quirks¶

XLA tensor internals are opaque. XLA tensors always appear to be contiguous and without storage. Networks should not try to check the strides of XLA tensors.
XLA tensors should be moved to the CPU before saving them. Saving XLA tensors directly causes them to be loaded back on the device(s) they were saved from. If a device is unavailable at load time then the load will fail. Moving XLA tensors to the CPU before saving them lets you decide which device(s) to put the loaded tensors on. This is necessary if you want to load the tensors on a machine without XLA devices. Care should be taken moving the XLA tensors to the CPU before saving them, however, as moving tensors across device types does not preserve view relationships. Instead, views should be reconstructed as necessary after the tensors are loaded.
Copying an XLA Tensor with Python’s copy.copy returns a deep copy, not a shallow copy. Use a view of an XLA tensor to get a shallow copy of it.
Handling shared weights. Modules can share weights by setting the Parameters of one module to another. This “tying” of module weights should be done AFTER the modules are moved to an XLA device. Otherwise two independent copies of the shared tensor will be made on the XLA device.

More Debugging Tools¶

We don’t expect users to use tools in this section to debug their models. But we might ask for them when you submit a bug report since they provide additional information that metrics report doesn’t have.

Environment Variables¶

There are also a number of environment variables which control the behavior of the PyTorch/XLA software stack.

Setting such variables will cause different degrees of performance degradation, so they should only be enabled for debugging.

XLA_IR_DEBUG: Enables the Python stack trace to be captured where creating IR nodes, hence allowing to understand which PyTorch operation was responsible for generating the IR.
XLA_HLO_DEBUG: Enables the Python stack frame captured when _XLA_IRDEBUG is active, to be propagated to the XLA HLO metadata.
XLA_SAVE_TENSORS_FILE: The path to a file which will be used to dump the IR graphs during execution. Note that the file can become really big if the option is left enabled and the PyTorch program let run for long time. The graphs are appended to the file, so to have a clean sheet from run to run, the file should be explicitly removed.
XLA_SAVE_TENSORS_FMT: The format of the graphs stored within the _XLA_SAVE_TENSORSFILE file. Can be text (the default), dot (the Graphviz format) or hlo.
XLA_METRICS_FILE: If set, the path to a local file where the internal metrics will be saved at every step. Metrics will be appended to the file, if already existing.
XLA_SAVE_HLO_FILE: If set, the path to a local file where, in case of compilation/execution error, the offending HLO graph will be saved.
XLA_GET_TENSORS_OPBYOP: Enables pure OpByOp dispatch. The PyTorch/XLA software tries to fuse together many PyTorch operations into a single computation graph, but sometimes, either for debugging, or in case the PyTorch code have a very dynamic nature (in shapes or graph terms), it is better to force the execution in OpByOp mode (every IR node is lowered into a separate XLA computation, and chain-executed). This environment variable, if set to 1, enables OpByOp during the “get tensors” operation (the operation used by PyTorch/XLA to fetch intermediate values back from the TPU device into PyTorch CPU tensors).
XLA_SYNC_TENSORS_OPBYOP: The same as _XLA_GET_TENSORSOPBYOP but for “sync tensors” operation (the operation used at the end of a step, to flush pending IR computations and materialize them into TPU device data).
XLA_SYNC_WAIT: Forces the XLA tensor sync operation to wait for its completion, before moving to the next step.
XLA_USE_BF16: If set to 1, tranforms all the PyTorch Float values into BiFloat16 when sending to the TPU device. Note that when using XLA_USE_BF16=1 tensor arithmetic will be done in reduced precision and so tensors will not be accurate if accumulated over time. For example:
```
# In reduced bfloat16 precision
>>> torch.tensor(4096, dtype=torch.bfloat16) + torch.tensor(1, dtype=torch.bfloat16)
tensor(4096., dtype=torch.bfloat16)
# Whereas in full float32 precision
>>> torch.tensor(4096) + torch.tensor(1)
tensor(4097)
```
So to get accurate metrics such as average loss value over many steps, use manual mixed precision where metrics stay in FP32.
XLA_USE_F16: If set to 1, tranforms all the PyTorch Float values into Float16 (PyTorch Half type) when sending to devices which supports them.
XLA_USE_32BIT_LONG: If set to 1, maps PyTorch Long types to XLA 32bit type. On the versions of the TPU HW at the time of writing, 64bit integer computations are expensive, so setting this flag might help. It should be verified by the user that truncating to 32bit values is a valid operation according to the use of PyTorch Long values in it.
TF_CPP_LOG_THREAD_ID: If set to 1, the TF logs will show the thread ID helping with debugging multithreaded processes.

TF_CPP_VMODULE: Environment variable used for TF VLOGs and takes the form of TF_CPP_VMODULE=name=value,.... Note that for VLOGs you must set TF_CPP_MIN_LOG_LEVEL=0. For PyTorch/XLA using a configuration like TF_CPP_VMODULE=tensor=5 would enable logging such as:

2019-10-03 17:23:56.419040: I   27891 torch_xla/csrc/tensor.cpp:1104]
Executing IR graph hash 4211381954965020633 on device TPU:3 done!
2019-10-03 17:23:56.419448: I   27890 torch_xla/csrc/tensor.cpp:1104]
Executing IR graph hash 15483856951158150605 on device TPU:5 done!
2019-10-03 17:23:56.419539: I   27896 torch_xla/csrc/tensor.cpp:1104]
Executing IR graph hash 4211381954965020633 on device TPU:4 done!
...

TF_CPP_MIN_LOG_LEVEL: Level to print messages for. TF_CPP_MIN_LOG_LEVEL=0 will turn on INFO logging, TF_CPP_MIN_LOG_LEVEL=1 WARNING and so on. Our PyTorch/XLA TF_VLOG uses tensorflow::INFO level by default so to see VLOGs set TF_CPP_MIN_LOG_LEVEL=0.
XLA_DUMP_HLO_GRAPH: If set to =1 in case of a compilation or execution error the offending HLO graph will be dumped as part of the runtime error raised by xla_util.cc.

Retrieving Stack Traces¶

In the event that the PyTorch process is hanging, it might be useful to include the stack traces together with the GitHub issue.

First thing is to find out which PID the PyTorch process is associated with. Using the ps command it is possible to find that information. It will be a python process running your main python file.

In order to allow GDB to attach a user process the following command should be run as root:

echo 0 > /proc/sys/kernel/yama/ptrace_scope

The above command remains active until the machine is rebooted.

The, given the PID, it is possible to grab the stack traces with the following command:

./scripts/dump_stacks.py PID > /tmp/stack-traces.log

Using debug_run.py To Collect Debug Information¶

A utility is provided in scripts/debug_run.py which can be used to create a tar.gz archive with the information required to debug PyTorch/XLA executions.

Example:

./scripts/debug_run.py --outfile /tmp/debug_run.tar.gz -- python -u SCRIPT [ARGS...]

The python -u flag is suggested to disable buffering so that captured logs are correctly interleaved (otherwise STDOUT will be rendered after all STDERR).

The above command line example will leave the temporary folder containing the archived information on the filesystem. Use the --tidy flag to have that removed on exit:

./scripts/debug_run.py --tidy --outfile /tmp/debug_run.tar.gz -- python -u SCRIPT [ARGS...]

The debug_run.tar.gz file should then be attached to bug reports when necessary.

Since the script will collect a lot of data, it should usually be let run for no more than hundred steps or so.

If the SCRIPT has arguments to control the number of steps, those should be used, otherwise hitting CTRL^C will interrupt the run.

It is also sugested to run in single-core mode, to minimize the amount of data. Running in single-core mode is also strongly suggested when debugging execution issues.

Common Issues¶

Missing XLA configuration error message: You need to set XRT_TPU_CONFIG if using TPUs. If using GPUs set GPU_NUM_DEVICES=N for N number of GPUs. If using CPUs set XRT_DEVICE_MAP="CPU:0;/job:localservice/replica:0/task:0/device:XLA_CPU:0" and XRT_WORKERS="localservice:0;grpc://localhost:9002"

PJRT Runtime (Beta)¶

This document reflects the current state of PJRT support in current nightly builds. See the same document on the r2.0 branch for the status in the latest stable release.

The PyTorch/XLA team is currently migrating from the currently-supported XRT runtime to the PJRT runtime used by JAX.

PJRT is available for preview in PyTorch/XLA 2.0. We are planning to make PJRT our officially supported runtime, so we encourage all users to experiment with it. We aim to make PJRT stable in release 2.1, so if you encounter a bug with PJRT, please file an issue on GitHub with the runtime tag.

New features in PyTorch/XLA r2.0:

PJRT will be configured by default if you don’t pass in any other runtime configuration. If you continue to set XRT configuration (XRT_TPU_CONFIG), this change has no impact
New TPU runtime implementation in libtpu improves performance by up to 30%.
New xm.rendezvous implementation that scales to thousands of TPU cores
[experimental] torch.distributed support for TPU v2 and v3, including pjrt:// init_method
[experimental] Single-host GPU support in PJRT. Multi-host support coming soon!

TL;DR¶

To use the PJRT preview runtime, set the PJRT_DEVICE environment variable to CPU, TPU, orGPU`
In XRT, all distributed workloads are multiprocess, with one process per device. On TPU v2 and v3 in PJRT, workloads are multiprocess and multithreaded (4 processes with 2 threads each), so your workload should be thread-safe. See Multithreading on TPU v2/v3 and the Multiprocessing section of the API guide for more information. Key differences to keep in mind:
- To initialize a model in a thread-safe way, either broadcast the parameters across replicas after initialization (torch_xla.experimental.pjrt.broadcast_master_param) or load each replica’s parameters from a common checkpoint.
- For other random number generation, use torch.Generator where possible. The global torch RNG is not thread-safe, even if you set the same torch.manual_seed across replicas.
- To use torch.distributed, import torch_xla.experimental.pjrt_backend and use the pjrt:// init_method.
- These steps are optional for GPU and TPU v4.

Sample diff from XRT to PJRT:

 import os

 import torch
 import torch.nn as nn
 from torch.nn.parallel import DistributedDataParallel as DDP
 import torch.optim as optim
 import torch.distributed as dist
 import torch_xla.core.xla_model as xm
 import torch_xla.distributed.parallel_loader as pl
 import torch_xla.distributed.xla_backend
 import torch_xla.distributed.xla_multiprocessing as xmp
+import torch_xla.experimental.pjrt_backend
+import torch_xla.experimental.pjrt as pjrt


 def _mp_fn(index):
   device = xm.xla_device()
-  dist.init_process_group('xla', rank=xm.get_ordinal(), world_size=xm.xrt_world_size())
+  dist.init_process_group('xla', init_method='pjrt://')

   torch.manual_seed(42)
   model = nn.Linear(128, 10).to(device)

+  # Optional for TPU v4 and GPU
+  pjrt.broadcast_master_param(model)
   model = DDP(model, gradient_as_bucket_view=True)

   loss_fn = nn.MSELoss()
   optimizer = optim.SGD(model.parameters(), lr=.001)

   for i in range(10):
     data, target = torch.randn((128, 128), device=device), torch.randn((128, 10), device=device)

     optimizer.zero_grad()
     output = model(data)
     loss = loss_fn(output, target)
     loss.backward()

     optimizer.step()
     xm.mark_step()

   # Print mean parameters so we can confirm they're the same across replicas
   print([p.mean() for p in model.parameters()])

 if __name__ == '__main__':
-  os.environ['XRT_TPU_CONFIG'] = 'localservice;0;localhost:51011'
-  os.environ['MASTER_ADDR'] = 'localhost'
-  os.environ['MASTER_PORT'] = '12355'

+  # Recommended: set PJRT_DEVICE to your local device type
+  os.environ['PJRT_DEVICE'] = 'TPU'

   xmp.spawn(_mp_fn)

Benefits¶

Simple runtime configuration: just set PJRT_DEVICE to TPU, CPU, or GPU and start using XLA! Or, let PJRT select a device automatically based on your environment.
Improved performance: reduced overhead from gRPC means faster end-to-end execution. On TorchBench 2.0, we observed a >35% improvement in training time on TPU v4.
Easy pod execution: just copy your code to each TPU worker, and execute them all at the same time with gcloud compute tpus tpuvm ssh --worker=all.
Better scaling: removes XRT’s limitation on parameter sizes and supports up to 2048 TPU chips.

Quickstart¶

To start using PJRT with PyTorch/XLA, all you need to do is set the PJRT_DEVICE environment variable. If you’re working on a TPU v2 or v3, keep reading to learn about the differences between TPU v2 and v3 and v4.

CPU¶

On any machine with PyTorch/XLA installed, you can run our MNIST example on CPU like this:

PJRT_DEVICE=CPU python3 xla/test/test_train_mp_mnist.py --fake_data

TPU¶

To create a new TPU with PyTorch/XLA r1.13 installed:

gcloud alpha compute tpus tpu-vm create $USER-pjrt --accelerator-type=v4-8 --version=tpu-vm-v4-pt-1.13 --zone=us-central2-b --project=$PROJECT

On a v4-8, you can run our ResNet50 example like this:

git clone --depth=1 --branch r1.13 https://github.com/pytorch/xla.git
PJRT_DEVICE=TPU python3 xla/test/test_train_mp_imagenet.py --fake_data --batch_size=256 --num_epochs=1

By default, PJRT will use all TPU chips. To use only one TPU chip, configure TPU_PROCESS_BOUNDS and TPU_VISIBLE_CHIPS:

TPU_PROCESS_BOUNDS=1,1,1 TPU_VISIBLE_CHIPS=0 PJRT_DEVICE=TPU python3 xla/test/test_train_mp_imagenet.py --fake_data --batch_size=256 --num_epochs=1

Pods¶

On TPU Pods, use gcloud to run your command on each TPU in parallel:

gcloud alpha compute tpus tpu-vm ssh $USER-pjrt --zone=us-central2-b --project=$PROJECT --worker=all --command="git clone --depth=1 --branch r1.13 https://github.com/pytorch/xla.git"
gcloud alpha compute tpus tpu-vm ssh $USER-pjrt --zone=us-central2-b --project=$PROJECT --worker=all --command="PJRT_DEVICE=TPU python3 xla/test/test_train_mp_imagenet.py --fake_data --batch_size=256 --num_epochs=1"

Docker¶

You can also use Docker to run your workload in a container with PyTorch/XLA preinstalled:

export DOCKER_IMAGE=gcr.io/...

# Optional: authenticate docker if your image is in a private GCP repository
gcloud compute tpus tpu-vm ssh $USER-pjrt --zone=us-central2-b --project=$PROJECT --worker=all --command "sudo gcloud auth configure-docker"

# Run your workload
gcloud compute tpus tpu-vm ssh $USER-pjrt --zone=us-central2-b --project=$PROJECT --worker=all --command "sudo docker run --rm --privileged --net=host -e PJRT_DEVICE=TPU $DOCKER_IMAGE python pytorch/xla/test/test_train_mp_imagenet.py --fake_data"

Note that docker run requires privileged access to the host (--privileged) to expose the TPU device to the container. Docker on TPU pods is only supported with host networking --net=host at this time. See the Cloud TPU documentation for more information.

GPU¶

Warning: GPU support is still highly experimental!

To use GPUs with PJRT, simply set PJRT_DEVICE=GPU and configure GPU_NUM_DEVICES to the number of devices on the host. For example:

PJRT_DEVICE=GPU GPU_NUM_DEVICES=4 python3 xla/test/test_train_mp_imagenet.py --fake_data --batch_size=128 --num_epochs=1

Currently, only a single host is supported, and multi-host GPU cluster support will be added in an future release.

Differences from XRT¶

Although in most cases we expect PJRT and XRT to work mostly interchangeably from the end-user’s perspective (especially on TPU v4), there are some subtle differences that are important to keep in mind. Importantly, XRT was designed around the TPU Node architecture, so it will always spawn a client and a server process, even on TPU VMs. Thus, every batch of inputs has additional latency from serializing and deserializing data to send it over the network.

PJRT uses the local device directly with no intermediate server process. In the default configuration, PJRT will create one process per TPU chip, or 4 processes per TPU host. See the Cloud TPU documentation for more information about TPU architecture.

Performance gains are possible for workloads constrained overhead from .
Under XRT, the server process is the only process that interacts with the TPU devices, and client processes don’t have direct access to the TPU devices. When profiling a single-host TPU (e.g. v3-8 or v4-8), you would normally see 8 device traces (one for each TPU core). With PJRT, each process has one chip, and a profile from that process will show only 2 TPU cores.
- For the same reason, profiling does not work on TPU Pods with XRT, because the server process runs independently from the user’s model code. PJRT does not have that constraint, so it is possible to profile 2 TPU cores per process in a TPU Pod.
PJRT only supports the TPU VM architecture and we have no plans to support the TPU Node architecture with PJRT.
Runtime configuration is significantly simpler with PJRT. xla_dist is not required to run TPU Pod workloads. Instead, copy your code to each TPU host ([gcloud compute tpus tpu-vm scp](https://cloud.google.com/sdk/gcloud/reference/alpha/compute/tpus/tpu-vm/scp)) and run the code on each host in parallel (e.g. [gcloud compute tpus tpu-vm ssh --workers=all --command="PJRT_DEVICE=TPU python run.py"](https://cloud.google.com/sdk/gcloud/reference/alpha/compute/tpus/tpu-vm/ssh))
xm.rendezvous has been reimplemented using XLA-native collective communication to enhance stability on large TPU pods. See below for more details.

Multithreading on TPU v2/v3¶

On TPU v2 and v3, distributed workloads always run multithreaded, since each TPU core exposes two TPU cores as devices and only one process may open a TPU chip at a time. In its default configuration, xmp.spawn automatically spawns as many processes as possible (4 per TPU host) and creates two threads per process (one per TPU core).

Note: on TPU v4, each TPU chip is represented as one PyTorch device, so distributed workloads will run across 4 processes, each with only one thread. This is identical to XRT’s behavior.

In most cases, this will not require substantial changes to your existing code. The main change you will have to make in most cases is to model initialization. Because torch’s global RNG is shared between threads, results will vary between threads and runs even if you set torch.manual_seed to the same value in every replica. To get consistent parameters between replicas, either use torch_xla.experimental.pjrt.broadcast_master_param to broadcast one replica’s parameters to all other replicas, or load each replica’s parameters from a common checkpoint.

Changes to xm.rendezvous¶

New in PyTorch/XLA r2.0

With XRT, worker 0 runs a mesh master service, and all processes on all workers connect to that service over gRPC. In practice, we found that running a single mesh master process was unreliable on TPU pods with thousands of chips due to the number of inbound connections to worker 0. A single client process timing out could cause a failure and force the entire workload to restart.

Thus, we have reimplemented xm.rendezvous with native XLA collective communication, which is much more stable and well-tested on large TPU pods. This imposes two new constraints compared to the XRT implementation:

Because the payload has to become part of the XLA graph, xm.mark_step is called both before and after the data is transferred. Calling xm.rendezvous in the middle of model code may force an unwanted compilation.
Because XLA does not permit collective operations to run on a subset of workers, all workers must participate in the rendezvous.

If you require the old behavior of xm.rendezvous (i.e. communicating data without altering the XLA graph and/or synchronizing a subset of workers), consider using ``torch.distributed.barrier` <https://pytorch.org/docs/stable/distributed.html#torch.distributed.barrier>`_ or [torch.distributed.all_gather_object](https://pytorch.org/docs/stable/distributed.html#torch.distributed.all_gather_object) with a gloo process group. If you are also using the xla torch.distributed backend, you can use torch.new_group to create a gloo subgroup. See this example from the PyTorch documentation. Keep in mind these constraints:

torch.distributed is not fully supported on TPU v2/v3. Only a subset of operations with the xla backend are implemented, and gloo will likely not work as expected in a multiprocessing context.
In our experiments, gloo does not scale well to thousands of TPU chips, so expect this alternative to be less reliable than using xm.rendezvous with PJRT at large scales.

PJRT and torch.distributed¶

New in PyTorch/XLA r2.0

When using PJRT with torch.distributed and [torch.nn.parallel.DistributedDataParallel](https://github.com/pytorch/xla/blob/master/docs/ddp.md) we strongly recommend using the new pjrt:// init_method, which automatically finds the replica IDs, world size, and master IP by querying the runtime. For example:

import torch
import torch.distributed as dist
import torch_xla.core.xla_model as xm
import torch_xla.distributed.xla_multiprocessing as xmp
from torch_xla.experimental import pjrt

# Required for `pjrt://` init_method
import torch_xla.experimental.pjrt_backend

def _all_gather(index: int):
  # No need to pass in `rank` or `world_size`
  dist.init_process_group('xla', init_method='pjrt://')

  t = torch.tensor([index], dtype=torch.int32, device=xm.xla_device())
  output = [torch.zeros_like(t) for _ in range(dist.get_world_size())]
  dist.all_gather(output, t)

  xm.mark_step()
  print(output)

if __name__ == '__main__':
  xmp.spawn(_all_gather)

Note: Although the pjrt:// init_method is not required on TPU v4, it is still recommended. If you use env://, MASTER_ADDR must be set to IP host that has device 0, which is not always worker 0. The pjrt:// init_method finds this IP automatically and supports TPU v2/v3.

For more information about using DistributedDataParallel on PyTorch/XLA, see ``ddp.md` <./ddp.md>`_ on TPU V4. For an example that uses DDP and PJRT together, run the following example script on a TPU:

PJRT_DEVICE=TPU python xla/test/test_train_mp_mnist.py --ddp --pjrt_distributed --fake_data --num_epochs 1

Performance¶

TorchBench shows improvements in average training time across tasks with PJRT compared to XRT, with an average improvement of over 35% on TPU v4-8. The benefits vary significantly by task and model type, ranging from 0% to 175%. The following chart shows the breakdown by task:

New TPU runtime¶

New in PyTorch/XLA r2.0

The PyTorch/XLA r2.0 release introduces support for the PJRT Plugin API, used to access the new TFRT-based TPU runtime in libtpu. This is now the default runtime when PJRT_DEVICE=TPU is set. The legacy StreamExecutor-based TPU runtime used in 1.13 will still be available with PJRT_DEVICE=TPU_LEGACY in the 2.0 release, but it will be removed in a future version. If you encounter an issue that only happens on TPU and not TPU_LEGACY, please file an issue on GitHub.

In most cases, we expect performance to be similar between the two runtimes, but in some cases, the new runtime may be up to 30% faster. The following chart shows the breakdown by task:

Note: the improvements shown in this chart are also included in the PJRT vs XRT comparison.

TorchDynamo(torch.compile) integration in PyTorch XLA¶

Torchdynamo is a Python-level JIT compiler designed to make unmodified PyTorch programs faster. It provides a clean API for compiler backends to hook in and its biggest feature is to dynamically modify Python bytecode right before it is executed. In the pytorch/xla 2.0 release, PyTorch/XLA provided an experimental backend for the TorchDynamo for both inference and training.

The way that XLA bridge works is that Dynamo will provide a TorchFX graph when it recognizes a model pattern and PyTorch/XLA will use existing Lazy Tensor technology to compile the FX graph and return the compiled function.

Inference¶

Here is a small code example of running resnet18 with torch.compile

import torch
imprt torchvision
import torch_xla.core.xla_model as xm

def eval_model(loader):
  device = xm.xla_device()
  xla_resnet18 = torchvision.models.resnet18().to(device)
  xla_resnet18.eval()
  dynamo_resnet18 = torch.compile(
      xla_resnet18, backend='torchxla_trace_once')
  for data, _ in loader:
    output = dynamo_resnet18(data)

NOTE: inference backend name torchxla_trace_once is subject to change.

With the torch.compile you will see that PyTorch/XLA only traces the resent18 model once during the init time and executes the compiled binary everytime dynamo_resnet18 is invoked, instead of tracing the model every time. Note that currently Dynamo does not support fallback so if there is an op that can not be traced by XLA, it will error out. We will fix this issue in the upcoming 2.1 release. Here is a inference speed analysis to compare Dynamo and Lazy using torch bench on Cloud TPU v4-8

Training¶

PyTorch/XLA also supports Dynamo for training, but it is very experimental and we are working with the PyTorch Compiler team to iterate on the implementation. On the 2.0 release it only supports forward and backward pass but not the optimizer. Here is an example of training a resnet18 with torch.compile

import torch
imprt torchvision
import torch_xla.core.xla_model as xm

def train_model(model, data, target):
  loss_fn = torch.nn.CrossEntropyLoss()
  pred = model(data)
  loss = loss_fn(pred, target)
  loss.backward()
  return pred

def train_model_main(loader):
  device = xm.xla_device()
  xla_resnet18 = torchvision.models.resnet18().to(device)
  xla_resnet18.train()
  dynamo_train_model = torch.compile(
        train_model, backend='aot_torchxla_trace_once')
  for data, target in loader:
    output = dynamo_train_model(xla_resnet18, data, target)

NOTE: Backend we used here is aot_torchxla_trace_once(subject to change) instead of torchxla_trace_once

We expect to extract and execute 3 graphs per training step instead of one training step if you use the Lazy tensor. Here is a training speed analysis to compare Dynamo and Lazy using a torch bench on Cloud TPU v4-8.

NOTE: We run each model’s fwd and bwd for a single step and then collect the e2e time. In the real world we will run multiple steps at each training job which can easily hide the tracing cost from execution(since it is async). Lazy Tensor will have much better performance in that scenario.

We are currently working on the optimizer support and that will be availiable on nightly soon but won’t be in the 2.0 release.

Take away¶

TorchDynamo provides a really promising way for the compiler backend to hide the complexity from the user and easily retrieve the modeling code in a graph format. Compared with PyTorch/XLA’s traditional Lazy Tensor way of extracting the graph, TorchDynamo can skip the graph tracing for every iteration hence provide a much better inference response time. However TorchDynamo does not trace the communication ops(like all_reduce and all_gather) yet and it provides separate graphs for the forward and the backward which hurts xla performance. These feature gaps compared to Lazy Tensor makes it less efficient in real world training use cases, especially the tracing cost can be overlapped with the execution in training. The PyTorch/XLA team will keep investing in TorchDynamo and work with upstream to mature the training story.

Fully Sharded Data Parallel (FSDP) in PyTorch XLA¶

Fully Sharded Data Parallel (FSDP) in PyTorch XLA is a utility for sharding Module parameters across data-parallel workers.

Example usage:

import torch
import torch_xla.core.xla_model as xm
from torch_xla.distributed.fsdp import XlaFullyShardedDataParallel as FSDP

model = FSDP(my_module)
optim = torch.optim.Adam(model.parameters(), lr=0.0001)
output = model(x, y)
loss = output.sum()
loss.backward()
optim.step()

It is also possible to shard individual layers separately and have an outer wrapper handle any leftover parameters.

Notes:

The XlaFullyShardedDataParallel class supports both the ZeRO-2 optimizer (sharding gradients and optimizer states) and the ZeRO-3 optimizer (sharding parameters, gradients, and optimizer states) in https://arxiv.org/abs/1910.02054.
- The ZeRO-3 optimizer should be implemented via nested FSDP with reshard_after_forward=True. See test/test_train_mp_mnist_fsdp_with_ckpt.py and test/test_train_mp_imagenet_fsdp.py for an example.
- For large models that cannot fit into a single TPU memory or the host CPU memory, one should interleave submodule construction with inner FSDP wrapping. See ``FSDPViTModel` <https://github.com/ronghanghu/vit_10b_fsdp_example/blob/master/run_vit_training.py>`_ for an example.
a simple wrapper checkpoint_module is provided (based on torch_xla.utils.checkpoint.checkpoint from https://github.com/pytorch/xla/pull/3524) to perform gradient checkpointing over a given nn.Module instance. See test/test_train_mp_mnist_fsdp_with_ckpt.py and test/test_train_mp_imagenet_fsdp.py for an example.
Auto-wrapping submodules: instead of manually nested FSDP wrapping, one can also specify an auto_wrap_policy argument to automatically wrap the submodules with inner FSDP. size_based_auto_wrap_policy in torch_xla.distributed.fsdp.wrap is an example of auto_wrap_policy callable, this policy wraps layers with the number of parameters larger than 100M. transformer_auto_wrap_policy in torch_xla.distributed.fsdp.wrap is an example of auto_wrap_policy callable for transformer-like model architectures.

For example, to automatically wrap all torch.nn.Conv2d submodules with inner FSDP, one can use:

from torch_xla.distributed.fsdp.wrap import transformer_auto_wrap_policy
auto_wrap_policy = partial(transformer_auto_wrap_policy, transformer_layer_cls={torch.nn.Conv2d})

Additionally, one can also specify an auto_wrapper_callable argument to use a custom callable wrapper for the submodules (the default wrapper is just the XlaFullyShardedDataParallel class itself). For example, one can use the following to apply gradient checkpointing (i.e. activation checkpointing/rematerialization) to each auto-wrapped submodule.

from torch_xla.distributed.fsdp import checkpoint_module
auto_wrapper_callable = lambda m, *args, **kwargs: XlaFullyShardedDataParallel(
    checkpoint_module(m), *args, **kwargs)

When stepping the optimizer, directly call optimizer.step and do not call xm.optimizer_step. The latter reduces the gradient across ranks, which is not needed for FSDP (where the parameters are already sharded).
When saving model and optimizer checkpoints during training, each training process needs to save its own checkpoint of the (sharded) model and optimizer state dicts (use master_only=False and set different paths for each rank in xm.save). When resuming, it needs to load the checkpoint for the corresponding rank.
Please also save model.get_shard_metadata() along with model.state_dict() as follows and use consolidate_sharded_model_checkpoints to stitch the sharded model checkpoints together into a full model state dict. See test/test_train_mp_mnist_fsdp_with_ckpt.py for an example. .. code-block:: python3

ckpt = {
‘model’: model.state_dict(), ‘shard_metadata’: model.get_shard_metadata(), ‘optimizer’: optimizer.state_dict(),

} ckpt_path = f’/tmp/rank-{xm.get_ordinal()}-of-{xm.xrt_world_size()}.pth’ xm.save(ckpt, ckpt_path, master_only=False)
The checkpoint consolidation script can also be launched from the command line as follows. .. code-block:: bash

# consolidate the saved checkpoints via command line tool python3 -m torch_xla.distributed.fsdp.consolidate_sharded_ckpts –ckpt_prefix /path/to/your_sharded_checkpoint_files –ckpt_suffix “_rank--of-.pth”

The implementation of this class is largely inspired by and mostly follows the structure of fairscale.nn.FullyShardedDataParallel in https://fairscale.readthedocs.io/en/stable/api/nn/fsdp.html. One of the biggest differences from fairscale.nn.FullyShardedDataParallel is that in XLA we don’t have explicit parameter storage, so here we resort to a different approach to free full parameters for ZeRO-3.

Example training scripts on MNIST and ImageNet¶

MNIST: ``test/test_train_mp_mnist_fsdp_with_ckpt.py` <https://github.com/pytorch/xla/blob/master/test/test_train_mp_mnist_fsdp_with_ckpt.py>`_ (it also tests checkpoint consolidation)
ImageNet: ``test/test_train_mp_imagenet_fsdp.py` <https://github.com/pytorch/xla/blob/master/test/test_train_mp_imagenet_fsdp.py>`_

Installation¶

FSDP is available on PyTorch/XLA 1.12 release and newer nightly. Please refer to https://github.com/pytorch/xla#-available-images-and-wheels for installation guide.

Clone PyTorch/XLA repo¶

git clone --recursive https://github.com/pytorch/pytorch
cd pytorch/
git clone --recursive https://github.com/pytorch/xla.git
cd ~/

Train MNIST on v3-8 TPU¶

It gets around 98.9 accuracy for 2 epochs:

python3 ~/pytorch/xla/test/test_train_mp_mnist_fsdp_with_ckpt.py \
  --batch_size 16 --drop_last --num_epochs 2 \
  --use_nested_fsdp --use_gradient_checkpointing

This script automatically tests checkpoint consolidation at the end. You can also manually consolidate the sharded checkpoints via

# consolidate the saved checkpoints via command line tool
python3 -m torch_xla.distributed.fsdp.consolidate_sharded_ckpts \
  --ckpt_prefix /tmp/mnist-fsdp/final_ckpt \
  --ckpt_suffix "_rank-*-of-*.pth"

Train ImageNet with ResNet-50 on v3-8 TPU¶

It gets around 75.9 accuracy for 100 epochs; download ImageNet-1k to /datasets/imagenet-1k:

python3 ~/pytorch/xla/test/test_train_mp_imagenet_fsdp.py \
  --datadir /datasets/imagenet-1k --drop_last \
  --model resnet50 --test_set_batch_size 64 --eval_interval 10 \
  --lr 0.4 --batch_size 128 --num_warmup_epochs 5 --lr_scheduler_divide_every_n_epochs 30 --lr_scheduler_divisor 10 --num_epochs 100 \
  --use_nested_fsdp

You can also add --use_gradient_checkpointing (which needs to be used along with --use_nested_fsdp or --auto_wrap_policy) to apply gradient checkpointing on the residual blocks.

Example training scripts on TPU pod (with 10 billion parameters)¶

To train large models that cannot fit into a single TPU, one should apply auto-wrap or manually wrap the submodules with inner FSDP when building the entire model to implement the ZeRO-3 algorithm.

Please see https://github.com/ronghanghu/vit_10b_fsdp_example for an example of sharded training of a Vision Transformer (ViT) model using this XLA FSDP PR.

How to do `DistributedDataParallel`¶

This document shows how to use torch.nn.parallel.DistributedDataParallel in xla, and further describes its difference against the native xla data parallel approach.

Background / Motivation¶

Customers have long requested the ability to use PyTorch’s DistributedDataParallel API with xla. And here we enable it as an experimental feature.

How to use DistributedDataParallel¶

For those who switched from the PyTorch eager mode to XLA, here are all the changes you need to do to convert your eager DDP model into XLA model. We assume that you already know how to use XLA on a single device.

Import xla specific distributed packages:

import torch_xla.core.xla_model as xm
import torch_xla.distributed.xla_backend

Init xla process group similar to other process groups such as nccl and gloo.

dist.init_process_group("xla", rank=rank, world_size=world_size)

Use xla specific APIs to get rank and world_size if you need to.

new_rank = xm.get_ordinal()
world_size = xm.xrt_world_size()

Pass gradient_as_bucket_view=True to the DDP wrapper.

ddp_model = DDP(model, gradient_as_bucket_view=True)

Finally launch your model with xla specific launcher.

xmp.spawn(demo_fn)

Here we have put everything together (the example is actually taken from the DDP tutorial). The way you code it is pretty similar to the eager experience. Just with xla specific touches on a single device plus the above five changes to your script.

import os
import sys
import tempfile
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.optim as optim

from torch.nn.parallel import DistributedDataParallel as DDP

# additional imports for xla
import torch_xla.core.xla_model as xm
import torch_xla.distributed.xla_backend
import torch_xla.distributed.xla_multiprocessing as xmp

def setup(rank, world_size):
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = '12355'

    # initialize the xla process group
    dist.init_process_group("xla", rank=rank, world_size=world_size)

def cleanup():
    dist.destroy_process_group()

class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(10, 1000000)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(1000000, 5)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))

def demo_basic(rank):
    # xla specific APIs to get rank, world_size.
    new_rank = xm.get_ordinal()
    assert new_rank == rank
    world_size = xm.xrt_world_size()

    print(f"Running basic DDP example on rank {rank}.")
    setup(rank, world_size)

    # create model and move it to XLA device
    device = xm.xla_device()
    model = ToyModel().to(device)
    # currently, graident_as_bucket_view is needed to make DDP work for xla
    ddp_model = DDP(model, gradient_as_bucket_view=True)

    loss_fn = nn.MSELoss()
    optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)

    optimizer.zero_grad()
    outputs = ddp_model(torch.randn(20, 10).to(device))
    labels = torch.randn(20, 5).to(device)
    loss_fn(outputs, labels).backward()
    optimizer.step()
    # xla specific API to execute the graph
    xm.mark_step()

    cleanup()


def run_demo(demo_fn):
    # xla specific launcher
    xmp.spawn(demo_fn)

if __name__ == "__main__":
    run_demo(demo_basic)

Benchmarking¶

Resnet50 with fake data¶

The following results are collected with the command: python test/test_train_mp_imagenet.py --fake_data --model=resnet50 --num_epochs=1 on a TPU VM V3-8 environment with ToT PyTorch and PyTorch/XLA. And the statistical metrics are produced by using the script in this pull request. The unit for the rate is images per second.

Type	Mean	Median	90th %	Std Dev	CV
xm.optimizer_step	418.54	419.22	430.40	9.76	0.02
DDP	395.97	395.54	407.13	7.60	0.02

The performance difference between our native approach for distributed data parallel and DistributedDataParallel wrapper is: 1 - 395.97 / 418.54 = 5.39%. This result seems reasonable given the DDP wrapper introduces extra overheads on tracing the DDP runtime.

MNIST with fake data¶

The following results are collected with the command: python test/test_train_mp_mnist.py --fake_data on a TPU VM V3-8 environment with ToT PyTorch and PyTorch/XLA. And the statistical metrics are produced by using the script in this pull request. The unit for the rate is images per second.

Type	Mean	Median	90th %	Std Dev	CV
xm.optimizer_step	17864.19	20108.96	24351.74	5866.83	0.33
DDP	10701.39	11770.00	14313.78	3102.92	0.29

The performance difference between our native approach for distributed data parallel and DistributedDataParallel wrapper is: 1 - 14313.78 / 24351.74 = 41.22%. Here we compare 90th % instead since the dataset is small and first a few rounds are heavily impacted by data loading. This slowdown is huge but makes sense given the model is small. The additional DDP runtime tracing overhead is hard to amortize.

MNIST with real data¶

The following results are collected with the command: python test/test_train_mp_mnist.py --logdir mnist/ on a TPU VM V3-8 environment with ToT PyTorch and PyTorch/XLA.

And we can observe that the DDP wrapper converges slower than the native XLA approach even though it still achieves a high accuracy rate at 97.48% at the end. (The native approach achieves 99%.)

Disclaimer¶

This feature is still experimental and under active development. Use it in cautions and feel free to file any bugs to the xla github repo. For those who are interested in the native xla data parallel approach, here is the tutorial.

Here are some of the known issues that are under investigation:

gradient_as_bucket_view=True needs to be enforced.
There are some issues while being used with torch.utils.data.DataLoader. test_train_mp_mnist.py with real data crashes before exiting.

How to run with PyTorch/XLA:GPU¶

PyTorch/XLA enables PyTorch users to utilize the XLA compiler which supports accelerators including TPU, GPU, CPU and … This doc will go over the basic steps to run PyTorch/XLA on a nvidia gpu instance

Create a GPU instance¶

Pytorch/XLA currently publish prebuilt docker images and wheels with cuda11.2 and python 3.7/3.8. We recommend users to create a GPU instance with corresponding config. For a full list of docker images and wheels, please refer to this doc.

Environment Setup¶

Docker¶

sudo docker pull gcr.io/tpu-pytorch/xla:nightly_3.8_cuda_11.2
sudo apt-get install -y apt-transport-https ca-certificates curl gnupg-agent    software-properties-common
distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add -
curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list
sudo apt-get update && sudo apt-get install -y nvidia-container-toolkit
sudo systemctl restart docker
sudo docker run --gpus all -it -d gcr.io/tpu-pytorch/xla:nightly_3.7\8_cuda_11.2 bin/bash
sudo docker exec -it $(sudo docker ps | awk 'NR==2 { print $1 }') /bin/bash

Note that you need to restart the docker to make gpu devices visible in the docker container. After logging into the docker, you can use nvidia-smi to verify the device is setup correctly.

(pytorch) root@20ab2c7a2d06:/# nvidia-smi
Thu Dec  8 06:24:29 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.47.03    Driver Version: 510.47.03    CUDA Version: 11.6     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  Off  | 00000000:00:04.0 Off |                    0 |
| N/A   36C    P0    38W / 300W |      0MiB / 16384MiB |      1%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+

Wheel¶

pip3 install torch=1.13
pip3 install https://storage.googleapis.com/tpu-pytorch/wheels/cuda/112/torch_xla-1.13-cp37-cp37m-linux_x86_64.whl

Run a simple model¶

In order to run below examples, you need to clone the pytorch/xla repo to access the imagenet example(We already clone it in our docker).

(pytorch) root@20ab2c7a2d06:/# export GPU_NUM_DEVICES=1
(pytorch) root@20ab2c7a2d06:/# python pytorch/xla/test/test_train_mp_imagenet.py --fake_data
==> Preparing data..
Epoch 1 train begin 06:12:38
2022-12-08 06:13:12.452874: W      79 tensorflow/compiler/xla/service/gpu/gpu_conv_algorithm_picker.cc:729] None of the algorithms provided by cuDNN heuristics worked; trying fallback algorithms.  Conv: (f32[128,256,28,28]{3,2,1,0}, u8[0]{0}) custom-call(f32[128,256,14,14]{3,2,1,0}, f32[3,3,256,256]{1,0,2,3}), window={size=3x3 stride=2x2 pad=1_1x1_1}, dim_labels=bf01_01io->bf01, custom_call_target="__cudnn$convBackwardInput", backend_config="{\"conv_result_scale\":1,\"activation_mode\":\"0\",\"side_input_scale\":0}"
2022-12-08 06:13:13.780992: W      79 tensorflow/compiler/xla/service/gpu/gpu_conv_algorithm_picker.cc:729] None of the algorithms provided by cuDNN heuristics worked; trying fallback algorithms.  Conv: (f32[128,128,56,56]{3,2,1,0}, u8[0]{0}) custom-call(f32[128,128,28,28]{3,2,1,0}, f32[3,3,128,128]{1,0,2,3}), window={size=3x3 stride=2x2 pad=1_1x1_1}, dim_labels=bf01_01io->bf01, custom_call_target="__cudnn$convBackwardInput", backend_config="{\"conv_result_scale\":1,\"activation_mode\":\"0\",\"side_input_scale\":0}"
| Training Device=xla:0/0 Epoch=1 Step=0 Loss=6.89059 Rate=2.82 GlobalRate=2.82 Time=06:13:23
| Training Device=xla:0/0 Epoch=1 Step=20 Loss=6.79297 Rate=117.16 GlobalRate=45.84 Time=06:13:36
| Training Device=xla:0/0 Epoch=1 Step=40 Loss=6.43628 Rate=281.16 GlobalRate=80.49 Time=06:13:43
| Training Device=xla:0/0 Epoch=1 Step=60 Loss=5.83108 Rate=346.88 GlobalRate=108.82 Time=06:13:49
| Training Device=xla:0/0 Epoch=1 Step=80 Loss=4.99023 Rate=373.62 GlobalRate=132.43 Time=06:13:56
| Training Device=xla:0/0 Epoch=1 Step=100 Loss=3.92699 Rate=384.33 GlobalRate=152.40 Time=06:14:02
| Training Device=xla:0/0 Epoch=1 Step=120 Loss=2.68816 Rate=388.35 GlobalRate=169.49 Time=06:14:09

AMP (AUTOMATIC MIXED PRECISION)¶

AMP is very useful on GPU training and PyTorch/XLA reuse Cuda’s AMP rule. You can checkout our mnist example and imagenet example. Note that we also used a modified version of optimizers to avoid the additional sync between device and host.

PyTorch on XLA Devices¶

Creating an XLA Tensor¶

XLA Tensors are PyTorch Tensors¶

Running Models on XLA Devices¶

Running on a Single XLA Device¶

Running on Multiple XLA Devices with Multi-processing¶

Running on TPU Pods¶

XLA Tensor Deep Dive¶

XLA Tensors are Lazy¶

XLA Tensors and bFloat16¶

Memory Layout¶

Moving XLA Tensors to and from the CPU¶

Saving and Loading XLA Tensors¶

Further Reading¶

PyTorch/XLA API¶

xla_model¶

distributed¶

utils¶

test¶

Troubleshooting¶

Perform A Auto-Metrics Analysis¶

Get A Metrics Report¶

Understand The Metrics Report¶

Clar The Metrics Report¶

Performance Profiling¶

Known Performance Caveats¶

XLA Tensor Quirks¶

More Debugging Tools¶

Environment Variables¶

Retrieving Stack Traces¶

Using debug_run.py To Collect Debug Information¶

Common Issues¶

PJRT Runtime (Beta)¶

TL;DR¶

Benefits¶

Quickstart¶

CPU¶

TPU¶

Pods¶

Docker¶

GPU¶

Differences from XRT¶

Multithreading on TPU v2/v3¶

Changes to xm.rendezvous¶

PJRT and torch.distributed¶

Performance¶

New TPU runtime¶

TorchDynamo(torch.compile) integration in PyTorch XLA¶

Inference¶

Training¶

Take away¶

Fully Sharded Data Parallel (FSDP) in PyTorch XLA¶

Example training scripts on MNIST and ImageNet¶

Installation¶

Clone PyTorch/XLA repo¶

Train MNIST on v3-8 TPU¶

Train ImageNet with ResNet-50 on v3-8 TPU¶

Example training scripts on TPU pod (with 10 billion parameters)¶

How to do DistributedDataParallel¶

Background / Motivation¶

How to use DistributedDataParallel¶

Benchmarking¶

Resnet50 with fake data¶

MNIST with fake data¶

MNIST with real data¶

Disclaimer¶

How to run with PyTorch/XLA:GPU¶

Create a GPU instance¶

Environment Setup¶

Docker¶

Wheel¶

Run a simple model¶

AMP (AUTOMATIC MIXED PRECISION)¶

Docs

Tutorials

Resources

How to do `DistributedDataParallel`¶