Facilitating New Backend Integration by PrivateUse1

Created On: Oct 03, 2023 | Last Updated: May 07, 2024 | Last Verified: Nov 05, 2024

In this tutorial we will walk through some necessary steps to integrate a new backend living outside pytorch/pytorch repo by PrivateUse1. Note that this tutorial assumes that you already have a basic understanding of PyTorch. you are an advanced user of PyTorch.

Note

This tutorial only involves the parts related to the PrivateUse1 mechanism that facilitates the integration of new devices, and other parts will not be covered. At the same time, not all the modules involved in this tutorial are required, and you can choose the modules that are helpful to you according to your actual needs.

What is PrivateUse1?

Prior to Pytorch 2.0, PyTorch provided three reserved dispatch keys (and their corresponding Autograd keys) for prototyping out-of-tree backend extensions, the three dispatch keys are as follows:

• PrivateUse1/AutogradPrivateUse1

• PrivateUse2/AutogradPrivateUse2

• PrivateUse3/AutogradPrivateUse3

After the prototype verification is passed, you can apply for a private key for the new backend, such as CUDA, XLA, MPS, and so on.

However, with the rapid development of PyTorch, more and more hardware manufacturers are trying to integrate their backends into PyTorch, which might cause the following problems:

• Every new backend integration involves a lot of file modification

• There is currently a hard limit on the number of Dispatch Keys (DispatchKeySet 64-bit limit)

Note

There is also a problem with integrating the new backend into PyTorch through the PrivateUse1 Key, as it is impossible to integrate many backends at the same time. Fortunately, these out-of-tree backends are rarely used simultaneously.

In view of the above reasons, the community began to recommend new backend to be integrated into the PyTorch via PrivateUse1.

However, the previous PrivateUse1 mechanism is not fully capable of integrating with the new backend, because it lacks some related support in certain modules, such as Storage, AMP, Distributed, and so on.

With the arrival of Pytorch 2.1.0, a series of optimizations and enhancements have been made for PrivateUse1 in terms of new backend integration, and it is now possible to support the integration of new devices rapidly and efficiently.

How to integrate new backend via PrivateUse1

In this section, we will discuss the details of integrating the new backend into Pytorch via PrivateUse1, which mainly consists of the following parts:

1. Register kernels for the new backend.

2. Register generator for the new backend.

3. Register device guard for the new backend.

4. Register serialization and deserialization functions for new backend metadata.

5. Other Modules.

Register kernels for the new backend

The new backend may have some high-performance implementations of operator, which can be registered to the dispatcher by TORCH_LIBRARY_IMPL API described in Registering a Dispatched Operator in C++. This involves several situations:

1. Register all the forward operators supported by the new backend to the dispatcher, and register the fallback at the same time, so that when the new backend does not support some operators, these operators can fall back to the CPU for execution to ensure the availability of functions.

at::Tensor wrapper_Custom_Tensor_add(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha) {
  // Implementation of add kernel in new backend
  ...
}

TORCH_LIBRARY_IMPL(aten, PrivateUse1, m) {
  ...
  m.impl("add.Tensor", TORCH_FN(wrapper_Custom_Tensor_add));
  ...
}

void custom_cpu_fallback(const c10::OperatorHandle& op, torch::jit::Stack* stack) {
  // Add some hints about new devices that do not support and need to fall back to cpu
  at::native::cpu_fallback(op, stack);
}

TORCH_LIBRARY_IMPL(_, PrivateUse1, m) {
  m.fallback(torch::CppFunction::makeFromBoxedFunction<&custom_cpu_fallback>());
}

2. Register kernels from torch::autograd::Function to the dispatcher by AutogradPrivateUse1, if it is necessary for new backend to override PyTorch Autograd layer, the dispatcher and autograd system will automatically call the forward and backward implementations of these operators.

class CumtomSeluFunction : public torch::autograd::Function<CumtomSeluFunction> {
  // Implementation of selu kernel in new backend
}

at::Tensor wrapper_AutogradCumstom__selu(const at::Tensor & self) {
  return CumtomSeluFunction::apply(self);
}

TORCH_LIBRARY_IMPL(aten, AutogradPrivateUse1, m) {
  ...
  m.impl("selu", TORCH_FN(wrapper_AutogradCustom__selu));
  ...
}

3. Register kernels which want to support automatic mixed precision (AMP) and fallback mechanism to the dispatcher by AutocastPrivateUse1, the autocast system will automatically call these kernels when needed.

TORCH_LIBRARY_IMPL(aten, AutocastPrivateUse1, m) {
  ...
  KERNEL_PRIVATEUSEONE(<operator>, <policy>)
  ...
}

TORCH_LIBRARY_IMPL(_, AutocastPrivateUse1, m) {
  m.fallback(torch::CppFunction::makeFallthrough());
}

What needs to be added is that if you want to support AMP in a new backend, you need to register a new BackendModule by torch._register_device_module("backend_name", BackendModule), and the BackendModule needs to have the following APIs:

• get_amp_supported_dtype() -> List[torch.dtype]

  get the supported dtypes on the new backend in AMP, which might support one more dtype.

• is_autocast_enabled() -> bool

  check the AMP is enabled or not on the new backend.

• get_autocast_dtype() -> torch.dtype

  get the supported dtype on the new backend in AMP, which is set by set_autocast_dtype or the default dtype, and the default dtype is torch.float16.

• set_autocast_enabled(bool) -> None

  enable or disable AMP on the new backend.

• set_autocast_dtype(dtype) -> None

  set the supported dtype on the new backend in AMP, and the dtype be contained in the dtypes got from get_amp_supported_dtype.

Register generator for the new backend

It is necessary to support generators corresponding to new devices. Currently, PrivateUse1 can dynamically register custom generators, which are mainly divided into the following steps.

1. Inherit the GeneratorImpl class to implement the generator class corresponding to the new backend, and implement various general methods.

2. Define a new backend builder with a single parameter: device index.

3. Call REGISTER_GENERATOR_PRIVATEUSE1 macro to complete dynamic registration.

struct CustomGeneratorImpl : public c10::GeneratorImpl {
  // Implementation of generator in new backend
}

at::Generator make_custom_generator(c10::DeviceIndex device_index) {
  return at::make_generator<CustomGeneratorImpl>(device_index);
}

REGISTER_GENERATOR_PRIVATEUSE1(make_cumstom_generator)

Register device guard for the new backend

PyTorch provides functionalities related to device, stream, and event switching via DeviceGuard. This function is also applicable to PrivateUse1 Key.

1. Inherit the DeviceGuardImplInterface class to implement the various general methods corresponding to the new backend.

2. Call C10_REGISTER_GUARD_IMPL macro to complete dynamic registration.

struct CustomGuardImpl final : public c10::impl::DeviceGuardImplInterface {
  // Implementation of guard in new backend
}

C10_REGISTER_GUARD_IMPL(PrivateUse1, CustomGuardImpl);

Register serialization and deserialization functions for new backend metadata

PyTorch is currently able to dynamically register serialization/deserialization functions to support the serialization and deserialization of new backend additional metadata named backend_meta_ in class TensorImpl.ExtraMeta. You can refer to the following steps:

1. Inherit the BackendMeta class to implement CustomBackendMetadata corresponding to the new backend and various fields of the new backend can be customized in the class.

2. Implement the serialization and deserialization functions of the new backend, the function signatures are void(const at::Tensor&, std::unordered_map<std::string, bool>&).

3. Call the TensorBackendMetaRegistry macro to complete dynamic registration.

struct CustomBackendMetadata : public c10::BackendMeta {
  // Implementation of backend metadata in new backend
}

void for_serialization(const at::Tensor& t, std::unordered_map<std::string, bool>& m) {
  // Implementation of serialization
}

void for_deserialization(const at::Tensor& t, std::unordered_map<std::string, bool>& m) {
  // Implementation of deserialization
}

TensorBackendMetaRegistry(c10::DeviceType::PrivateUse1, &for_serialization, &for_deserialization);

Other Modules

In addition to the above-mentioned parts, there are some other modules that can be expanded through PrivateUse1, such as distributed collective communication, benchmark timer, and others, which will be added in the future. One example about PrivateUse1 integration is Ascend NPU.

How to Improve User Experience with Privateuse1

The primary goal of integrating new devices through PrivateUse1 is to meet the basic functional requirements, and the next thing to do is to improve usability, which mainly involves the following aspects.

1. Register new backend module to Pytorch.

2. Rename PrivateUse1 to a custom name for the new backend.

3. Generate methods and properties related to the new backend.

Register new backend module to Pytorch

Some CUDA-related interfaces in PyTorch can be called through the following form: torch.cuda.xxx. Therefore, in order to comply with user habits, the new backend implemented through the PrivateUse1 mechanism should also provide similar interfaces.

For example, using Ascend NPU:

torch._register_device_module('npu', torch_npu.npu)

After doing the above operations, users can call some exclusive APIs of Ascend NPU through torch.npu.xxx

Rename PrivateUse1 to a custom name for the new backend

PrivateUse1 Key is the internal mechanism of the new backend integrated into PyTorch. For users, compared with PrivateUse1, the custom name strongly related to the new backend should be more friendly.

Taking the Ascend NPU as an example, the first usage will be more user-friendly.

torch.rand((2,2),device='npu:0')
torch.rand((2,2),device='privateuse1:0')

Now, PyTorch provides a new C++/Python API for the self-named PrivateUse1 backend, which is very simple to use.

PYTHON

torch.rename_privateuse1_backend("npu")

C++

c10::register_privateuse1_backend("npu")

Generate methods and properties related to the new backend

After renaming PrivateUse1 to a custome name, automatically generate properties and methods related to the new backend name in the Tensor, nn, Storage modules for the new backend.

Here is an example for Ascend NPU:

torch.rename_privateuse1_backend("npu")
unsupported_dtype = [torch.quint8]
torch.utils.generate_methods_for_privateuse1_backend(for_tensor=True, for_module=True, for_storage=True, unsupported_dtype=unsupported_dtype)

Then, you can use the following methods and properties:

torch.Tensor.npu()
torch.Tensor.is_npu
torch.Storage.npu()
torch.Storage.is_npu
...

Future Work

The improvement of the PrivateUse1 mechanism is still in progress, so the integration method of PrivateUse1 of the new module will be added in turn. Here are a few items that we are actively working on:

• Add the integration method of distributed collective communication.

• Add the integration method of benchmark timer.

Conclusion

This tutorial walked you through the process of integrating new backends into PyTorch via PrivateUse1, including but not limited to operator registration, generator registration, device guard registration, and so on. At the same time, some methods are introduced to improve the user experience.