text
stringlengths 0
1.73k
| source
stringlengths 35
119
| category
stringclasses 2
values |
|---|---|---|
Extracting bigger segments of the model into an intermediate representation makes it possible to do sophisticated whole-program optimizations and to offload computation to specialized AI accelerators which operate on graphs of computation. We have already been developing the beginnings of these optimizations, including passes that fuse GPU operations together to improve the performance of smaller RNN models.
It also allows us to use existing high-performance backends available in Caffe2 today to run the model efficiently. Additionally, @script functions (and modules!) can be fully exported to ONNX in a way that retains their dynamic nature, such that you can easily run them in a Python-free environment using the model executors from Caffe2 or by transferring the model to any other framework supporting ONNX.
Usability | https://pytorch.org/blog/the-road-to-1_0/ | pytorch blogs |
Usability
We care deeply about maintaining our current level of usability and we know that execution of the code not directly in Python leads to harder debugging, but this is something that we think about a lot, and we're making sure that you're not getting locked in to a completely different programming language. | https://pytorch.org/blog/the-road-to-1_0/ | pytorch blogs |
First, we follow the principle of pay for what you use — if you don't need to optimize or export your model, you do not have to use these new features and won't see any downsides. Furthermore, use of traced or @script modules/functions can be done incrementally. For instance, all of these behaviors are allowed: You can trace part of your model and use the trace in a larger non-traced model. You can use tracing for 90% of your model, and use @script for the one sub-module that actually has some control flow in it. You can write a function using @script and have it call a native python function. If something appears incorrect in an @script function, you can remove the annotation and the code will execute in native python where it is easy to debug using your favorite tools and methods. Think of tracing and @script like type annotations using MyPy or TypeScript — each additional annotation can be tested incrementally, and none are required until you want to optimize or productionize. | https://pytorch.org/blog/the-road-to-1_0/ | pytorch blogs |
Most importantly, these modes will be built into the core of PyTorch so that mixing and matching them with your existing code can happen seamlessly.
Note: The name JIT for these components is a bit of a misnomer and comes from historical reasons. The tracing/function execution in PyTorch started out as an optimizing JIT compiler that generated fused CUDA kernels but then grew to encompass optimization, @script, and export. When it is ready for release we will likely rename this functionality to the hybrid frontend, but we wanted to present it here as it is named in the code so that you can follow along as we develop it.
Other changes and improvements
Production support is the big feature for 1.0, but we will continue optimizing and fixing other parts of PyTorch as course of the standard release process. | https://pytorch.org/blog/the-road-to-1_0/ | pytorch blogs |
On the backend side of things, PyTorch will see some changes, which might affect user-written C and C++ extensions. We are replacing (or refactoring) the backend ATen library to incorporate features and optimizations from Caffe2.
Last Words
We aim to release 1.0 some time during the summer. You can follow-along our progress on the Pull Requests page.
You can read this from the perspective of the Caffe2 project at: https://caffe2.ai/blog/2018/05/02/Caffe2_PyTorch_1_0.html | https://pytorch.org/blog/the-road-to-1_0/ | pytorch blogs |
layout: blog_detail
title: 'PyTorch adds new tools and libraries, welcomes Preferred Networks to its community'
author: Team PyTorch
PyTorch continues to be used for the latest state-of-the-art research on display at the NeurIPS conference next week, making up nearly 70% of papers that cite a framework. In addition, we’re excited to welcome Preferred Networks, the maintainers of the Chainer framework, to the PyTorch community. Their teams are moving fully over to PyTorch for developing their ML capabilities and services.
This growth underpins PyTorch’s focus on building for the needs of the research community, and increasingly, supporting the full workflow from research to production deployment. To further support researchers and developers, we’re launching a number of new tools and libraries for large scale computer vision and elastic fault tolerant training. Learn more on GitHub and at our NeurIPS booth. | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
Preferred Networks joins the PyTorch community
Preferred Networks, Inc. (PFN) announced plans to move its deep learning framework from Chainer to PyTorch. As part of this change, PFN will collaborate with the PyTorch community and contributors, including people from Facebook, Microsoft, CMU, and NYU, to participate in the development of PyTorch.
PFN developed Chainer, a deep learning framework that introduced the concept of define-by-run (also referred to as eager execution), to support and speed up its deep learning development. Chainer has been used at PFN since 2015 to rapidly solve real-world problems with the latest, cutting-edge technology. Chainer was also one of the inspirations for PyTorch’s initial design, as outlined in the PyTorch NeurIPS paper. | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
PFN has driven innovative work with CuPy, ImageNet in 15 minutes, Optuna, and other projects that have pushed the boundaries of design and engineering. As part of the PyTorch community, PFN brings with them creative engineering capabilities and experience to help take the framework forward. In addition, PFN’s migration to PyTorch will allow it to efficiently incorporate the latest research results to accelerate its R&D activities, given PyTorch’s broad adoption with researchers, and to collaborate with the community to add support for PyTorch on MN-Core, a deep learning processor currently in development. | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
We are excited to welcome PFN to the PyTorch community, and to jointly work towards the common goal of furthering advances in deep learning technology. Learn more about the PFN’s migration to PyTorch here.
Tools for elastic training and large scale computer vision
PyTorch Elastic (Experimental)
Large scale model training is becoming commonplace with architectures like BERT and the growth of model parameter counts into the billions or even tens of billions. To achieve convergence at this scale in a reasonable amount of time, the use of distributed training is needed. | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
The current PyTorch Distributed Data Parallel (DDP) module enables data parallel training where each process trains the same model but on different shards of data. It enables bulk synchronous, multi-host, multi-GPU/CPU execution of ML training. However, DDP has several shortcomings; e.g. jobs cannot start without acquiring all the requested nodes; jobs cannot continue after a node fails due to error or transient issue; jobs cannot incorporate a node that joined later; and lastly; progress cannot be made with the presence of a slow/stuck node.
The focus of PyTorch Elastic, which uses Elastic Distributed Data Parallelism, is to address these issues and build a generic framework/APIs for PyTorch to enable reliable and elastic execution of these data parallel training workloads. It will provide better programmability, higher resilience to failures of all kinds, higher-efficiency and larger-scale training compared with pure DDP. | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
Elasticity, in this case, means both: 1) the ability for a job to continue after node failure (by running with fewer nodes and/or by incorporating a new host and transferring state to it); and 2) the ability to add/remove nodes dynamically due to resource availability changes or bottlenecks.
While this feature is still experimental, you can try it out on AWS EC2, with the instructions here. Additionally, the PyTorch distributed team is working closely with teams across AWS to support PyTorch Elastic training within services such as Amazon Sagemaker and Elastic Kubernetes Service (EKS). Look for additional updates in the near future.
New Classification Framework | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
New Classification Framework
Image and video classification are at the core of content understanding. To that end, you can now leverage a new end-to-end framework for large-scale training of state-of-the-art image and video classification models. It allows researchers to quickly prototype and iterate on large distributed training jobs at the scale of billions of images. Advantages include:
Ease of use - This framework features a modular, flexible design that allows anyone to train machine learning models on top of PyTorch using very simple abstractions. The system also has out-of-the-box integration with AWS on PyTorch Elastic, facilitating research at scale and making it simple to move between research and production.
High performance - Researchers can use the framework to train models such as Resnet50 on ImageNet in as little as 15 minutes.
| https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
You can learn more at the NeurIPS Expo workshop on Multi-Modal research to production or get started with the PyTorch Elastic Imagenet example here.
Come see us at NeurIPS
The PyTorch team will be hosting workshops at NeurIPS during the industry expo on 12/8. Join the sessions below to learn more, and visit the team at the PyTorch booth on the show floor and during the Poster Session. At the booth, we’ll be walking through an interactive demo of PyTorch running fast neural style transfer on a Cloud TPU - here’s a sneak peek. | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
We’re also publishing a paper that details the principles that drove the implementation of PyTorch and how they’re reflected in its architecture.
Multi-modal Research to Production - This workshop will dive into a number of modalities such as computer vision (large scale image classification and instance segmentation) and Translation and Speech (seq-to-seq Transformers) from the lens of taking cutting edge research to production. Lastly, we will also walk through how to use the latest APIs in PyTorch to take eager mode developed models into graph mode via Torchscript and quantize them for scale production deployment on servers or mobile devices. Libraries used include: | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
Classification Framework - a newly open sourced PyTorch framework developed by Facebook AI for research on large-scale image and video classification. It allows researchers to quickly prototype and iterate on large distributed training jobs. Models built on the framework can be seamlessly deployed to production.
Detectron2 - the recently released object detection library built by the Facebook AI Research computer vision team. We will articulate the improvements over the previous version including: 1) Support for latest models and new tasks; 2) Increased flexibility, to enable new computer vision research; 3) Maintainable and scalable, to support production use cases.
Fairseq - general purpose sequence-to-sequence library, can be used in many applications, including (unsupervised) translation, summarization, dialog and speech recognition.
| https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
Responsible and Reproducible AI - This workshop on Responsible and Reproducible AI will dive into important areas that are shaping the future of how we interpret, reproduce research, and build AI with privacy in mind. We will cover major challenges, walk through solutions, and finish each talk with a hands-on tutorial.
Reproducibility: As the number of research papers submitted to arXiv and conferences skyrockets, scaling reproducibility becomes difficult. We must address the following challenges: aid extensibility by standardizing code bases, democratize paper implementation by writing hardware agnostic code, facilitate results validation by documenting “tricks” authors use to make their complex systems function. To offer solutions, we will dive into tool like PyTorch Hub and PyTorch Lightning which are used by some of the top researchers in the world to reproduce the state of the art.
| https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
Interpretability: With the increase in model complexity and the resulting lack of transparency, model interpretability methods have become increasingly important. Model understanding is both an active area of research as well as an area of focus for practical applications across industries using machine learning. To get hands on, we will use the recently released Captum library that provides state-of-the-art algorithms to provide researchers and developers with an easy way to understand the importance of neurons/layers and the predictions made by our models.`
| https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
Private AI: Practical applications of ML via cloud-based or machine-learning-as-a-service platforms pose a range of security and privacy challenges. There are a number of technical approaches being studied including: homomorphic encryption, secure multi-party computation, trusted execution environments, on-device computation, and differential privacy. To provide an immersive understanding of how some of these technologies are applied, we will use the CrypTen project which provides a community based research platform to take the field of Private AI forward.
We’d like to thank the entire PyTorch team and the community for all their contributions to this work.
Cheers!
Team PyTorch | https://pytorch.org/blog/pytorch-adds-new-tools-and-libraries-welcomes-preferred-networks-to-its-community/ | pytorch blogs |
layout: blog_detail
title: "How Computational Graphs are Executed in PyTorch"
author: Preferred Networks
featured-img: ""
Welcome to the last entry into understanding the autograd engine of PyTorch series!
If you haven’t read parts 1 & 2 check them now to understand how PyTorch creates the computational graph for the backward pass!
This post is based on PyTorch v1.11, so some highlighted parts may differ across versions.
PyTorch autograd graph execution
The last post showed how PyTorch constructs the graph to calculate the outputs' derivatives w.r.t. the inputs when executing the forward pass. Now we will see how the execution of the backward pass is coordinated and done by looking at the whole process, starting from Python down to the lower C++ level internals.
What Happens when Calling backward()/grad() from Python
Using variable.backward() | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Using variable.backward()
After doing all our calculations with an input set to require the gradient, we call .backward() on the result to initiate the backward pass execution.
>>> x = torch.tensor([0.5, 0.75], requires_grad=True)
>>> y = torch.exp(x).sum()
>>> y.backward()
Calling .backward() on a tensor results in a call to torch.autograd.backward().
# torch/_tensor.py
def backward(self, gradient=None, retain_graph=None, create_graph=False, inputs=None):
…
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
torch.autograd.backward() checks the arguments and calls the autograd engine in the C++ layer.
``` python
def backward(
tensors: _TensorOrTensors, | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
def backward(
tensors: _TensorOrTensors,
grad_tensors: Optional[_TensorOrTensors] = None,
retain_graph: Optional[bool] = None,
create_graph: bool = False,
grad_variables: Optional[_TensorOrTensors] = None,
inputs: Optional[_TensorOrTensors] = None,
) -> None:
…
if inputs is not None and len(inputs) == 0:
raise RuntimeError("'inputs' argument to backward() cannot be empty.")
tensors = (tensors,) if isinstance(tensors, torch.Tensor) else tuple(tensors)
inputs = (inputs,) if isinstance(inputs, torch.Tensor) else \
tuple(inputs) if inputs is not None else tuple()
grad_tensors_ = _tensor_or_tensors_to_tuple(grad_tensors, len(tensors))
grad_tensors_ = _make_grads(tensors, grad_tensors_)
if retain_graph is None:
retain_graph = create_graph
Variable._execution_engine.run_backward(
tensors, grad_tensors_, retain_graph, create_graph, inputs,
allow_unreachable=True, accumulate_grad=True) # allow_unreachable flag
``` | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
``
First, whether thegrad_tensorsargument was specified or not, there is a call to the [_make_grads](https://github.com/pytorch/pytorch/blob/bc2c6edaf163b1a1330e37a6e34caf8c553e4755/torch/autograd/__init__.py#L30-L74) function. This is used to check the providedgrad_tensorsor to specify the default value for them by looking at thetensorsargument values’ shapes. Check the first blog post for details on the default value for thegrad_tensors` of the backward pass. This function just provides the vector of the vector jacobian product if it was not initially specified. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
In the above code, Variable has an _execution_engine attribute that is defined in torch.autograd.variable to be of type ImperativeEngine; the C++ engine exported to python and declared in torch/csrc/autograd/python_engine.cpp. In the following sections, we explain in detail how this object executes the backward pass.
Note that the torch.autograd.backward function has an inputs optional argument. This argument is used when we want to calculate the .grad field of only a subset of input tensors in the forward pass.
```python
x = torch.tensor([0.5, 0.75], requires_grad=True)
y = torch.tensor([0.1, 0.90], requires_grad=True)
z = torch.exp(x * y).sum()
torch.autograd.backward([z], inputs=[x])
x.grad
tensor([0.1051, 1.7676])
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
x.grad
tensor([0.1051, 1.7676])
y.grad # None
```
Using torch.autograd.grad
An alternative to backward() is to use torch.autograd.grad(). The main difference to backward() is that grad() returns a tuple of tensors with the gradients of the outputs w.r.t. the inputs kwargs instead of storing them in the .grad field of the tensors. As you can see, the grad() code shown below is very similar to backward.
```python
def grad(
outputs: _TensorOrTensors,
inputs: _TensorOrTensors,
grad_outputs: Optional[_TensorOrTensors] = None,
retain_graph: Optional[bool] = None,
create_graph: bool = False,
only_inputs: bool = True,
allow_unused: bool = False,
is_grads_batched: bool = False
) -> Tuple[torch.Tensor, ...]:
outputs = (outputs,) if isinstance(outputs, torch.Tensor) else tuple(outputs)
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
inputs = (inputs,) if isinstance(inputs, torch.Tensor) else tuple(inputs)
overridable_args = outputs + inputs
if has_torch_function(overridable_args):
return handle_torch_function(
grad,
overridable_args,
outputs,
inputs,
grad_outputs=grad_outputs,
retain_graph=retain_graph,
create_graph=create_graph,
only_inputs=only_inputs,
allow_unused=allow_unused,
)
grad_outputs_ = _tensor_or_tensors_to_tuple(grad_outputs, len(outputs))
grad_outputs_ = _make_grads(outputs, grad_outputs_)
if retain_graph is None:
retain_graph = create_graph
if is_grads_batched:
# …. It will not be covered here
else:
return Variable._execution_engine.run_backward(
outputs, grad_outputs_, retain_graph, create_graph, inputs,
allow_unused, accumulate_grad=False) # Calls into the C++ engine to run the backward pass
``` | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
```
Figure 1 shows the computational graph with the backward() and grad() arguments highlighted in red and blue, respectively:
Fgiure 1: Correspondence of `backward`/`grad` arguments in the graphs.
Going Inside the Autograd Engine
Refreshing Concepts: Nodes and Edges
As we saw in 2
The computational graph comprises Node and Edge objects. Please read that post if you haven’t done it yet.
Nodes | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Nodes
Node objects are defined in torch/csrc/autograd/function.h, and they provide an overload of operator() for the associated function and a list of edges to do the graph traversal. Note that Node is a base class that autograd functions inherit from and override the apply method to execute the backward function.
struct TORCH_API Node : std::enable_shared_from_this<Node> {
...
/// Evaluates the function on the given inputs and returns the result of the
/// function call.
variable_list operator()(variable_list&& inputs) {
...
}
protected:
/// Performs the `Node`'s actual operation.
virtual variable_list apply(variable_list&& inputs) = 0;
…
edge_list next_edges_;
uint64_t topological_nr_ = 0;
…
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
uint64_t topological_nr_ = 0;
…
```
There is an attribute called topological_nr_ in every node object. This number is used to optimize the graph execution as it allows to discard of graph branches under certain conditions. The topological number is the longest distance between this node and any leaf node and it is shown in Figure 2. Its main property is that for any pair of nodes x, y in a directed graph topo_nr(x) < topo_nr(y) means that there is no path from x to y. So this allows for reducing the number of paths in the graph in need of traversal. Check the topological_nr
) method comment for further details.
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Figure 2: Example of the Topological Number calculation
Edges
The Edge object links Nodes together, and its implementation is straightforward.
struct Edge {
...
/// The function this `Edge` points to.
std::shared_ptr<Node> function;
/// The identifier of a particular input to the function.
uint32_t input_nr;
};
It only requires a function pointer to the Node and an input number that is the index of the output from the forward function this edge points to. When preparing the set of gradients before calling "function", we know that what is flowing from this edge should be accumulated in the "input_nr"th argument. Note that the input/output name is flipped here and this is the input to the backward function. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Edge objects are constructed using the gradient_edge function method.
Edge gradient_edge(const Variable& self) {
if (const auto& gradient = self.grad_fn()) {
return Edge(gradient, self.output_nr());
} else {
return Edge(grad_accumulator(self), 0);
}
}
Entering the C++ Realm
Once that torch.autograd.backward() has been invoked, the
THPEngine_run_backward routine starts the graph traversal. Following is a schema of the function body:
```c++
PyObject THPEngine_run_backward(PyObject self, PyObject args, PyObject kwargs)
{
HANDLE_TH_ERRORS
PyObject tensors = nullptr;
PyObject grad_tensors = nullptr;
unsigned char keep_graph = 0;
unsigned char create_graph = 0;
PyObject *inputs = nullptr; | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
PyObject *inputs = nullptr;
// Convert the python arguments to C++ objects
const char accepted_kwargs[] = { // NOLINT
"tensors", "grad_tensors", "keep_graph", "create_graph", "inputs",
"allow_unreachable", "accumulate_grad", nullptr
};
if (!PyArg_ParseTupleAndKeywords(args, kwargs, "OObb|Obb", (char*)accepted_kwargs,
&tensors, &grad_tensors, &keep_graph, &create_graph, &inputs, &allow_unreachable, &accumulate_grad))
// Prepare arguments
for(const auto i : c10::irange(num_tensors)) {
// Check that the tensors require gradients
}
std::vector output_edges;
if (inputs != nullptr) {
// Prepare outputs
}
{
// Calls the actual autograd engine
pybind11::gil_scoped_release no_gil;
outputs = engine.execute(roots, grads, keep_graph, create_graph, accumulate_grad, output_edges);
}
// Clean up and finish
}
``` | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
// Clean up and finish
}
First, we prepare the input arguments after converting the `PyObject` arguments to actual C++ objects. The `tensors` list contains the tensors from which we start the backward pass. These tensors are converted to edges using `torch::autograd::impl::gradient_edge` and added to a list called `roots` where the graph traversal starts.
```c++
edge_list roots;
roots.reserve(num_tensors);
variable_list grads;
grads.reserve(num_tensors);
for(const auto i : c10::irange(num_tensors)) {
PyObject *_tensor = PyTuple_GET_ITEM(tensors, i);
const auto& variable = THPVariable_Unpack(_tensor);
auto gradient_edge = torch::autograd::impl::gradient_edge(variable);
roots.push_back(std::move(gradient_edge));
PyObject *grad = PyTuple_GET_ITEM(grad_tensors, i);
if (THPVariable_Check(grad)) {
const Variable& grad_var = THPVariable_Unpack(grad);
grads.push_back(grad_var);
}
}
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
grads.push_back(grad_var);
}
}
```
Now, if the inputs argument was specified in backward or we used the torch.autograd.grad api, the following code creates a list of edges to accumulate the gradients in the specified tensors at the end of the computation. The engine uses this later to optimize the execution as it doesn’t add the gradients in all the leaf nodes, just the specified ones.
```c++
std::vector output_edges;
if (inputs != nullptr) {
int num_inputs = PyTuple_GET_SIZE(inputs);
output_edges.reserve(num_inputs);
for (const auto i : c10::irange(num_inputs)) {
PyObject *input = PyTuple_GET_ITEM(inputs, i);
const auto& tensor = THPVariable_Unpack(input);
const auto output_nr = tensor.output_nr();
auto grad_fn = tensor.grad_fn();
if (!grad_fn) {
grad_fn = torch::autograd::impl::try_get_grad_accumulator(tensor);
}
if (accumulate_grad) {
tensor.retain_grad();
}
if (!grad_fn) { | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
if (!grad_fn) {
output_edges.emplace_back(std::make_shared(), 0);
} else {
output_edges.emplace_back(grad_fn, output_nr);
}
}
}
The next step is the actual graph traversal and node function execution, and finally, the cleanup and return.
```c++
{
// Calls the actual autograd engine
pybind11::gil_scoped_release no_gil;
auto& engine = python::PythonEngine::get_python_engine();
outputs = engine.execute(roots, grads, keep_graph, create_graph, accumulate_grad, output_edges);
}
// Clean up and finish
}
Starting the Real Execution
engine.executeis present in torch/csrc/autograd/engine.cpp
There are two differentiated steps here:
Analyze the graph to find dependencies between functions
Create worker threads that traverse the graph
Data Structures Used for the Execution
GraphTask | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
GraphTask
All the execution metadata is managed by the GraphTask class in torch/csrc/autograd/engine.h
struct GraphTask: std::enable_shared_from_this<GraphTask> {
std::atomic<uint64_t> outstanding_tasks_{0};
// …
std::unordered_map<Node*, InputBuffer> not_ready_;
std::unordered_map<Node*, int> dependencies_;
struct ExecInfo {
// …
};
std::unordered_map<Node*, ExecInfo> exec_info_;
std::vector<Variable> captured_vars_;
// …
std::shared_ptr<ReadyQueue> cpu_ready_queue_;
};
Here we see a series of variables dedicated to maintaining the execution state. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
outstanding_tasks_ tracks the number of tasks left to be executed for the backward pass to complete. not_ready_ holds the input arguments for the Nodes that are not ready to be executed. dependencies_ track the number of predecessors that a Node has. As the count reaches 0, the Node is ready for execution; it is placed in a ready queue to be retrieved and executed later.
exec_info_ and the associated ExecInfo struct are used only when the inputs argument is specified or it is a call to autograd.grad(). They allow filter paths on the graph that are not needeed since only the gradients are calculated only for the variables in the inputs list.
captured_vars_ is where the results of the graph execution are temporarily stored if we used the torch.autograd.grad() api instead of torch.autograd.backward() since grad() returns the gradients as tensors instead of just filling the .grad field of the inputs.
NodeTask | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
NodeTask
The NodeTask struct is a basic class that holds an fn_ pointer to the node to execute, and an inputs_ buffer to store the input arguments to this function. Note that the functions executed by the backward pass are the derivatives specified in the derivatives.yaml file. or the user provided backward function when using custom functions as described in the second blog post.
The inputs_ buffer is also where the output gradients of the previously executed functions are aggregated, and it is defined as a std::vector<Variable> container with facilities to accumulate values at a given position.
```c++
struct NodeTask {
std::weak_ptr base_;
std::shared_ptr fn_;
// This buffer serves as an implicit "addition" node for all of the | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
// gradients flowing here. Once all the dependencies are finished, we
// use the contents of this buffer to run the function.
InputBuffer inputs_;
};
### GraphRoot
The [`GraphRoot`](https://github.com/pytorch/pytorch/blob/bc2c6edaf163b1a1330e37a6e34caf8c553e4755/torch/csrc/autograd/functions/basic_ops.h#L72-L89) is a special function used to hold multiple input variables in a single place. The code is pretty simple as it only acts as a container of variables.
```c++
struct TORCH_API GraphRoot : public Node {
GraphRoot(edge_list functions, variable_list inputs)
: Node(std::move(functions)),
outputs(std::move(inputs)) {
for (const auto& t : outputs) {
add_input_metadata(t);
}
}
variable_list apply(variable_list&& inputs) override {
return outputs;
}
AccumulateGrad
This function is set during the graph creation in gradient_edge when the Variable object doesn’t have a grad_fn. This is, it is a leaf node.
```c++ | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
if (const auto& gradient = self.grad_fn()) {
// …
} else {
return Edge(grad_accumulator(self), 0);
}
The function body is defined in torch/csrc/autograd/functions/accumulate_grad.cpp and it essentially accumulates the input grads in the object’s .grad attribute.
auto AccumulateGrad::apply(variable_list&& grads) -> variable_list {
check_input_variables("AccumulateGrad", grads, 1, 0);
…
at::Tensor new_grad = callHooks(variable, std::move(grads[0]));
std::lock_guard<std::mutex> lock(mutex_);
at::Tensor& grad = variable.mutable_grad();
accumulateGrad(
variable,
grad,
new_grad,
1 + !post_hooks().empty() /* num_expected_refs */,
[&grad](at::Tensor&& grad_update) { grad = std::move(grad_update); });
return variable_list();
}
}} // namespace torch::autograd
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
}} // namespace torch::autograd
```
accumulateGrad
does several checks on the tensors format and eventually performs the variable_grad += new_grad; accumulation.
Preparing the graph for execution
Now, let’s walk through Engine::execute. The first thing to do besides arguments consistency checks is to create the actual GraphTask object we described above. This object keeps all the metadata of the graph execution.
```c++
auto Engine::execute(const edge_list& roots,
const variable_list& inputs,
bool keep_graph,
bool create_graph,
bool accumulate_grad,
const edge_list& outputs) -> variable_list { | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
validate_outputs(roots, const_cast(inputs), {
return msg;
});
// Checks
auto graph_task = std::make_shared(
/ keep_graph / keep_graph,
/ create_graph / create_graph,
/ depth / not_reentrant_backward_call ? 0 : total_depth + 1,
/ cpu_ready_queue / local_ready_queue);
// If we receive a single root, skip creating extra root node
// …
// Prepare graph by computing dependencies
// …
// Queue the root
// …
// launch execution
// …
}
After creating the `GraphTask`, we use its associated function if we only have one root node. If we have multiple root nodes, we create a special `GraphRoot` object as described before.
```c++
bool skip_dummy_node = roots.size() == 1;
auto graph_root = skip_dummy_node ?
roots.at(0).function :
std::make_shared<GraphRoot>(roots, inputs);
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
The next step is to fill the `dependencies_` map in the `GraphTask` object since the engine must know when it can execute a task. The `outputs` here is the `inputs` argument passed to the `torch.autograd.backward()` call in Python. But here, we have reversed the names since the gradients w.r.t. the inputs of the forward pass are now the outputs of the backward pass. And from now on, there is no concept of forward/backward, but only graph traversal and execution.
```c++
auto min_topo_nr = compute_min_topological_nr(outputs);
// Now compute the dependencies for all executable functions
compute_dependencies(graph_root.get(), *graph_task, min_topo_nr);
if (!outputs.empty()) {
graph_task->init_to_execute(*graph_root, outputs, accumulate_grad, min_topo_nr);
}
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
```
Here we preprocess the graph for the execution of the nodes. First, compute_min_topological_nr is called to to obtain the minimum topological number of the tensors specified in outputs (0 if no inputs kwarg was supplied to .backward or input for .grad). This computation prunes paths in the graph that lead to input variables of which we don’t want/need to calculate the grads. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Second, is the compute_dependencies call. This function is a very simple graph traversal that starts with the root Node, and for each of the edges in node.next_edges() it increments the counter in dependencies_. Figure 3 shows the result of the dependencies calculation for the example graph. Note that the number of dependencies of any node is just the number of edges arriving at it.
Figure 3: Number of dependencies for each node
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Finally, the init_to_execute call, this is the one that populates the GraphTask::exec_info_ map in case that inputs were specified in the python backward call. It iterates the graph again, starting from the root, and records in the exec_info_ map the intermediate nodes needed to calculate only the given inputs gradients.
```c++
// Queue the root
if (skip_dummy_node) {
InputBuffer input_buffer(roots.at(0).function->num_inputs());
auto input = inputs.at(0);
input_buffer.add(roots.at(0).input_nr,
std::move(input),
input_stream,
opt_next_stream);
execute_with_graph_task(graph_task, graph_root, std::move(input_buffer));
} else {
execute_with_graph_task(graph_task, graph_root, InputBuffer(variable_list()));
} | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
// Avoid a refcount bump for the Future, since we check for refcount in
// DistEngine (see TORCH_INTERNAL_ASSERT(futureGrads.use_count() == 1)
// in dist_engine.cpp).
auto& fut = graph_task->future_result_;
fut->wait();
return fut->value().toTensorVector();
}
```
And now, we are ready to start the actual execution by creating the InputBuffer. In case we only have one root variable, we begin by copying the value of the inputs tensor (this is the gradients passed to python backward) in position 0 of the input_buffer. This is a small optimization that avoids running the RootNode for no reason. Also, if the rest of the graph is not on the cpu, we directly start on that worker while the RootNode is always placed on the cpu ready queue. Details of the workers and ready queues are explained in the section below.
On the other hand, if we have multiple roots, the GraphRoot object also holds the inputs, so it is enough to pass it an empty InputBuffer. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Graph Traversal and Node Execution
Devices, Threads and Queues
Before diving into the actual execution, we need to see how the engine is structured.
First of all, the engine is multithreaded with one thread per device. For example, the caller thread is associated with the CPU while additional threads are created and associated with each GPU or other devices available in the system. Each thread tracks its device using thread-local storage in the worker_device variable. In addition, the threads have a queue of tasks to be executed also located in thread-local storage, the local_ready_queue. This is where work is queued for this thread to execute in the thread_main function that is explained later. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
You will wonder how the device where a task should be executed is decided. The InputBuffer class has a device() function that returns the first non-cpu device of all its tensors.
This function is used together with Engine::ready_queue to select the queue to queue a task.
auto Engine::ready_queue(std::shared_ptr<ReadyQueue> cpu_ready_queue, at::Device device) -> std::shared_ptr<ReadyQueue>{
if (device.type() == at::kCPU || device.type() == at::DeviceType::Meta) {
return cpu_ready_queue;
} else {
// See Note [Allocating GPUs to autograd threads]
return device_ready_queues_.at(device.index());
}
}
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
}
```
The ReadyQueue object is defined in torch/csrc/autograd/engine.h and it is a simple wrapper over std::priority_queue that allows a thread to wait for a task if it’s empty. One interesting property of the ReadyQueue is that it increases the GraphTask::outstanding_tasks_ value used to determine if the execution has completed or not.
```c++
auto ReadyQueue::push(NodeTask item, bool incrementOutstandingTasks) -> void {
{
std::lock_guard lock(mutex_);
if (incrementOutstandingTasks) {
std::shared_ptr graph_task = item.base_.lock();
++graph_task->outstanding_tasks_;
} | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
++graph_task->outstanding_tasks_;
}
heap_.push(std::move(item));
}
not_empty_.notify_one();
}
auto ReadyQueue::pop() -> NodeTask {
std::unique_lock lock(mutex_);
not_empty_.wait(lock, [this]{ return !heap_.empty(); });
auto task = std::move(const_cast(heap_.top())); heap_.pop();
return task;
}
```
Reentrant Backward
A reentrant backward happens when one of the tasks in a backward pass calls again backward. It is not a very common case, but it can be used to reduce memory utilization as it could potentially avoid saving intermediate results. For more information, check this PyTorch forum post.
```python
class ReentrantBackward(torch.autograd.Function):
@staticmethod
def forward(ctx, input):
return input.sum()
@staticmethod
def backward(ctx, input):
# Let's compute the backward by using autograd
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
input = input.detach().requires_grad_()
with torch.enable_grad():
out = input.sum()
out.backward() # REENTRANT CALL!!
return out.detach()
```
Here, we call backward() inside backward() for a user custom-defined autograd function.
This situation can lead to deadlocks because the first backward needs to wait for the second one to complete. But some internal implementation details can prevent the second backward from completing as it is explained in the dedicated subsection.
Thread Initialization
execute_with_graph_task is in charge of initializing the threads taking care of the computation and placing the root node in the queue of the device that produced it.
```c++
c10::intrusive_ptr Engine::execute_with_graph_task(
const std::shared_ptr& graph_task,
std::shared_ptr graph_root, | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
std::shared_ptr graph_root,
InputBuffer&& input_buffer) {
initialize_device_threads_pool();
// Lock mutex for GraphTask.
std::unique_lock lock(graph_task->mutex_);
auto queue = ready_queue(graph_task->cpu_ready_queue_, input_buffer.device());
if (worker_device == NO_DEVICE) {
set_device(CPU_DEVICE);
graph_task->owner_ = worker_device;
queue->push(NodeTask(graph_task, std::move(graph_root), std::move(input_buffer)));
lock.unlock();
thread_main(graph_task);
worker_device = NO_DEVICE;
} else {
// This deals with reentrant backwards, we will see it later.
}
return graph_task->future_result_;
}
```
First, this function initializes several threads (one per device) calling initialize_device_threads_pool() where several things happen:
One ReadyQueue per device is created.
One thread per non-cpu device is created. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
One thread per non-cpu device is created.
A thread local worker_device variable is set to track the current device associated with the thread.
thread_main function is called, and threads wait for tasks to be put in their queues.
Then it retrieves the queue to place the root node based on the device that holds the tensors present in the input_buffer using the ready_queue function. Now, the main thread (the one also executing the Python interpreter) has its worker_device set to NO_DEVICE, and it is in charge of executing functions with all its tensors living in the cpu. If worker_device is set to any other value, the graph execution is already started, and .backward() was called inside a running Node, creating a reentrant backward call. This is explained later. For now,
the main thread places the task in the queue and call thread_main.
Where the Magic Happens | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Where the Magic Happens
It’s been a long way, but finally, we are ready to traverse the graph and execute the nodes. Each of the spawned threads, and the main thread call thread_main.
```c++
auto Engine::thread_main(const std::shared_ptr& graph_task) -> void {
while (graph_task == nullptr || !graph_task->future_result_->completed()) {
std::shared_ptr local_graph_task;
{
NodeTask task = local_ready_queue->pop();
if (task.isShutdownTask_) {
break;
}
if (!(local_graph_task = task.base_.lock())) {
// GraphTask for function is no longer valid, skipping further
// execution.
continue;
}
if (task.fn_ && !local_graph_task->has_error_.load()) {
at::ThreadLocalStateGuard tls_guard(local_graph_task->thread_locals_);
try {
GraphTaskGuard guard(local_graph_task);
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
GraphTaskGuard guard(local_graph_task);
NodeGuard ndguard(task.fn_);
{
evaluate_function(
local_graph_task,
task.fn_.get(),
task.inputs_,
local_graph_task->cpu_ready_queue_);
}
} catch (std::exception& e) {
thread_on_exception(local_graph_task, task.fn_, e);
}
}
}
// Decrement the outstanding tasks.
--local_graph_task->outstanding_tasks_;
// Check if we've completed execution.
if (local_graph_task->completed()) {
local_graph_task->mark_as_completed_and_run_post_processing();
auto base_owner = local_graph_task->owner_;
if (worker_device != base_owner) {
std::atomic_thread_fence(std::memory_order_release);
ready_queue_by_index(local_graph_task->cpu_ready_queue_, base_owner)
->push(NodeTask(local_graph_task, nullptr, InputBuffer(0)));
}
}
}
}
``` | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
}
}
}
```
The code here is simple, given the local_ready_queue assigned to each thread in thread-local storage. The threads loop until there are no tasks left to execute in the graph. Note that for device-associated threads, the passed graph_task argument is nullptr, and they block in local_ready_queue->pop() until a task is pushed in their queue. After some consistency checks (the task type is shutdown, or the graph is still valid). We get to the actual function invocation in evaluate_function.
```c++
try {
GraphTaskGuard guard(local_graph_task);
NodeGuard ndguard(task.fn_);
{
evaluate_function(
local_graph_task,
task.fn_.get(),
task.inputs_,
local_graph_task->cpu_ready_queue_);
}
} catch (std::exception& e) { | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
} catch (std::exception& e) {
thread_on_exception(local_graph_task, task.fn_, e);
}
}
```
After calling evaluate_function, we check if the graph_task execution is complete by looking the outstanding_tasks_ number. This number increases when a task is pushed to a queue and is decreased in local_graph_task->completed() when a task is executed. When the execution is done, we return the results that are be in the captured_vars_ in case we called torch.autograd.grad() instead of torch.autograd.backward() as this function returns tensors instead of storing them in the .grad attribute of the inputs. Finally we wake up the main thread if it’s waiting by sending a dummy task.
```c++
// Decrement the outstanding tasks.
--local_graph_task->outstanding_tasks_;
// Check if we've completed execution.
if (local_graph_task->completed()) {
local_graph_task->mark_as_completed_and_run_post_processing();
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
auto base_owner = local_graph_task->owner_;
if (worker_device != base_owner) {
std::atomic_thread_fence(std::memory_order_release);
ready_queue_by_index(local_graph_task->cpu_ready_queue_, base_owner)
->push(NodeTask(local_graph_task, nullptr, InputBuffer(0)));
}
}
```
Calling the Function and Unlocking New Tasks
evaluate_function serves three purposes:
Run the function.
Accumulate its results in the next node InputBuffers.
Decrease the dependencies counter of the next nodes and enqueues the tasks reaching 0 to be executed.
```c++
void Engine::evaluate_function(
std::shared_ptr& graph_task,
Node* func,
InputBuffer& inputs,
const std::shared_ptr& cpu_ready_queue) {
// If exec_info_ is not empty, we have to instrument the execution
auto& exec_info_ = graph_task->exec_info_; | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
auto& exec_info_ = graph_task->exec_info_;
if (!exec_info_.empty()) {
// Checks if the function needs to be executed
if (!fn_info.needed_) {
// Skip execution if we don't need to execute the function.
return;
}
}
auto outputs = call_function(graph_task, func, inputs);
auto& fn = *func;
if (!graph_task->keep_graph_) {
fn.release_variables();
}
```
Initially, we check the exec_info_ map of the GraphTask structure to determine if the current node needs to be executed. Remember that if this map is empty, all the nodes are executed because we are calculating the grads for all the inputs of the forward pass.
After this check, the function is executed by running call_function. Its implementation is very straightforward and calls the actual derivative function and registered hooks if any.
```c++
int num_outputs = outputs.size(); | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
int num_outputs = outputs.size();
if (num_outputs == 0) {
// Records leaf stream (if applicable)
return;
}
if (AnomalyMode::is_enabled()) {
// check for nan values in result
}
Next, we check the outputs of the function after call_function is done. If the number of outputs is 0, there are no following nodes to be executed so we can safely return. This is the case of the AccumulateGrad node associated with the leaf nodes.
Also, the check for NaN values in the gradients is done here if requested.
std::lock_guard<std::mutex> lock(graph_task->mutex_);
for (const auto i : c10::irange(num_outputs)) {
auto& output = outputs[i];
const auto& next = fn.next_edge(i);
if (!next.is_valid()) continue;
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
if (!next.is_valid()) continue;
```
We have now executed a grad_fn that has returned one gradient per each of the associated forward pass function inputs. As we saw in the previous blog post, we have an Edge object per each of these input tensors, and the grad_fn of the function producing them in the forward pass. Essentially, Output[0] of the node in the backward pass, corresponds to the first argument of the forward pass associated function. Figure 4 shows how the outputs of a backward function are related to the inputs of the forward function. See that the outputs of grad_fn C are the gradients of z w.r.t. the inputs of Function C
Figure 4: Correspondence between forward and backward functions inputs and outputs
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
We now iterate through these edges and check if the associated functions are ready to be executed.
// Check if the next function is ready to be computed
bool is_ready = false;
auto& dependencies = graph_task->dependencies_;
auto it = dependencies.find(next.function.get());
if (it == dependencies.end()) {
auto name = next.function->name();
throw std::runtime_error(std::string("dependency not found for ") + name);
} else if (--it->second == 0) {
dependencies.erase(it);
is_ready = true;
}
auto& not_ready = graph_task->not_ready_;
auto not_ready_it = not_ready.find(next.function.get());
For this, we check the graph_task->dependencies_ map. We decrement the counter, and if it reaches 0, we mark the function pointed by the edge ready to be executed. Following, we prepare the input buffers of the tasks indicated by the next edges.
```c++
if (not_ready_it == not_ready.end()) {
if (!exec_info_.empty()) { | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
if (!exec_info_.empty()) {
// Skip functions that aren't supposed to be executed
}
// Creates an InputBuffer and moves the output to the corresponding input position
InputBuffer input_buffer(next.function->num_inputs());
input_buffer.add(next.input_nr,
std::move(output),
opt_parent_stream,
opt_next_stream);
if (is_ready) {
auto queue = ready_queue(cpu_ready_queue, input_buffer.device());
queue->push(
NodeTask(graph_task, next.function, std::move(input_buffer)));
} else {
not_ready.emplace(next.function.get(), std::move(input_buffer));
}
``` | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
Here, we look for the task in the `graph_task->not_ready_` map. If it is not present, we create a new `InputBuffer` object and set the current output in the `input_nr` position of the buffer associated with the edge. If the task is ready to be executed, we enqueue it in the appropriate device `ready_queue` and complete the execution. However, if the task is not ready and we have seen it before, it is present in the `not_ready_map_`.
```c++
} else {
// The function already has a buffer
auto &input_buffer = not_ready_it->second;
// Accumulates into buffer
input_buffer.add(next.input_nr,
std::move(output),
opt_parent_stream,
opt_next_stream);
if (is_ready) {
auto queue = ready_queue(cpu_ready_queue, input_buffer.device());
queue->push(NodeTask(graph_task, next.function, std::move(input_buffer)));
not_ready.erase(not_ready_it);
}
}
}
}
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
}
}
}
```
In this case, we accumulate the output in the existing input_buffer instead of creating a new one. Once all the tasks are processed, the worker thread exits the loop and complete.
All this process is summarized in the animation in Figure 5. We see how a thread peeks at the tasks in the ready queue and decrements the next nodes' dependencies, unlocking them for execution.
Figure 5: Animation of the execution of the computational graph
Flow with Reentrant Backward
As we saw above, the reentrant backward problem is when the currently executed function does a nested call to backward. When this happens, the thread running this function goes all the way down to execute_with_graph_task as in the non-reentrant case, but here is when things are different.
```c++
c10::intrusive_ptr Engine::execute_with_graph_task( | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
const std::shared_ptr& graph_task,
std::shared_ptr graph_root,
InputBuffer&& input_buffer) {
initialize_device_threads_pool();
// Lock mutex for GraphTask.
std::unique_lock lock(graph_task->mutex_);
auto queue = ready_queue(graph_task->cpu_ready_queue_, input_buffer.device());
if (worker_device == NO_DEVICE) {
//Regular case
} else {
// If worker_device is any devices (i.e. CPU, CUDA): this is a re-entrant
// backward call from that device.
graph_task->owner_ = worker_device;
// Now that all the non-thread safe fields of the graph_task have been populated,
// we can enqueue it.
queue->push(NodeTask(graph_task, std::move(graph_root), std::move(input_buffer)));
if (current_depth >= max_recursion_depth_) {
// If reached the max depth, switch to a different thread
add_thread_pool_task(graph_task);
} else {
++total_depth;
++current_depth;
lock.unlock();
thread_main(graph_task);
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
thread_main(graph_task);
--current_depth;
--total_depth;
}
}
return graph_task->future_result_;
}
```
Here, execute_with_graph_task detects this as a reentrant call and then looks for the current number of nested calls. If it exceeds the limit, we create a new thread to take care of the execution of this graph, and if not, we execute this reentrant call regularly.
The limit of nested calls was originally set to avoid stack overflow due to reentrant calls creating very large call stacks. However, the number was further reduced when sanitizer tests were added because of the maximum amount of locks a thread can hold at a given moment. This can be seen in torch/csrc/autograd/engine.h. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
When this maximum depth is exceeded, a new thread is created with the add_thread_pool_task function.
void Engine::add_thread_pool_task(const std::weak_ptr<GraphTask>& graph_task) {
std::unique_lock<std::mutex> lck(thread_pool_shared_->mutex_);
// if we have pending graph_task objects to be processed, create a worker.
bool create_thread = (thread_pool_shared_->num_workers_ <= thread_pool_shared_->graphtasks_queue_.size());
thread_pool_shared_->graphtasks_queue_.push(graph_task);
lck.unlock();
if (create_thread) {
std::thread t(&Engine::reentrant_thread_init, this);
t.detach();
}
thread_pool_shared_->work_.notify_one();
}
| https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
```
Before going in-depth, let's look at the thread_pool_shared_ object in the Engine which manages all the information related to the threads associated to the reentrant backward calls.
```c++
struct ThreadPoolShared {
unsigned int num_workers_;
std::condition_variable work_;
std::mutex mutex_;
std::queue> graphtasks_queue_;
// NOLINTNEXTLINE(cppcoreguidelines-pro-type-member-init)
ThreadPoolShared() : num_workers_(0) {}
};
```
ThreadPoolShared is a simple container holding a queue of GraphTask objects with synchronization mechanisms and the number of current workers. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
Now it is easy to understand how add_thread_pool_task creates a thread when there are graph_task objects enqueued and insufficient workers to process them.
add_thread_pool_task initializes a thread by executing reentrant_thread_init
```c++
void Engine::reentrant_thread_init() {
at::init_num_threads();
auto tp_shared = thread_pool_shared_;
while(true) {
std::unique_lock lk(tp_shared->mutex_);
++thread_pool_shared_->num_workers_;
tp_shared->work_.wait(lk, [&tp_shared]{ return !tp_shared->graphtasks_queue_.empty();});
--thread_pool_shared_->num_workers_;
auto task = tp_shared->graphtasks_queue_.front();
tp_shared->graphtasks_queue_.pop();
lk.unlock();
std::shared_ptr graph_task;
if (!(graph_task = task.lock())) {
continue;
}
set_device(graph_task->owner_); | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
}
set_device(graph_task->owner_);
// set the local_ready_queue to the ready queue on the graph_task->owner_ device
local_ready_queue = ready_queue_by_index(graph_task->cpu_ready_queue_, graph_task->owner_);
total_depth = graph_task->reentrant_depth_;
thread_main(graph_task);
}
}
```
The code is straightforward. The newly created thread waits on the thread_pool_shared->graphtasks_queue_ for reentrant backward graphs to be available and executes them. Notice that this thread uses the task-ready queue associated with the device of the thread that started this call by accessing the graph_task->owner_ field set in the execute_with_graph_task function.
Error Handling
Whenever an error happens in one of the worker threads. It will be propagated to the backward calling thread. | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
To achieve this, there is a try/catch block in the thread_main that catches any exception in the Node function call and sets it to the associated GraphTask object.
try {
…
GraphTaskGuard guard(local_graph_task);
NodeGuard ndguard(task.fn_);
{
evaluate_function(
…
}
} catch (std::exception& e) {
thread_on_exception(local_graph_task, task.fn_, e);
}
}
}
thread_on_exception and the functions it calls end up setting the exception in the local_graph_task object.
```c++
void Engine::thread_on_exception( | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
void Engine::thread_on_exception(
std::shared_ptr<GraphTask> graph_task,
const std::shared_ptr<Node>& fn,
std::exception& e) {
graph_task->set_exception(std::current_exception(), fn);
}
void GraphTask::set_exception_without_signal(const std::shared_ptr<Node>& fn) {
if (!has_error_.exchange(true)) {
if (AnomalyMode::is_enabled() && fn) {
fn->metadata()->print_stack(fn->name());
}
}
}
void GraphTask::set_exception(
std::exception_ptr eptr,
const std::shared_ptr<Node>& fn) {
set_exception_without_signal(fn);
if (!future_completed_.exchange(true)) {
// NOLINTNEXTLINE(performance-move-const-arg)
future_result_->setError(std::move(eptr));
}
}
In set_exception it sets the has_error_ flag to true and it calls the setError | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
function of the future_result_ object. This will make the error to be re-thrown at the caller thread when future_result_->value() is accessed.
IValue value() {
std::unique_lock<std::mutex> lock(mutex_);
AT_ASSERT(completed());
if (eptr_) {
std::rethrow_exception(eptr_);
}
return value_;
}
Closing Remarks
This has been the last post of this series covering how PyTorch does the auto differentiation. We hope you enjoyed reading it and that now you are familiar enough with PyTorch internals to start contributing in PyTorch development! | https://pytorch.org/blog/how-computational-graphs-are-executed-in-pytorch/ | pytorch blogs |
layout: blog_detail
title: 'PyTorch 1.6 released w/ Native AMP Support, Microsoft joins as maintainers for Windows'
author: Team PyTorch
Today, we’re announcing the availability of PyTorch 1.6, along with updated domain libraries. We are also excited to announce the team at Microsoft is now maintaining Windows builds and binaries and will also be supporting the community on GitHub as well as the PyTorch Windows discussion forums.
The PyTorch 1.6 release includes a number of new APIs, tools for performance improvement and profiling, as well as major updates to both distributed data parallel (DDP) and remote procedure call (RPC) based distributed training.
A few of the highlights include: | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
A few of the highlights include:
Automatic mixed precision (AMP) training is now natively supported and a stable feature (See here for more details) - thanks for NVIDIA’s contributions;
Native TensorPipe support now added for tensor-aware, point-to-point communication primitives built specifically for machine learning;
Added support for complex tensors to the frontend API surface;
New profiling tools providing tensor-level memory consumption information;
Numerous improvements and new features for both distributed data parallel (DDP) training and the remote procedural call (RPC) packages.
| https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Additionally, from this release onward, features will be classified as Stable, Beta and Prototype. Prototype features are not included as part of the binary distribution and are instead available through either building from source, using nightlies or via compiler flag. You can learn more about what this change means in the post here. You can also find the full release notes here.
Performance & Profiling
[Stable] Automatic Mixed Precision (AMP) Training | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
AMP allows users to easily enable automatic mixed precision training enabling higher performance and memory savings of up to 50% on Tensor Core GPUs. Using the natively supported torch.cuda.amp API, AMP provides convenience methods for mixed precision, where some operations use the torch.float32 (float) datatype and other operations use torch.float16 (half). Some ops, like linear layers and convolutions, are much faster in float16. Other ops, like reductions, often require the dynamic range of float32. Mixed precision tries to match each op to its appropriate datatype.
Design doc (Link)
Documentation (Link)
Usage examples (Link)
[Beta] Fork/Join Parallelism | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
[Beta] Fork/Join Parallelism
This release adds support for a language-level construct as well as runtime support for coarse-grained parallelism in TorchScript code. This support is useful for situations such as running models in an ensemble in parallel, or running bidirectional components of recurrent nets in parallel, and allows the ability to unlock the computational power of parallel architectures (e.g. many-core CPUs) for task level parallelism.
Parallel execution of TorchScript programs is enabled through two primitives: torch.jit.fork and torch.jit.wait. In the below example, we parallelize execution of foo:
```python
import torch
from typing import List
def foo(x):
return torch.neg(x)
@torch.jit.script
def example(x):
futures = [torch.jit.fork(foo, x) for _ in range(100)]
results = [torch.jit.wait(future) for future in futures]
return torch.sum(torch.stack(results))
print(example(torch.ones([])))
```
Documentation (Link)
| https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
[Beta] Memory Profiler
The torch.autograd.profiler API now includes a memory profiler that lets you inspect the tensor memory cost of different operators inside your CPU and GPU models.
Here is an example usage of the API:
```python
import torch
import torchvision.models as models
import torch.autograd.profiler as profiler
model = models.resnet18()
inputs = torch.randn(5, 3, 224, 224)
with profiler.profile(profile_memory=True, record_shapes=True) as prof:
model(inputs)
NOTE: some columns were removed for brevity
print(prof.key_averages().table(sort_by="self_cpu_memory_usage", row_limit=10))
--------------------------- --------------- --------------- ---------------
Name CPU Mem Self CPU Mem Number of Calls
--------------------------- --------------- --------------- ---------------
empty 94.79 Mb 94.79 Mb 123
resize_ 11.48 Mb 11.48 Mb 2 | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
addmm 19.53 Kb 19.53 Kb 1
empty_strided 4 b 4 b 1
conv2d 47.37 Mb 0 b 20
--------------------------- --------------- --------------- ---------------
```
PR (Link)
Documentation (Link)
Distributed Training & RPC
[Beta] TensorPipe backend for RPC | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
[Beta] TensorPipe backend for RPC
PyTorch 1.6 introduces a new backend for the RPC module which leverages the TensorPipe library, a tensor-aware point-to-point communication primitive targeted at machine learning, intended to complement the current primitives for distributed training in PyTorch (Gloo, MPI, ...) which are collective and blocking. The pairwise and asynchronous nature of TensorPipe lends itself to new networking paradigms that go beyond data parallel: client-server approaches (e.g., parameter server for embeddings, actor-learner separation in Impala-style RL, ...) and model and pipeline parallel training (think GPipe), gossip SGD, etc.
# One-line change needed to opt in
torch.distributed.rpc.init_rpc(
...
backend=torch.distributed.rpc.BackendType.TENSORPIPE,
)
# No changes to the rest of the RPC API
torch.distributed.rpc.rpc_sync(...)
Design doc (Link)
| https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Documentation (Link)
[Beta] DDP+RPC
PyTorch Distributed supports two powerful paradigms: DDP for full sync data parallel training of models and the RPC framework which allows for distributed model parallelism. Previously, these two features worked independently and users couldn’t mix and match these to try out hybrid parallelism paradigms.
Starting in PyTorch 1.6, we’ve enabled DDP and RPC to work together seamlessly so that users can combine these two techniques to achieve both data parallelism and model parallelism. An example is where users would like to place large embedding tables on parameter servers and use the RPC framework for embedding lookups, but store smaller dense parameters on trainers and use DDP to synchronize the dense parameters. Below is a simple code snippet.
```python
// On each trainer
remote_emb = create_emb(on="ps", ...)
ddp_model = DDP(dense_model)
for data in batch:
with torch.distributed.autograd.context(): | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
with torch.distributed.autograd.context():
res = remote_emb(data)
loss = ddp_model(res)
torch.distributed.autograd.backward([loss])
```
DDP+RPC Tutorial (Link)
Documentation (Link)
Usage Examples (Link)
[Beta] RPC - Asynchronous User Functions | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
[Beta] RPC - Asynchronous User Functions
RPC Asynchronous User Functions supports the ability to yield and resume on the server side when executing a user-defined function. Prior to this feature, when a callee processes a request, one RPC thread waits until the user function returns. If the user function contains IO (e.g., nested RPC) or signaling (e.g., waiting for another request to unblock), the corresponding RPC thread would sit idle waiting for these events. As a result, some applications have to use a very large number of threads and send additional RPC requests, which can potentially lead to performance degradation. To make a user function yield on such events, applications need to: 1) Decorate the function with the @rpc.functions.async_execution decorator; and 2) Let the function return a torch.futures.Future and install the resume logic as callbacks on the Future object. See below for an example:
```python
@rpc.functions.async_execution
def async_add_chained(to, x, y, z): | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
def async_add_chained(to, x, y, z):
return rpc.rpc_async(to, torch.add, args=(x, y)).then(
lambda fut: fut.wait() + z
)
ret = rpc.rpc_sync(
"worker1",
async_add_chained,
args=("worker2", torch.ones(2), 1, 1)
)
print(ret) # prints tensor([3., 3.])
```
Tutorial for performant batch RPC using Asynchronous User Functions (Link)
Documentation (Link)
Usage examples (Link)
Frontend API Updates
[Beta] Complex Numbers | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
[Beta] Complex Numbers
The PyTorch 1.6 release brings beta level support for complex tensors including torch.complex64 and torch.complex128 dtypes. A complex number is a number that can be expressed in the form a + bj, where a and b are real numbers, and j is a solution of the equation x^2 = −1. Complex numbers frequently occur in mathematics and engineering, especially in signal processing and the area of complex neural networks is an active area of research. The beta release of complex tensors will support common PyTorch and complex tensor functionality, plus functions needed by Torchaudio, ESPnet and others. While this is an early version of this feature, and we expect it to improve over time, the overall goal is provide a NumPy compatible user experience that leverages PyTorch’s ability to run on accelerators and work with autograd to better support the scientific community.
Mobile Updates | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Mobile Updates
PyTorch 1.6 brings increased performance and general stability for mobile on-device inference. We squashed a few bugs, continued maintenance and added few new features while improving fp32 and int8 performance on a large variety of ML model inference on CPU backend.
[Beta] Mobile Features and Performance
Stateless and stateful XNNPACK Conv and Linear operators
Stateless MaxPool2d + JIT optimization passes
JIT pass optimizations: Conv + BatchNorm fusion, graph rewrite to replace conv2d/linear with xnnpack ops, relu/hardtanh fusion, dropout removal
QNNPACK integration removes requantization scale constraint
Per-channel quantization for conv, linear and dynamic linear
Disable tracing for mobile client to save ~600 KB on full-jit builds
Updated Domain Libraries
torchvision 0.7 | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Updated Domain Libraries
torchvision 0.7
torchvision 0.7 introduces two new pretrained semantic segmentation models, FCN ResNet50 and DeepLabV3 ResNet50, both trained on COCO and using smaller memory footprints than the ResNet101 backbone. We also introduced support for AMP (Automatic Mixed Precision) autocasting for torchvision models and operators, which automatically selects the floating point precision for different GPU operations to improve performance while maintaining accuracy.
Release notes (Link)
torchaudio 0.6
torchaudio now officially supports Windows. This release also introduces a new model module (with wav2letter included), new functionals (contrast, cvm, dcshift, overdrive, vad, phaser, flanger, biquad), datasets (GTZAN, CMU), and a new optional sox backend with support for TorchScript.
Release notes (Link)
| https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Additional updates
HACKATHON
The Global PyTorch Summer Hackathon is back! This year, teams can compete in three categories virtually:
PyTorch Developer Tools: Tools or libraries designed to improve productivity and efficiency of PyTorch for researchers and developers
Web/Mobile Applications powered by PyTorch: Applications with web/mobile interfaces and/or embedded devices powered by PyTorch
PyTorch Responsible AI Development Tools: Tools, libraries, or web/mobile apps for responsible AI development
This is a great opportunity to connect with the community and practice your machine learning skills.
Join the hackathon
Watch educational videos
LPCV Challenge | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
LPCV Challenge
The 2020 CVPR Low-Power Vision Challenge (LPCV) - Online Track for UAV video submission deadline is coming up shortly. You have until July 31, 2020 to build a system that can discover and recognize characters in video captured by an unmanned aerial vehicle (UAV) accurately using PyTorch and Raspberry Pi 3B+.
Prototype Features
To reiterate, Prototype features in PyTorch are early features that we are looking to gather feedback on, gauge the usefulness of and improve ahead of graduating them to Beta or Stable. The following features are not part of the PyTorch 1.6 release and instead are available in nightlies with separate docs/tutorials to help facilitate early usage and feedback.
Distributed RPC/Profiler | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Distributed RPC/Profiler
Allow users to profile training jobs that use torch.distributed.rpc using the autograd profiler, and remotely invoke the profiler in order to collect profiling information across different nodes. The RFC can be found here and a short recipe on how to use this feature can be found here.
TorchScript Module Freezing
Module Freezing is the process of inlining module parameters and attributes values into the TorchScript internal representation. Parameter and attribute values are treated as final value and they cannot be modified in the frozen module. The PR for this feature can be found here and a short tutorial on how to use this feature can be found here.
Graph Mode Quantization | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Graph Mode Quantization
Eager mode quantization requires users to make changes to their model, including explicitly quantizing activations, module fusion, rewriting use of torch ops with Functional Modules and quantization of functionals are not supported. If we can trace or script the model, then the quantization can be done automatically with graph mode quantization without any of the complexities in eager mode, and it is configurable through a qconfig_dict. A tutorial on how to use this feature can be found here.
Quantization Numerical Suite | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
Quantization Numerical Suite
Quantization is good when it works, but it’s difficult to know what's wrong when it doesn't satisfy the expected accuracy. A prototype is now available for a Numerical Suite that measures comparison statistics between quantized modules and float modules. This is available to test using eager mode and on CPU only with more support coming. A tutorial on how to use this feature can be found here.
Cheers!
Team PyTorch | https://pytorch.org/blog/pytorch-1.6-released/ | pytorch blogs |
layout: blog_detail
title: "Accelerating Large Language Models with Accelerated Transformers"
author: Lucas Pasqualin, Driss Guessous, Christian Puhrsch, Bertrand Maher, Michael Gschwind
| https://pytorch.org/blog/accelerating-large-language-models/ | pytorch blogs |
TL;DR. We show how to use Accelerated PyTorch 2.0 Transformers and the newly introduced torch.compile() method to accelerate Large Language Models on the example of nanoGPT, a compact open-source implementation of the GPT model from Andrej Karpathy. Using the new scaled dot product attention operator introduced with Accelerated PT2 Transformers, we select the flash_attention custom kernel and achieve faster training time per batch (measured with Nvidia A100 GPUs), going from a ~143ms/batch baseline to ~113 ms/batch. In addition, the enhanced implementation using the SDPA operator offers better numerical stability. Finally, further optimizations are achieved using padded inputs, which when combined with flash attention lead to ~87ms/batch. | https://pytorch.org/blog/accelerating-large-language-models/ | pytorch blogs |
Recent times have seen exponential adoption of large language models (LLMs) and Generative AI in everyday life. Tightly coupled with these ever-growing models is the ever-growing training cost - in terms of both time and hardware utilization. The PyTorch team has tackled these challenges head on with Accelerated PyTorch 2 Transformers (previously known as “Better Transformer”) and JIT Compilation in PyTorch 2.0. | https://pytorch.org/blog/accelerating-large-language-models/ | pytorch blogs |
In this blog post, we explore training optimizations gained by utilizing custom kernel implementations of SDPA - also known as scaled dot product attention - a critical layer in transformer models. The custom kernel for SDPA replaces several discrete sequential operations with one globally optimized kernel which avoids allocating a large amount of intermediate CUDA memory. This approach offers a number of advantages, including but not limited to: higher performance computation of SDPA by reducing memory bandwidth bottleneck, reduced memory footprint to support larger batch sizes, and finally added numerical stability by prescaling input tensors. These optimizations are demonstrated on nanoGPT, an open-source implementation of GPT from Andrej Karpathy.
Background
Scaled dot product attention is the fundamental building block of multihead attention, as introduced in “Attention is All You Need”, and has a wide range of applications in LLM and Generative AI models. | https://pytorch.org/blog/accelerating-large-language-models/ | pytorch blogs |
Figure 1: The Transformer model architecture based on “Attention is All You Need”. With the new PyTorch SDPA operator, Multi-Head Attention is efficiently implemented by a linear layer for the in-projection, the SDPA operator, and a linear layer for the out-projection.
With the new scaled_dot_product_attention operator, multihead attention can be implemented in just 3 steps: in projection with a linear layer, SDPA, and out projection with a linear layer.
```
In Projection
variable descriptions:
q,k,v = Query, Key, Value tensors
bsz = batch size
num_heads = Numner of heads for Multihead Attention
tgt_len = Target length
src_len = Source Length
head_dim: Head Dimension | https://pytorch.org/blog/accelerating-large-language-models/ | pytorch blogs |
head_dim: Head Dimension
q, k, v = _in_projection(query, key, value, q_proj_weight, k_proj_weight, v_proj_weight, b_q, b_k, b_v)
q = q.view(bsz, num_heads, tgt_len, head_dim)
k = k.view(bsz, num_heads, src_len, head_dim)
v = v.view(bsz, num_heads, src_len, head_dim)
# Scaled Dot Product Attention
attn_output = scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
# Out Projection
attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
attn_output = linear(attn_output, out_proj_weight, out_proj_bias)
attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
``` | https://pytorch.org/blog/accelerating-large-language-models/ | pytorch blogs |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.