Features

Device-Agnostics

Ease-of-use Python API

Intel® Extension for PyTorch* provides simple frontend Python APIs and utilities to get performance optimizations such as operator optimization.

Check the API Documentation for details of API functions and Examples for helpful usage tips.

Here are detailed discussions of specific feature topics, summarized in the rest of this document:

Channels Last

Compared with the default NCHW memory format, using channels_last (NHWC) memory format can further accelerate convolutional neural networks. In Intel® Extension for PyTorch*, NHWC memory format has been enabled for most key CPU and GPU operators. More detailed information is available at Channels Last.

Intel® Extension for PyTorch* automatically converts a model to channels last memory format when users optimize the model with ipex.optimize(model). With this feature users won’t need to manually apply model=model.to(memory_format=torch.channels_last) any more. More detailed information is available at Auto Channels Last.

Auto Mixed Precision (AMP)

Benefiting from less memory usage and computation, low precision data types typically speed up both training and inference workloads. Furthermore, accelerated by Intel® native hardware instructions, including Intel® Deep Learning Boost (Intel® DL Boost) on the 3rd Generation Xeon® Scalable Processors (aka Cooper Lake), as well as the Intel® Advanced Matrix Extensions (Intel® AMX) instruction set on the 4th next generation of Intel® Xeon® Scalable Processors (aka Sapphire Rapids), low precision data type, bfloat 16 and float16, provide further boosted performance. We recommend to use AMP for accelerating convolutional and matmul based neural networks.

The support of Auto Mixed Precision (AMP) with BFloat16 on CPU and BFloat16 optimization of operators has been enabled in Intel® Extension for PyTorch*, and partially upstreamed to PyTorch master branch. These optimizations will be landed in PyTorch master through PRs that are being submitted and reviewed. On GPU side, support of BFloat16 and Float16 are both available in Intel® Extension for PyTorch*. BFloat16 is the default low precision floating data type when AMP is enabled.

Detailed information of AMP for GPU and CPU are available at Auto Mixed Precision (AMP) on GPU and Auto Mixed Precision (AMP) on CPU respectively.

INT8 Quantization

Intel® Extension for PyTorch* provides built-in quantization recipes to deliver good statistical accuracy for most popular DL workloads including CNN, NLP and recommendation models on CPU side. On top of that, if users would like to tune for a higher accuracy than what the default recipe provides, a recipe tuning API powered by Intel® Neural Compressor is provided for users to try.

Check more detailed information for INT8 Quantization [CPU] and INT8 recipe tuning API guide (Experimental, *NEW feature in 1.13.0* on CPU) on CPU side.

On Intel® GPUs, quantization usages follow PyTorch default quantization APIs. Check sample codes at Examples page.

Distributed Training

To meet demands of large scale model training over multiple devices, distributed training on Intel® GPUs and CPUs are supported. Two alternative methodologies are available. Users can choose either to use PyTorch native distributed training module, Distributed Data Parallel (DDP), with Intel® oneAPI Collective Communications Library (oneCCL) support via Intel® oneCCL Bindings for PyTorch (formerly known as torch_ccl) or use Horovod with Intel® oneAPI Collective Communications Library (oneCCL) support (Experimental).

For more detailed information, check DDP and Horovod (Experimental).

GPU-Specific

DLPack Solution

DLPack defines a stable in-memory data structure for sharing tensors among frameworks. It allows to share tensor data without copy when interoparate with other libraries. Intel® Extension for PyTorch* extends DLPack support in PyTorch* for XPU device particularly.

For more detailed information, check DLPack Solution.

DPC++ Extension

Intel® Extension for PyTorch* provides C++ APIs to get SYCL queue and configure floating-point math mode.

Check the API Documentation for the details of API functions. DPC++ Extension describes how to write customized DPC++ kernels with a practical example and build it with setuptools and CMake.

Advanced Configuration

The default settings for Intel® Extension for PyTorch* are sufficient for most use cases. However, if users want to customize Intel® Extension for PyTorch*, advanced configuration is available at build time and runtime.

For more detailed information, check Advanced Configuration.

Legacy Profiler Tool (Experimental)

The legacy profiler tool is an extension of PyTorch* legacy profiler for profiling operators’ overhead on XPU devices. With this tool, users can get the information in many fields of the run models or code scripts. User should build Intel® Extension for PyTorch* with profiler support as default and enable this tool by a with statement before the code segment.

For more detailed information, check Legacy Profiler Tool.

Simple Trace Tool (Experimental)

Simple Trace is a built-in debugging tool that lets you control printing out the call stack for a piece of code. Once enabled, it can automatically print out verbose messages of called operators in a stack format with indenting to distinguish the context.

For more detailed information, check Simple Trace Tool.

CPU-Specific

Operator Optimization

Intel® Extension for PyTorch* also optimizes operators and implements several customized operators for performance boosts. A few ATen operators are replaced by their optimized counterparts in Intel® Extension for PyTorch* via the ATen registration mechanism. Some customized operators are implemented for several popular topologies. For instance, ROIAlign and NMS are defined in Mask R-CNN. To improve performance of these topologies, Intel® Extension for PyTorch* also optimized these customized operators.

class intel_extension_for_pytorch.nn.FrozenBatchNorm2d(num_features: int, eps: float = 1e-05)

BatchNorm2d where the batch statistics and the affine parameters are fixed

Parameters: num_features (int) – \(C\) from an expected input of size \((N, C, H, W)\)

Shape

Input: \((N, C, H, W)\)
Output: \((N, C, H, W)\) (same shape as input)

intel_extension_for_pytorch.nn.functional.interaction(*args)

Get the interaction feature beyond different kinds of features (like gender or hobbies), used in DLRM model.

For now, we only optimized “dot” interaction at DLRM Github repo. Through this, we use the dot product to represent the interaction feature between two features.

For example, if feature 1 is “Man” which is represented by [0.1, 0.2, 0.3], and feature 2 is “Like play football” which is represented by [-0.1, 0.3, 0.2].

The dot interaction feature is ([0.1, 0.2, 0.3] * [-0.1, 0.3, 0.2]^T) = -0.1 + 0.6 + 0.6 = 1.1

Parameters: *args – Multiple tensors which represent different features

Shape

Input: \(N * (B, D)\), where N is the number of different kinds of features, B is the batch size, D is feature size
Output: \((B, D + N * ( N - 1 ) / 2)\)

Auto kernel selection is a feature that enables users to tune for better performance with GEMM operations. It is provided as parameter –auto_kernel_selection, with boolean value, of the ipex.optimize() function. By default, the GEMM kernel is computed with oneMKL primitives. However, under certain circumstances oneDNN primitives run faster. Users are able to set –auto_kernel_selection to True to run GEMM kernels with oneDNN primitives.” -> “We aim to provide good default performance by leveraging the best of math libraries and enabled weights_prepack, and it has been verified with broad set of models. If you would like to try other alternatives, you can use auto_kernel_selection toggle in ipex.optimize to switch, and you can disable weights_preack in ipex.optimize if you are concerning the memory footprint more than performance gain. However in majority cases, keeping default is what we recommend.

Runtime Extension

Intel® Extension for PyTorch* Runtime Extension provides PyTorch frontend APIs for users to get finer-grained control of the thread runtime and provides:

Multi-stream inference via the Python frontend module MultiStreamModule.
Spawn asynchronous tasks from both Python and C++ frontend.
Program core bindings for OpenMP threads from both Python and C++ frontend.

Note

Intel® Extension for PyTorch* Runtime extension is still in the experimental stage. The API is subject to change. More detailed descriptions are available in the API Documentation.

For more detailed information, check Runtime Extension.

Codeless Optimization (Experimental, NEW feature in 1.13.*)

This feature enables users to get performance benefits from Intel® Extension for PyTorch* without changing Python scripts. It hopefully eases the usage and has been verified working well with broad scope of models, though in few cases there could be small overhead comparing to applying optimizations with Intel® Extension for PyTorch* APIs.

For more detailed information, check Codeless Optimization.

Graph Capture (Experimental, NEW feature in 1.13.0*)

Since graph mode is key for deployment performance, this feature automatically captures graphs based on set of technologies that PyTorch supports, such as TorchScript and TorchDynamo. Users won’t need to learn and try different PyTorch APIs to capture graphs, instead, they can turn on a new boolean flag –graph_mode (default off) in ipex.optimize to get the best of graph optimization.

For more detailed information, check Graph Capture.

HyperTune (Experimental, NEW feature in 1.13.0*)

HyperTune is an experimental feature to perform hyperparameter/execution configuration searching. The searching is used in various areas such as optimization of hyperparameters of deep learning models. The searching is extremely useful in real situations when the number of hyperparameters, including configuration of script execution, and their search spaces are huge that manually tuning these hyperparameters/configuration is impractical and time consuming. Hypertune automates this process of execution configuration searching for the launcher and Intel® Extension for PyTorch*.

For more detailed information, check HyperTune.