Table of Contents
Shortcuts
recipes/intel_extension_for_pytorch

Intel® 用于PyTorch的扩展

Created On: Nov 09, 2021 | Last Updated: Jul 25, 2024 | Last Verified: Nov 05, 2024

Intel® 用于PyTorch的扩展通过提供最新的性能优化功能，进一步提升了在Intel硬件上的性能。这些优化利用了AVX-512向量神经网络指令（AVX512 VNNI）和Intel®高级矩阵扩展（Intel® AMX）以及Intel离散GPU上的Intel Xe矩阵扩展（XMX）AI引擎。此外，通过PyTorch xpu 设备，Intel® 用于PyTorch的扩展为Intel离散GPU的PyTorch加速提供了便捷。

Intel® 用于PyTorch的扩展已作为开源项目发布在 Github。

功能

Intel® 用于PyTorch的扩展在CPU和GPU之间共享大部分功能。

  • 易用的Python API： Intel® 用于PyTorch的扩展通过简单的前端Python API和实用工具，为用户提供了通过少量代码更改获得性能优化（如图优化和操作符优化）的可能性。通常，仅需在原始代码中添加2至3行代码。

  • Channels Last： 与默认的NCHW内存格式相比，channels_last（NHWC）内存格式可以进一步加速卷积神经网络。在Intel® 用于PyTorch的扩展中，大多数CPU关键操作符已经启用了NHWC内存格式，尽管其中的一部分尚未合并到PyTorch主分支中。预计这些功能很快会完全合并到PyTorch上游。

  • 自动混合精度（AMP）： 低精度数据类型BFloat16已经在第3代Xeon可扩展服务器（又称Cooper Lake）中原生支持，并将被支持Intel® Xeon®可扩展处理器的新一代中（配备Intel®高级矩阵扩展（Intel® AMX）指令集），其性能将进一步提高。在Intel® 用于PyTorch的扩展中，已经广泛支持了用BFloat16进行的自动混合精度（AMP）以及操作符优化。这些功能的一部分已被部分上游合并到PyTorch主分支中。

  • 图优化： 为了通过torchscript进一步优化性能，Intel® 用于PyTorch的扩展支持常用的操作符模式的融合，例如Conv2D+ReLU、Linear+ReLU等。这些优化以透明的方式传递给用户。支持的详细融合模式可在 这里 找到。

  • 操作符优化： Intel® 用于PyTorch的扩展优化了操作符，并为性能实现了一些定制化操作符。例如，在Mask R-CNN模型中定义的ROIAlign和NMS。为提升这些模型的性能，Intel® 用于PyTorch的扩展还优化了这些定制操作符。

快速上手

用户需要对代码进行少量更改即可开始使用Intel® 用于PyTorch的扩展。支持PyTorch命令式模式和TorchScript模式。本节介绍了如何在命令式模式和TorchScript模式下使用Intel® 用于PyTorch的扩展API功能。

您只需要导入Intel® 用于PyTorch的扩展包并将其优化功能应用到模型对象上。如果是训练工作量，还需要对优化器对象应用优化功能。

对于使用BFloat16数据类型进行训练和推理，PyTorch上游已启用`torch.cpu.amp`，支持便捷的混合精度。此外，Intel® 用于PyTorch的扩展注册的`torch.xpu.amp`，允许在Intel离散GPU上轻松使用BFloat16和Float16数据类型。

示例 - CPU

本节展示了在CPU上使用Intel® 用于PyTorch的扩展进行训练和推理的示例。

使用Intel® 用于PyTorch的扩展所需的代码更改已被突出显示。

训练

Float32

import torch
import torchvision
import intel_extension_for_pytorch as ipex

LR = 0.001
DOWNLOAD = True
DATA = 'datasets/cifar10/'

transform = torchvision.transforms.Compose([
    torchvision.transforms.Resize((224, 224)),
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
train_dataset = torchvision.datasets.CIFAR10(
        root=DATA,
        train=True,
        transform=transform,
        download=DOWNLOAD,
)
train_loader = torch.utils.data.DataLoader(
        dataset=train_dataset,
        batch_size=128
)

model = torchvision.models.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr = LR, momentum=0.9)
model.train()
model, optimizer = ipex.optimize(model, optimizer=optimizer)

for batch_idx, (data, target) in enumerate(train_loader):
    optimizer.zero_grad()
    output = model(data)
    loss = criterion(output, target)
    loss.backward()
    optimizer.step()
    print(batch_idx)
torch.save({
     'model_state_dict': model.state_dict(),
     'optimizer_state_dict': optimizer.state_dict(),
     }, 'checkpoint.pth')

BFloat16

import torch
import torchvision
import intel_extension_for_pytorch as ipex

LR = 0.001
DOWNLOAD = True
DATA = 'datasets/cifar10/'

transform = torchvision.transforms.Compose([
    torchvision.transforms.Resize((224, 224)),
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
train_dataset = torchvision.datasets.CIFAR10(
        root=DATA,
        train=True,
        transform=transform,
        download=DOWNLOAD,
)
train_loader = torch.utils.data.DataLoader(
        dataset=train_dataset,
        batch_size=128
)

model = torchvision.models.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr = LR, momentum=0.9)
model.train()
model, optimizer = ipex.optimize(model, optimizer=optimizer, dtype=torch.bfloat16)

for batch_idx, (data, target) in enumerate(train_loader):
    optimizer.zero_grad()
    with torch.cpu.amp.autocast():
        output = model(data)
        loss = criterion(output, target)
        loss.backward()
    optimizer.step()
    print(batch_idx)
torch.save({
     'model_state_dict': model.state_dict(),
     'optimizer_state_dict': optimizer.state_dict(),
     }, 'checkpoint.pth')

推理 - 命令式模式

Float32

import torch
import torchvision.models as models

model = models.resnet50(pretrained=True)
model.eval()
data = torch.rand(1, 3, 224, 224)

#################### code changes ####################
import intel_extension_for_pytorch as ipex
model = ipex.optimize(model)
######################################################

with torch.no_grad():
  model(data)

BFloat16

import torch
from transformers import BertModel

model = BertModel.from_pretrained(args.model_name)
model.eval()

vocab_size = model.config.vocab_size
batch_size = 1
seq_length = 512
data = torch.randint(vocab_size, size=[batch_size, seq_length])

#################### code changes ####################
import intel_extension_for_pytorch as ipex
model = ipex.optimize(model, dtype=torch.bfloat16)
######################################################

with torch.no_grad():
  with torch.cpu.amp.autocast():
    model(data)

推理 - TorchScript模式

TorchScript模式使图优化成为可能，因此可以提升一些模型的性能。Intel® 用于PyTorch的扩展启用了大多数常用的操作符模式融合，用户无需额外代码更改即可获得性能收益。

Float32

import torch
import torchvision.models as models

model = models.resnet50(pretrained=True)
model.eval()
data = torch.rand(1, 3, 224, 224)

#################### code changes ####################
import intel_extension_for_pytorch as ipex
model = ipex.optimize(model)
######################################################

with torch.no_grad():
  d = torch.rand(1, 3, 224, 224)
  model = torch.jit.trace(model, d)
  model = torch.jit.freeze(model)

  model(data)

BFloat16

import torch
from transformers import BertModel

model = BertModel.from_pretrained(args.model_name)
model.eval()

vocab_size = model.config.vocab_size
batch_size = 1
seq_length = 512
data = torch.randint(vocab_size, size=[batch_size, seq_length])

#################### code changes ####################
import intel_extension_for_pytorch as ipex
model = ipex.optimize(model, dtype=torch.bfloat16)
######################################################

with torch.no_grad():
  with torch.cpu.amp.autocast():
    d = torch.randint(vocab_size, size=[batch_size, seq_length])
    model = torch.jit.trace(model, (d,), check_trace=False, strict=False)
    model = torch.jit.freeze(model)

    model(data)

示例 - GPU

本节展示了在GPU上使用Intel® 用于PyTorch的扩展进行训练和推理的示例。

需要更改的代码在上一行用注释突出显示。

训练

Float32

import torch
import torchvision
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

LR = 0.001
DOWNLOAD = True
DATA = 'datasets/cifar10/'

transform = torchvision.transforms.Compose([
    torchvision.transforms.Resize((224, 224)),
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
train_dataset = torchvision.datasets.CIFAR10(
        root=DATA,
        train=True,
        transform=transform,
        download=DOWNLOAD,
)
train_loader = torch.utils.data.DataLoader(
        dataset=train_dataset,
        batch_size=128
)

model = torchvision.models.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr = LR, momentum=0.9)
model.train()
#################################### code changes ################################
model = model.to("xpu")
model, optimizer = ipex.optimize(model, optimizer=optimizer, dtype=torch.float32)
#################################### code changes ################################

for batch_idx, (data, target) in enumerate(train_loader):
    ########## code changes ##########
    data = data.to("xpu")
    target = target.to("xpu")
    ########## code changes ##########
    optimizer.zero_grad()
    output = model(data)
    loss = criterion(output, target)
    loss.backward()
    optimizer.step()
    print(batch_idx)
torch.save({
     'model_state_dict': model.state_dict(),
     'optimizer_state_dict': optimizer.state_dict(),
     }, 'checkpoint.pth')

BFloat16

import torch
import torchvision
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

LR = 0.001
DOWNLOAD = True
DATA = 'datasets/cifar10/'

transform = torchvision.transforms.Compose([
    torchvision.transforms.Resize((224, 224)),
    torchvision.transforms.ToTensor(),
    torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
train_dataset = torchvision.datasets.CIFAR10(
        root=DATA,
        train=True,
        transform=transform,
        download=DOWNLOAD,
)
train_loader = torch.utils.data.DataLoader(
        dataset=train_dataset,
        batch_size=128
)

model = torchvision.models.resnet50()
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr = LR, momentum=0.9)
model.train()
##################################### code changes ################################
model = model.to("xpu")
model, optimizer = ipex.optimize(model, optimizer=optimizer, dtype=torch.bfloat16)
##################################### code changes ################################

for batch_idx, (data, target) in enumerate(train_loader):
    optimizer.zero_grad()
    ######################### code changes #########################
    data = data.to("xpu")
    target = target.to("xpu")
    with torch.xpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
    ######################### code changes #########################
        output = model(data)
        loss = criterion(output, target)
    loss.backward()
    optimizer.step()
    print(batch_idx)
torch.save({
     'model_state_dict': model.state_dict(),
     'optimizer_state_dict': optimizer.state_dict(),
     }, 'checkpoint.pth')

推理 - 命令式模式

Float32

import torch
import torchvision.models as models
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

model = models.resnet50(pretrained=True)
model.eval()
data = torch.rand(1, 3, 224, 224)

model = model.to(memory_format=torch.channels_last)
data = data.to(memory_format=torch.channels_last)

#################### code changes ################
model = model.to("xpu")
data = data.to("xpu")
model = ipex.optimize(model, dtype=torch.float32)
#################### code changes ################

with torch.no_grad():
  model(data)

BFloat16

import torch
import torchvision.models as models
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

model = models.resnet50(pretrained=True)
model.eval()
data = torch.rand(1, 3, 224, 224)

model = model.to(memory_format=torch.channels_last)
data = data.to(memory_format=torch.channels_last)

#################### code changes #################
model = model.to("xpu")
data = data.to("xpu")
model = ipex.optimize(model, dtype=torch.bfloat16)
#################### code changes #################

with torch.no_grad():
  ################################# code changes ######################################
  with torch.xpu.amp.autocast(enabled=True, dtype=torch.bfloat16, cache_enabled=False):
  ################################# code changes ######################################
    model(data)

Float16

import torch
import torchvision.models as models
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

model = models.resnet50(pretrained=True)
model.eval()
data = torch.rand(1, 3, 224, 224)

model = model.to(memory_format=torch.channels_last)
data = data.to(memory_format=torch.channels_last)

#################### code changes ################
model = model.to("xpu")
data = data.to("xpu")
model = ipex.optimize(model, dtype=torch.float16)
#################### code changes ################

with torch.no_grad():
  ################################# code changes ######################################
  with torch.xpu.amp.autocast(enabled=True, dtype=torch.float16, cache_enabled=False):
  ################################# code changes ######################################
    model(data)

推理 - TorchScript模式

TorchScript模式使图优化成为可能，因此可以提升一些模型的性能。Intel® 用于PyTorch的扩展启用了大多数常用的操作符模式融合，用户无需额外代码更改即可获得性能收益。

Float32

import torch
from transformers import BertModel
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

model = BertModel.from_pretrained(args.model_name)
model.eval()

vocab_size = model.config.vocab_size
batch_size = 1
seq_length = 512
data = torch.randint(vocab_size, size=[batch_size, seq_length])

#################### code changes ################
model = model.to("xpu")
data = data.to("xpu")
model = ipex.optimize(model, dtype=torch.float32)
#################### code changes ################

with torch.no_grad():
  d = torch.randint(vocab_size, size=[batch_size, seq_length])
  ##### code changes #####
  d = d.to("xpu")
  ##### code changes #####
  model = torch.jit.trace(model, (d,), check_trace=False, strict=False)
  model = torch.jit.freeze(model)

  model(data)

BFloat16

import torch
from transformers import BertModel
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

model = BertModel.from_pretrained(args.model_name)
model.eval()

vocab_size = model.config.vocab_size
batch_size = 1
seq_length = 512
data = torch.randint(vocab_size, size=[batch_size, seq_length])

#################### code changes #################
model = model.to("xpu")
data = data.to("xpu")
model = ipex.optimize(model, dtype=torch.bfloat16)
#################### code changes #################

with torch.no_grad():
  d = torch.randint(vocab_size, size=[batch_size, seq_length])
  ################################# code changes ######################################
  d = d.to("xpu")
  with torch.xpu.amp.autocast(enabled=True, dtype=torch.bfloat16, cache_enabled=False):
  ################################# code changes ######################################
    model = torch.jit.trace(model, (d,), check_trace=False, strict=False)
    model = torch.jit.freeze(model)

    model(data)

Float16

import torch
from transformers import BertModel
############# code changes ###############
import intel_extension_for_pytorch as ipex
############# code changes ###############

model = BertModel.from_pretrained(args.model_name)
model.eval()

vocab_size = model.config.vocab_size
batch_size = 1
seq_length = 512
data = torch.randint(vocab_size, size=[batch_size, seq_length])

#################### code changes ################
model = model.to("xpu")
data = data.to("xpu")
model = ipex.optimize(model, dtype=torch.float16)
#################### code changes ################

with torch.no_grad():
  d = torch.randint(vocab_size, size=[batch_size, seq_length])
  ################################# code changes ######################################
  d = d.to("xpu")
  with torch.xpu.amp.autocast(enabled=True, dtype=torch.float16, cache_enabled=False):
  ################################# code changes ######################################
    model = torch.jit.trace(model, (d,), check_trace=False, strict=False)
    model = torch.jit.freeze(model)

    model(data)

C++（仅限CPU）

为了支持libtorch（PyTorch的C++库），Intel® 用于PyTorch的扩展还提供了其C++动态库。C++库主要用于处理推理工作，例如服务部署。

示例程序文件（example-app.cpp）

#include <torch/script.h>
#include <iostream>
#include <memory>

int main(int argc, const char* argv[]) {
    torch::jit::script::Module module;
    try {
        module = torch::jit::load(argv[1]);
    }
    catch (const c10::Error& e) {
        std::cerr << "error loading the model\n";
        return -1;
    }
    std::vector<torch::jit::IValue> inputs;
    // make sure input data are converted to channels last format
    inputs.push_back(torch::ones({1, 3, 224, 224}).to(c10::MemoryFormat::ChannelsLast));

    at::Tensor output = module.forward(inputs).toTensor();

    return 0;
}

构建脚本（CMakeLists.txt）

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(example-app)

find_package(intel_ext_pt_cpu REQUIRED)

add_executable(example-app example-app.cpp)
target_link_libraries(example-app "${TORCH_LIBRARIES}")

set_property(TARGET example-app PROPERTY CXX_STANDARD 14)

编译命令

$ cmake -DCMAKE_PREFIX_PATH=<LIBPYTORCH_PATH> ..
$ make

若 Found INTEL_EXT_PT_CPU 显示为 TRUE，则说明扩展已链接到二进制文件中。

$ cmake -DCMAKE_PREFIX_PATH=/workspace/libtorch ..
-- The C compiler identification is GNU 9.3.0
-- The CXX compiler identification is GNU 9.3.0
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD
-- Performing Test CMAKE_HAVE_LIBC_PTHREAD - Failed
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE
-- Found Torch: /workspace/libtorch/lib/libtorch.so
-- Found INTEL_EXT_PT_CPU: TRUE
-- Configuring done
-- Generating done
-- Build files have been written to: /workspace/build

$ ldd example-app
        ...
        libtorch.so => /workspace/libtorch/lib/libtorch.so (0x00007f3cf98e0000)
        libc10.so => /workspace/libtorch/lib/libc10.so (0x00007f3cf985a000)
        libintel-ext-pt-cpu.so => /workspace/libtorch/lib/libintel-ext-pt-cpu.so (0x00007f3cf70fc000)
        libtorch_cpu.so => /workspace/libtorch/lib/libtorch_cpu.so (0x00007f3ce16ac000)
        ...
        libdnnl_graph.so.0 => /workspace/libtorch/lib/libdnnl_graph.so.0 (0x00007f3cde954000)
        ...

模型库（仅限CPU）

已由Intel工程师优化的用例可在 Intel® 架构下的Model Zoo 中找到。

教程

官方Intel® 用于PyTorch的扩展文档中提供了更加详细的教程：

文档

访问 PyTorch 的详细开发者文档

查看文档

教程

获取针对初学者和高级开发人员的深入教程

查看教程

资源

查找开发资源并获得问题的解答

查看资源

为了分析流量并优化您的体验，我们在此网站上提供Cookie。点击或浏览即表示您同意我们使用Cookie。 作为本网站的当前维护者，Facebook的Cookie政策适用。了解更多信息，包括可用的控件： Cookies政策.