AIACC-Inference(AIACC推理加速)支持优化基于Torch框架搭建的模型,能够显著提升推理性能。本文介绍如何手动安装AIACC-Inference(AIACC推理加速)Torch版并提供示例体验推理加速效果。

前提条件

已创建阿里云GPU实例:
  • 实例规格:配备NVIDIA A100、A10、V100或T4 GPU。
    说明 更多信息,请参见实例规格族
  • 实例镜像:Ubuntu 16.04 LTS或CentOS 7.x。

背景信息

AIACC-Inference(AIACC推理加速)Torch版通过对模型的计算图进行切割,执行层间融合,以及高性能OP实现,大幅度提升PyTorch的推理性能。您无需指定精度和输入尺寸,即可通过JIT编译的方式对PyTorch框架下的深度学习模型进行推理优化。

AIACC-Inference(AIACC推理加速)Torch版通过调用aiacctorch.compile(model)接口即可实现推理性能加速。您只需先使用torch.jit.script或者torch.jit.trace接口,将PyTorch模型转换为TorchScript模型,更多信息,请参见PyTorch官方文档。本文将为您提供分别使用torch.jit.scripttorch.jit.trace接口实现推理性能加速的示例。

准备并安装AIACC-Inference(AIACC推理加速)Torch版软件包

AIACC-Inference(AIACC推理加速)Torch版为您提供了Conda一键安装包以及whl包两种软件包,您可以根据自身业务场景选择一种进行安装。

  • Conda安装包

    Conda一键安装包中已经预装了大部分依赖包,您只需手动安装CUDA驱动,再安装Conda包即可。具体操作如下:

    注意 请勿随意更改Conda安装包中的预装依赖包信息,否则可能会因为版本不匹配导致Demo运行报错。
    1. 远程登录实例
    2. 自行安装CUDA 470.57.02或以上版本的驱动。
    3. 下载Conda安装包。
      wget https://aiacc-inference-public.oss-cn-beijing.aliyuncs.com/aiacc-inference-torch/aiacc-inference-torch-miniconda-latest.tar.bz2
    4. 解压Conda安装包。
      mkdir ./aiacc-inference-miniconda && tar -xvf ./aiacc-inference-torch-miniconda-latest.tar.bz2 -C ./aiacc-inference-miniconda
    5. 加载Conda安装包。
      source ./aiacc-inference-miniconda/bin/activate
  • whl安装包
    您需要手动安装相关依赖包后再安装whl软件包。具体操作如下:
    1. 远程登录实例
    2. 选择以下任一方式安装相关依赖包。由于whl软件包依赖大量不同的软件组合,请您谨慎设置。
      • 方式一
        1. 自行安装如下版本的依赖包:
          • CUDA 11.1
          • cuDNN 8.3.1.22
          • TensorRT 8.2.3.0
        2. 将TensorRT及CUDA的相关依赖库放置在系统LD_LIBRARY_PATH环境变量中。

          以下命令以CUDA的相关依赖库位于/usr/local/cuda/目录下,TensorRT的相关依赖库位于/usr/local/TensorRT/目录下为例,您需要根据实际情况替换。

          export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
          export LD_LIBRARY_PATH=/usr/local/TensorRT/lib:$LD_LIBRARY_PATH
        3. 执行环境变量。
          source ~/.bashrc
      • 方式二

        使用NVIDIA的pip包安装相关依赖包。

        pip install nvidia-pyindex && \ pip install nvidia-tensorrt==8.2.3.0
    3. 安装PyTorch 1.9.0+cu111。
      pip3 install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
    4. 下载并安装aiacctorch。
      pip install aiacctorch -f https://aiacc-inference-public.oss-cn-beijing.aliyuncs.com/aiacc-inference-torch/aiacctorch_stable.html -f https://download.pytorch.org/whl/torch_stable.html

基于ResNet50模型执行推理

以下示例将以安装了Conda软件包为例,基于ResNet50模型,并调用torch.jit.script接口执行推理任务,执行1000次后取平均时间,将推理耗时从3.68 ms降低至0.396 ms以内。

  • 原始版本

    原始代码如下所示:

    import time
    import torch
    import torchvision.models as models
    mod = models.resnet50(pretrained=True).eval()
    mod_jit = torch.jit.script(mod)
    mod_jit = mod_jit.cuda()
    
    in_t = torch.randn([1,3,224,224]).float().cuda()
    
    # Warming up
    for _ in range(10):
        mod_jit(in_t)
    
        inference_count = 1000
        # inference test
        start = time.time()
        for _ in range(inference_count):
            mod_jit(in_t)
            end = time.time()
            print(f"use {(end-start)/inference_count*1000} ms each inference")
    print(f"{inference_count/(end-start)} step/s")

    执行结果如下,显示推理耗时大约为3.68 ms。

    aiacc3.68
  • 加速版本

    您仅需要在原始版本的示例代码中增加如下两行内容,即可实现性能加速。

    import aiacctorch
    aiacctorch.compile(mod_jit)

    更新后的代码如下:

    import time
    import aiacctorch #import aiacc包
    import torch
    import torchvision.models as models
    mod = models.resnet50(pretrained=True).eval()
    mod_jit = torch.jit.script(mod)
    mod_jit = mod_jit.cuda()
    mod_jit = aiacctorch.compile(mod_jit) #进行编译
    
    in_t = torch.randn([1,3,224,224]).float().cuda()
    
    # Warming up
    for _ in range(10):
        mod_jit(in_t)
    
    inference_count = 1000
    # inference test
    start = time.time()
    for _ in range(inference_count):
        mod_jit(in_t)
    end = time.time()
    print(f"use {(end-start)/inference_count*1000} ms each inference")
    print(f"{inference_count/(end-start)} step/s")

    执行结果如下,显示推理耗时为0.396 ms。相较于之前的3.68 ms,推理性能有了显著提升。

    2022-04-19_17-53-05.png

基于Bert-Base模型执行推理

以下示例将基于Bert-Base模型,并调用torch.jit.trace接口执行推理任务,将推理耗时从4.95 ms降低至0.419 ms以内。

  1. 安装transformers包。
    pip install transformers
  2. 分别运行原始版本和加速版本的Demo,并查看运行结果。
    • 原始版本

      原始代码如下所示:

      from transformers import BertModel, BertTokenizer, BertConfig
      import torch
      import time
      
      enc = BertTokenizer.from_pretrained("bert-base-uncased")
      
      # Tokenizing input text
      text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
      tokenized_text = enc.tokenize(text)
      
      # Masking one of the input tokens
      masked_index = 8
      tokenized_text[masked_index] = '[MASK]'
      indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
      segments_ids = [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, ]
      
      # Creating a dummy input
      tokens_tensor = torch.tensor([indexed_tokens]).cuda()
      segments_tensors = torch.tensor([segments_ids]).cuda()
      dummy_input = [tokens_tensor, segments_tensors]
      
      # Initializing the model with the torchscript flag
      # Flag set to True even though it is not necessary as this model does not have an LM Head.
      config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
                          num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
      
      # Instantiating the model
      model = BertModel(config)
      
      # The model needs to be in evaluation mode
      model.eval()
      
      # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
      model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
      
      model = model.eval().cuda()
      
      # Creating the trace
      traced_model = torch.jit.trace(model, dummy_input)
      
      # Warming up
      for _ in range(10):
          all_encoder_layers, pooled_output = traced_model(*dummy_input)
      
          inference_count = 1000
          # inference test
          start = time.time()
          for _ in range(inference_count):
              traced_model(*dummy_input)
              end = time.time()
              print(f"use {(end-start)/inference_count*1000} ms each inference")
      print(f"{inference_count/(end-start)} step/s")

      执行结果如下,显示推理耗时大约为4.95 ms。

      2022-04-19_17-54-40.png
    • 加速版本

      您仅需要在原始版本的示例代码中增加如下两行内容,即可实现性能加速。

      import aiacctorch
      aiacctorch.compile(traced_model)

      更新后的代码如下:

      from transformers import BertModel, BertTokenizer, BertConfig
      import torch
      import aiacctorch #import aiacc包
      import time
      
      enc = BertTokenizer.from_pretrained("bert-base-uncased")
      
      # Tokenizing input text
      text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
      tokenized_text = enc.tokenize(text)
      
      # Masking one of the input tokens
      masked_index = 8
      tokenized_text[masked_index] = '[MASK]'
      indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
      segments_ids = [1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, ]
      
      # Creating a dummy input
      tokens_tensor = torch.tensor([indexed_tokens]).cuda()
      segments_tensors = torch.tensor([segments_ids]).cuda()
      dummy_input = [tokens_tensor, segments_tensors]
      
      # Initializing the model with the torchscript flag
      # Flag set to True even though it is not necessary as this model does not have an LM Head.
      config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
          num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
      
      # Instantiating the model
      model = BertModel(config)
      
      # The model needs to be in evaluation mode
      model.eval()
      
      # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
      model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
      
      model = model.eval().cuda()
      
      # Creating the trace
      traced_model = torch.jit.trace(model, dummy_input)
      traced_model = aiacctorch.compile(traced_model) #进行编译
      
      # Warming up
      for _ in range(10):
          all_encoder_layers, pooled_output = traced_model(*dummy_input)
      
      inference_count = 1000
      # inference test
      start = time.time()
      for _ in range(inference_count):
          traced_model(*dummy_input)
      end = time.time()
      print(f"use {(end-start)/inference_count*1000} ms each inference")
      print(f"{inference_count/(end-start)} step/s")

      执行结果如下,显示推理耗时为0.419 ms。相较于之前的4.95 ms,推理性能有了显著提升。

      2022-04-19_17-56-13.png

基于ResNet50模型执行动态尺寸推理

在AIACC-Inference-Torch中,我们无需关心动态尺寸的问题,AIACC-Inference-Torch能够支持不同的输入尺寸。以下示例基于ResNet50模型,输入3个不同的长宽尺寸,带您体验使用AIACC-Inference-Torch进行加速的过程。

import time
import aiacctorch #import aiacc包
import torch
import torchvision.models as models
mod = models.resnet50(pretrained=True).eval()
mod_jit = torch.jit.script(mod)
mod_jit = mod_jit.cuda()
mod_jit = aiacctorch.compile(mod_jit) #进行编译

in_t = torch.randn([1,3,224,224]).float().cuda()
in_2t = torch.randn([1,3,448,448]).float().cuda()
in_3t = torch.randn([16,3,640,640]).float().cuda()

# Warming up
for _ in range(10):
    mod_jit(in_t)
    mod_jit(in_3t)

inference_count = 1000
# inference test
start = time.time()
for _ in range(inference_count):
    mod_jit(in_t)
    mod_jit(in_2t)
    mod_jit(in_3t)
end = time.time()
print(f"use {(end-start)/(inference_count*3)*1000} ms each inference")
print(f"{inference_count/(end-start)} step/s")
执行结果如下,显示推理耗时大约为9.84 ms。2022-04-19_18-00-12.png
说明 为了缩短模型编译的时间,应在warming up阶段推理最大及最小的tensor尺寸,避免在执行时重复编译。例如,已知推理尺寸在1×3×224×224至16×3×640×640之间时,应在warming up时推理这两个尺寸。

性能数据对比参考

以下数据为AIACC-Inference-HRT与PyTorch的性能对比结果,采用的环境配置如下:
  • 实例规格:配置NVIDIA A10的GPU实例。
  • CUDA版本:11.5。
  • CUDA Driver版本:470.57.02。
Model Input-Size AIACC-Inference-Torch (ms) Pytorch Half (ms) Pytorch Float (ms) 加速比
resnet50 1x3x224x224 0.46974873542785645 3.4382946491241455 2.9194235801696777 6.22
mobilenet-v2-100 1x3x224x224 0.23872756958007812 2.8045766353607178 2.0068271160125732 8.69
SRGAN-X4 1x3x272x480 23.070229649543762 35.863523721694946 132.00348043441772 5.74
YOLO-V3 1x3x640x640 3.869319200515747 8.807475328445435 15.704705834388735 4.06
bert-base-uncased 1x128,1x128 0.9421144723892212 3.1525989770889282 3.761411190032959 4.00
bert-large-uncased 1x128,1x128 1.3300731182098389 6.11789083480835 7.110695481300354 5.34