使用 Better Transformer 进行快速 Transformer 推断¶
Author: Michael Gschwind
This tutorial introduces Better Transformer (BT) as part of the PyTorch 1.12 release.
In this tutorial, we show how to use Better Transformer for production
inference with torchtext. Better Transformer is a production ready fastpath to
accelerate deployment of Transformer models with high performance on CPU and GPU.
The fastpath feature works transparently for models based either directly on
PyTorch core nn.module or with torchtext.
Models which can be accelerated by Better Transformer fastpath execution are those
using the following PyTorch core torch.nn.module classes TransformerEncoder,
TransformerEncoderLayer, and MultiHeadAttention. In addition, torchtext has
been updated to use the core library modules to benefit from fastpath acceleration.
(Additional modules may be enabled with fastpath execution in the future.)
Better Transformer offers two types of acceleration:
Native multihead attention (MHA) implementation for CPU and GPU to improve overall execution efficiency.
Exploiting sparsity in NLP inference. Because of variable input lengths, input tokens may contain a large number of padding tokens for which processing may be skipped, delivering significant speedups.
Fastpath execution is subject to some criteria. Most importantly, the model must be executed in inference mode and operate on input tensors that do not collect gradient tape information (e.g., running with torch.no_grad).
To follow this example in Google Colab, click here.
Better Transformer Features in This Tutorial¶
Load pretrained models (created before PyTorch version 1.12 without Better Transformer)
Run and benchmark inference on CPU with and without BT fastpath (native MHA only)
Run and benchmark inference on (configurable) DEVICE with and without BT fastpath (native MHA only)
Enable sparsity support
Run and benchmark inference on (configurable) DEVICE with and without BT fastpath (native MHA + sparsity)
Additional Information¶
Additional information about Better Transformer may be found in the PyTorch.Org blog A Better Transformer for Fast Transformer Inference.
Setup
1.1 Load pretrained models
We download the XLM-R model from the predefined torchtext models by following the instructions in torchtext.models. We also set the DEVICE to execute on-accelerator tests. (Enable GPU execution for your environment as appropriate.)
import torch
import torch.nn as nn
print(f"torch version: {torch.__version__}")
DEVICE = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
print(f"torch cuda available: {torch.cuda.is_available()}")
import torch, torchtext
from torchtext.models import RobertaClassificationHead
from torchtext.functional import to_tensor
xlmr_large = torchtext.models.XLMR_LARGE_ENCODER
classifier_head = torchtext.models.RobertaClassificationHead(num_classes=2, input_dim = 1024)
model = xlmr_large.get_model(head=classifier_head)
transform = xlmr_large.transform()
1.2 Dataset Setup
We set up two types of inputs: a small input batch and a big input batch with sparsity.
small_input_batch = [
"Hello world",
"How are you!"
]
big_input_batch = [
"Hello world",
"How are you!",
"""`Well, Prince, so Genoa and Lucca are now just family estates of the
Buonapartes. But I warn you, if you don't tell me that this means war,
if you still try to defend the infamies and horrors perpetrated by
that Antichrist- I really believe he is Antichrist- I will have
nothing more to do with you and you are no longer my friend, no longer
my 'faithful slave,' as you call yourself! But how do you do? I see
I have frightened you- sit down and tell me all the news.`
It was in July, 1805, and the speaker was the well-known Anna
Pavlovna Scherer, maid of honor and favorite of the Empress Marya
Fedorovna. With these words she greeted Prince Vasili Kuragin, a man
of high rank and importance, who was the first to arrive at her
reception. Anna Pavlovna had had a cough for some days. She was, as
she said, suffering from la grippe; grippe being then a new word in
St. Petersburg, used only by the elite."""
]
Next, we select either the small or large input batch, preprocess the inputs and test the model.
input_batch=big_input_batch
model_input = to_tensor(transform(input_batch), padding_value=1)
output = model(model_input)
output.shape
Finally, we set the benchmark iteration count:
ITERATIONS=10
Execution
2.1 Run and benchmark inference on CPU with and without BT fastpath (native MHA only)
We run the model on CPU, and collect profile information:
The first run uses traditional (“slow path”) execution.
The second run enables BT fastpath execution by putting the model in inference mode using model.eval() and disables gradient collection with torch.no_grad().
You can see an improvement (whose magnitude will depend on the CPU model) when the model is executing on CPU. Notice that the fastpath profile shows most of the execution time in the native TransformerEncoderLayer implementation aten::_transformer_encoder_layer_fwd.
print("slow path:")
print("==========")
with torch.autograd.profiler.profile(use_cuda=False) as prof:
for i in range(ITERATIONS):
output = model(model_input)
print(prof)
model.eval()
print("fast path:")
print("==========")
with torch.autograd.profiler.profile(use_cuda=False) as prof:
with torch.no_grad():
for i in range(ITERATIONS):
output = model(model_input)
print(prof)
2.2 Run and benchmark inference on (configurable) DEVICE with and without BT fastpath (native MHA only)
We check the BT sparsity setting:
model.encoder.transformer.layers.enable_nested_tensor
We disable the BT sparsity:
model.encoder.transformer.layers.enable_nested_tensor=False
We run the model on DEVICE, and collect profile information for native MHA execution on DEVICE:
The first run uses traditional (“slow path”) execution.
The second run enables BT fastpath execution by putting the model in inference mode using model.eval() and disables gradient collection with torch.no_grad().
When executing on a GPU, you should see a significant speedup, in particular for the small input batch setting:
model.to(DEVICE)
model_input = model_input.to(DEVICE)
print("slow path:")
print("==========")
with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(ITERATIONS):
output = model(model_input)
print(prof)
model.eval()
print("fast path:")
print("==========")
with torch.autograd.profiler.profile(use_cuda=True) as prof:
with torch.no_grad():
for i in range(ITERATIONS):
output = model(model_input)
print(prof)
2.3 Run and benchmark inference on (configurable) DEVICE with and without BT fastpath (native MHA + sparsity)
We enable sparsity support:
model.encoder.transformer.layers.enable_nested_tensor = True
We run the model on DEVICE, and collect profile information for native MHA and sparsity support execution on DEVICE:
The first run uses traditional (“slow path”) execution.
The second run enables BT fastpath execution by putting the model in inference mode using model.eval() and disables gradient collection with torch.no_grad().
When executing on a GPU, you should see a significant speedup, in particular for the large input batch setting which includes sparsity:
model.to(DEVICE)
model_input = model_input.to(DEVICE)
print("slow path:")
print("==========")
with torch.autograd.profiler.profile(use_cuda=True) as prof:
for i in range(ITERATIONS):
output = model(model_input)
print(prof)
model.eval()
print("fast path:")
print("==========")
with torch.autograd.profiler.profile(use_cuda=True) as prof:
with torch.no_grad():
for i in range(ITERATIONS):
output = model(model_input)
print(prof)
Summary¶
In this tutorial, we have introduced fast transformer inference with Better Transformer fastpath execution in torchtext using PyTorch core Better Transformer support for Transformer Encoder models. We have demonstrated the use of Better Transformer with models trained prior to the availability of BT fastpath execution. We have demonstrated and benchmarked the use of both BT fastpath execution modes, native MHA execution and BT sparsity acceleration.
