本文来源公众号“Coggle数据科学”,仅用于学术分享,侵权删,干货满满。
原文链接:Kaggle 知识点:Torch-TensorRT加速推理
Torch-TensorRT是一个专为PyTorch设计的推理编译器,它通过NVIDIA的TensorRT深度学习优化器和运行时,为NVIDIA GPU提供优化的推理代码。这个工具的目的是提高在NVIDIA GPU上运行PyTorch模型的效率,同时保持开发过程的简便性。
安装Torch-TensorRT
https://pytorch.org/TensorRT/getting_started/installation.html
Torch-TensorRT 2.x主要围绕Python进行开发,因此可以在pypi.org上找到预编译的版本,而不需要从源代码编译。
python -m pip install torch torch-tensorrt tensorrt
基础使用
方法1:使用torch.compile
import torch
import torch_tensorrt
model = MyModel().eval().cuda() # define your model here
x = torch.randn((1, 3, 224, 224)).cuda() # define what the inputs to the model will look like
optimized_model = torch.compile(model, backend="tensorrt")
optimized_model(x) # compiled on first run
optimized_model(x) # this will be fast!
方法2:导出模型torch.export
import torch
import torch_tensorrt
model = MyModel().eval().cuda() # define your model here
inputs = [torch.randn((1, 3, 224, 224)).cuda()] # define a list of representative inputs here
trt_gm = torch_tensorrt.compile(model, ir="dynamo", inputs)
torch_tensorrt.save(trt_gm, "trt.ep", inputs=inputs) # PyTorch only supports Python runtime for an ExportedProgram. For C++ deployment, use a TorchScript file
torch_tensorrt.save(trt_gm, "trt.ts", output_format="torchscript", inputs=inputs)
import torch
import torch_tensorrt
inputs = [torch.randn((1, 3, 224, 224)).cuda()] # your inputs go here
# You can run this in a new python session!
model = torch.export.load("trt.ep").module()
# model = torch_tensorrt.load("trt.ep").module() # this also works
model(*inputs)
处理尺寸输入
默认情况下,可以在PyTorch模型中使用不同输入形状,并在运行时确定输出形状。然而,Torch-TensorRT作为一个提前编译(AOT)编译器,需要一些关于输入形状的先验信息来编译和优化模型。
使用torch.export处理动态形状(AOT)
在动态输入形状的情况下,必须提供(min_shape, opt_shape, max_shape)参数,以便模型可以针对这个输入形状范围进行优化。以下是静态和动态形状的示例用法。
import torch
import torch_tensorrt
model = MyModel().eval().cuda()
# 使用静态形状编译
inputs = torch_tensorrt.Input(shape=[1, 3, 224, 224], dtype=torch.float32)
# 或使用动态形状编译
inputs = torch_tensorrt.Input(min_shape=[1, 3, 224, 224],
opt_shape=[4, 3, 224, 224],
max_shape=[8, 3, 224, 224],
dtype=torch.float32)
trt_gm = torch_tensorrt.compile(model, ir="dynamo", inputs)
自定义动态形状约束
给定输入x = torch_tensorrt.Input(min_shape, opt_shape, max_shape, dtype),Torch-TensorRT尝试在torch.export跟踪期间自动设置约束,通过相应地构建torch.export.Dim对象来实现。有时,可能需要设置额外的约束,如果未指定这些约束,Torchdynamo会报错。如果需要为模型设置任何自定义约束(使用torch.export.Dim),建议先导出程序,然后再用Torch-TensorRT编译。
import torch
import torch_tensorrt
class MatMul(torch.nn.Module):
def __init__(self):
super().__init__()
def forward(self, query, key):
attn_weight = torch.matmul(query, key.transpose(-1, -2))
return attn_weight
model = MatMul().eval().cuda()
inputs = [torch.randn(1, 12, 7, 64).cuda(), torch.randn(1, 12, 7, 64).cuda()]
seq_len = torch.export.Dim("seq_len", min=1, max=10)
dynamic_shapes=({2: seq_len}, {2: seq_len})
# 首先导出模型,自定义动态形状约束
exp_program = torch.export.export(model, tuple(inputs), dynamic_shapes=dynamic_shapes)
trt_gm = torch_tensorrt.dynamo.compile(exp_program, inputs)
# 运行推理
trt_gm(*inputs)
使用torch.compile处理动态形状(JIT)
torch_tensorrt.compile(model, inputs, ir="torch_compile")返回一个配置为TensorRT后端的torch.compile封装函数。在ir=torch_compile的情况下,用户可以使用torch._dynamo.mark_dynamic API为输入提供动态形状信息,以避免TensorRT引擎的重新编译。
import torch
import torch_tensorrt
model = MyModel().eval().cuda()
inputs = torch.randn((1, 3, 224, 224), dtype=float32)
# 这表明第0维是动态的,范围是[1, 8]
torch._dynamo.mark_dynamic(inputs, 0, min=1, max=8)
trt_gm = torch.compile(model, backend="tensorrt")
# 调用模型时发生编译
trt_gm(inputs)
# 修改批量大小时,不重新编译TRT引擎
inputs_bs2 = torch.randn((2, 3, 224, 224), dtype=torch.float32)
trt_gm(inputs_bs2)
进阶案例
案例1:ResNet
import torch
import torch_tensorrt
import torchvision.models as models
# Initialize model with half precision and sample inputs
model = models.resnet18(pretrained=True).half().eval().to("cuda")
inputs = [torch.randn((1, 3, 224, 224)).to("cuda").half()]
# Build and compile the model with torch.compile, using Torch-TensorRT backend
optimized_model = torch_tensorrt.compile(
model,
ir="torch_compile",
inputs=inputs,
enabled_precisions=enabled_precisions,
debug=debug,
workspace_size=workspace_size,
min_block_size=min_block_size,
torch_executed_ops=torch_executed_ops,
)
# Does not cause recompilation (same batch size as input)
new_inputs = [torch.randn((1, 3, 224, 224)).half().to("cuda")]
new_outputs = optimized_model(*new_inputs)
案例2:BERT
import torch
import torch_tensorrt
from transformers import BertModel
# Initialize model with float precision and sample inputs
model = BertModel.from_pretrained("bert-base-uncased").eval().to("cuda")
inputs = [
torch.randint(0, 2, (1, 14), dtype=torch.int32).to("cuda"),
torch.randint(0, 2, (1, 14), dtype=torch.int32).to("cuda"),
]
# Define backend compilation keyword arguments
compilation_kwargs = {
"enabled_precisions": enabled_precisions,
"debug": debug,
"workspace_size": workspace_size,
"min_block_size": min_block_size,
"torch_executed_ops": torch_executed_ops,
}
# Build and compile the model with torch.compile, using Torch-TensorRT backend
optimized_model = torch.compile(
model,
backend="torch_tensorrt",
dynamic=False,
options=compilation_kwargs,
)
optimized_model(*inputs)
THE END !
文章结束,感谢阅读。您的点赞,收藏,评论是我继续更新的动力。大家有推荐的公众号可以评论区留言,共同学习,一起进步。

4564

被折叠的 条评论
为什么被折叠?



