使用Cudagraphs导出Torch

这个交互式脚本旨在概述在ir=”dynamo”路径中使用Torch-TensorRT Cudagraphs集成的过程。该功能在torch.compile路径中也类似工作。

导入和模型定义

import torch
import torch_tensorrt
import torchvision.models as models

使用默认设置通过torch_tensorrt.compile进行编译

# We begin by defining and initializing a model
model = models.resnet18(pretrained=True).eval().to("cuda")

# Define sample inputs
inputs = torch.randn((16, 3, 224, 224)).cuda()

# Next, we compile the model using torch_tensorrt.compile
# We use the `ir="dynamo"` flag here, and `ir="torch_compile"` should
# work with cudagraphs as well.
opt = torch_tensorrt.compile(
    model,
    ir="dynamo",
    inputs=torch_tensorrt.Input(
        min_shape=(1, 3, 224, 224),
        opt_shape=(8, 3, 224, 224),
        max_shape=(16, 3, 224, 224),
        dtype=torch.float,
        name="x",
    ),
)

使用Cudagraphs集成进行推理

# We can enable the cudagraphs API with a context manager
with torch_tensorrt.runtime.enable_cudagraphs(opt) as cudagraphs_module:
    out_trt = cudagraphs_module(inputs)

# Alternatively, we can set the cudagraphs mode for the session
torch_tensorrt.runtime.set_cudagraphs_mode(True)
out_trt = opt(inputs)

# We can also turn off cudagraphs mode and perform inference as normal
torch_tensorrt.runtime.set_cudagraphs_mode(False)
out_trt = opt(inputs)

# If we provide new input shapes, cudagraphs will re-record the graph
inputs_2 = torch.randn((8, 3, 224, 224)).cuda()
inputs_3 = torch.randn((4, 3, 224, 224)).cuda()

with torch_tensorrt.runtime.enable_cudagraphs(opt) as cudagraphs_module:
    out_trt_2 = cudagraphs_module(inputs_2)
    out_trt_3 = cudagraphs_module(inputs_3)

包含图形中断的模块的Cuda图形

当CUDA Graphs应用于包含图断点的TensorRT模型时，每个断点都会引入额外的开销。这是因为图断点阻止了整个模型作为一个单一的、连续的优化单元执行。因此，CUDA Graphs通常提供的一些性能优势，如减少内核启动开销和提高执行效率，可能会减弱。 使用带有CUDA Graphs的封装运行时模块，您可以将操作序列封装成可以高效执行的图，即使存在图断点也是如此。 如果TensorRT模块有图断点，CUDA Graph上下文管理器会返回一个wrapped_module。该模块捕获整个执行图，通过减少内核启动开销和提高性能，在后续推理中实现高效重放。请注意，使用封装模块进行初始化涉及一个预热阶段，其中模块会执行多次。此预热确保内存分配和初始化不会记录在CUDA Graphs中，这有助于保持一致的执行路径并优化性能。

class SampleModel(torch.nn.Module):
    def forward(self, x):
        return torch.relu((x + 2) * 0.5)


model = SampleModel().eval().cuda()
input = torch.randn((1, 3, 224, 224)).to("cuda")

# The 'torch_executed_ops' compiler option is used in this example to intentionally introduce graph breaks within the module.
# Note: The Dynamo backend is required for the CUDA Graph context manager to handle modules in an Ahead-Of-Time (AOT) manner.
opt_with_graph_break = torch_tensorrt.compile(
    model,
    ir="dynamo",
    inputs=[input],
    min_block_size=1,
    pass_through_build_failures=True,
    torch_executed_ops={"torch.ops.aten.mul.Tensor"},
)

如果模块有图形中断，整个子模块将由cuda图形记录和重放

with torch_tensorrt.runtime.enable_cudagraphs(
    opt_with_graph_break
) as cudagraphs_module:
    cudagraphs_module(input)