Automatically Generate a Converter for a Custom Kernel

We are going to demonstrate how to automatically generate a converter for a custom kernel using Torch-TensorRT using the new Python based plugin system in TensorRT 10.7.

Torch-TensorRT supports falling back to PyTorch implementations of operations in the case that Torch-TensorRT does not know how to compile them in TensorRT. However, this comes at the cost of a graph break and will reduce the performance of the model. The easiest way to fix lack of support for ops is by adding a decomposition (see: Writing lowering passes for the Dynamo frontend) - which defines the operator in terms of PyTorch ops that are supported in Torch-TensorRT or a converter (see: Writing converters for the Dynamo frontend) - which defines the operator in terms of TensorRT operators.

In some cases there isn’t a great way to do either of these, perhaps because the operator is a custom kernel that is not part of standard PyTorch or TensorRT cannot support it natively.

For these cases, it is possible to use a TensorRT plugin to replace the operator inside the TensorRT engine, thereby avoiding the performance and resource overhead from a graph break.

Previously this involved a complex process in not only building a performant kernel but setting it up to run in TensorRT (see: Using Custom Kernels within TensorRT Engines with Torch-TensorRT). With TensorRT 10.7, there is a new Python native plugin system which greatly streamlines this process. This plugin system also allows Torch-TensorRT to automatically generate the necessary conversion code to convert the operation in PyTorch to TensorRT.

Writing Custom Operators in PyTorch

Pervious tutorials already cover creating custom operators in PyTorch which later get used with Torch-TensorRT.

Here we define a simple elementwise multiplication operator in Triton. This operator is then registered as a custom op in PyTorch. with its host launch code as well as a “meta-kernel”, A meta-kernel is a function that describes the shape and data type transformations that the operator will perform. This meta-kernel is used by Dynamo and Torch-TensorRT, so it is necessary to define.

from typing import Tuple

import tensorrt_bindings.plugin as trtp
import torch
import torch_tensorrt
import triton
import triton.language as tl


@triton.jit
def elementwise_mul_kernel(X, Y, Z, BLOCK_SIZE: tl.constexpr):
    # Program ID determines the block of data each thread will process
    pid = tl.program_id(0)
    # Compute the range of elements that this thread block will work on
    block_start = pid * BLOCK_SIZE
    # Range of indices this thread will handle
    offsets = block_start + tl.arange(0, BLOCK_SIZE)
    # Load elements from the X and Y tensors
    x_vals = tl.load(X + offsets)
    y_vals = tl.load(Y + offsets)
    # Perform the element-wise multiplication
    z_vals = x_vals * y_vals
    # Store the result in Z
    tl.store(Z + offsets, z_vals)


@torch.library.custom_op("torchtrt_ex::elementwise_mul", mutates_args=())  # type: ignore[misc]
def elementwise_mul(
    X: torch.Tensor, Y: torch.Tensor, b: float = 0.2, a: int = 2
) -> torch.Tensor:
    # Ensure the tensors are on the GPU
    assert X.is_cuda and Y.is_cuda, "Tensors must be on CUDA device."
    assert X.shape == Y.shape, "Tensors must have the same shape."

    # Create output tensor
    Z = torch.empty_like(X)

    # Define block size
    BLOCK_SIZE = 1024

    # Grid of programs
    grid = lambda meta: (X.numel() // meta["BLOCK_SIZE"],)

    # Launch the kernel
    elementwise_mul_kernel[grid](X, Y, Z, BLOCK_SIZE=BLOCK_SIZE)

    return Z

The meta kernel for an elementwise operation is just the shape and dtype of one of the inputs since we will not change the shape in the course of the operation.

@torch.library.register_fake("torchtrt_ex::elementwise_mul")
def _(x: torch.Tensor, y: torch.Tensor, b: float = 0.2, a: int = 2) -> torch.Tensor:
    return x

Writing Plugins for TensorRT using the Quick Deploy Plugin system

The quick deployment plugin system in TensorRT 10.7 allows for the creation of custom plugins in Python with significantly less boilerplate. It uses a similar system PyTorch where you define a function that describes the shape and data type transformations that the operator will perform and then define the code to launch the kernel given GPU memory handles.

Just like the PyTorch meta kernel, there is no transformation in shape or data type between the input and output so we can just tell TensorRT to expect the same shape as we get in

@trtp.register("torchtrt_ex::elementwise_mul")
def _(
    x: trtp.TensorDesc, y: trtp.TensorDesc, b: float, a: int
) -> Tuple[trtp.TensorDesc]:
    return x.like()

Here we reuse similar host launch code as PyTorch but we need to convert the TensorRT tensors into PyTorch tensors prior to launching the kernel These operations are also in-place, so the result must be put in the the output tensors provided by TensorRT.

@trtp.impl("torchtrt_ex::elementwise_mul")
def _(
    x: trtp.Tensor,
    y: trtp.Tensor,
    b: float,
    a: int,
    outputs: Tuple[trtp.Tensor],
    stream: int,
):
    # Define block size
    BLOCK_SIZE = 1024

    # Grid of programs
    grid = lambda meta: (x.numel() // meta["BLOCK_SIZE"],)

    x_t = torch.as_tensor(x, device="cuda")
    y_t = torch.as_tensor(y, device="cuda")
    z_t = torch.as_tensor(outputs[0], device="cuda")
    # Launch the kernel
    elementwise_mul_kernel[grid](x_t, y_t, z_t, BLOCK_SIZE=BLOCK_SIZE)

Generating the Converter

Given that we have defined the custom operator in PyTorch and TensorRT, we can now generate the converter for the operation. As long as the namespace and names match, the following function will automatically generate the converter for the operation.

torch_tensorrt.dynamo.conversion.plugins.generate_plugin_converter(
    "torchtrt_ex::elementwise_mul", supports_dynamic_shapes=True
)

Using our converter with a model

Now we can use our custom operator in a model and compile it with Torch-TensorRT. We can see that the custom operator is used as one of the operations in the forward pass of the model. The process of compiling the model at this point is identical to standard Torch-TensorRT usage.

class MyModel(torch.nn.Module):  # type: ignore[misc]
    def __init__(self):
        super().__init__()

    def forward(self, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
        z = torch.add(x, y)
        res = torch.ops.torchtrt_ex.elementwise_mul.default(x, z, a=1)

        return res


my_model = MyModel().to("cuda")
m = torch.full((64, 64), 2, device="cuda", dtype=torch.float)
n = torch.full((64, 64), 3, device="cuda", dtype=torch.float)

with torch_tensorrt.logging.errors():
    model_trt = torch_tensorrt.compile(
        my_model, inputs=[m, n], debug=True, min_block_size=1
    )
    for i in range(300):
        res = model_trt(m, n)
        assert torch.allclose(res, my_model(m, n))

print("Ran with custom plugin!")