Gelu schema change updates

[peri044]

During development of Pytorch 1.10, the schema of gelu layer changed from being
aten::gelu(Tensor self, bool approximate) -> (Tensor) to aten::gelu(Tensor self) -> (Tensor)

NVIDIA's NGC Pytorch containers 21.11 and 21.12 contains gelu of the following form (Pytorch version: 1.11.0a0+b6df043)

aten::gelu(Tensor self, bool approximate) -> (Tensor)

NGC container 22.01 (Pytorch version : 1.11.0a0+bfe5ad2) also contains gelu of the form

aten::gelu(Tensor self, bool approximate) -> (Tensor)

NVIDIA NGC container is using unmerged pull request pytorch/pytorch#61439.

Public pytorch 1.10.0 and 1.10.2 contains gelu op of the form

aten::gelu(Tensor self) -> (Tensor)

Torch-TensorRT available in master branch and shipped with 21.11, 21.12 and 22.01 currently supports the public pytorch's gelu schema . Hence, if your model (which has gelu layer) is converted into torchscript inside any of 21.11, 21.12 and 22.01 containers and compiled with Torch-TensorRT, you will face the following error

ERROR: [Torch-TensorRT] - Unsupported operator: aten::gelu(Tensor self, bool approximate) -> (Tensor)

because Torch-TensorRT is expecting aten::gelu(Tensor self).

We are working on this to resolve it in future NGC containers. But here is a temporary solution

Solution:

Rebuild the Torch-TensorRT inside the docker container with this patch #845.
You can follow the instructions here https://github.com/NVIDIA/Torch-TensorRT/tree/master/docker

Say if you're using existing 22.01 container, include the changes in the patch to your current Torch-TensorRT directory and build an image by running

docker build --build-arg BASE=22.01 -f docker/Dockerfile -t torch_tensorrt1.0:latest .

This will build a new docker container very similar to Pytorch NGC 22.01 container with Torch-TensorRT reinstalled (make sure the patch is applied before building). This new docker container would support both schema variants of gelu.