GroupNormalization scale and bias issue

[hrishi121]

Description

I have a ONNX model (opset21) that has a groupnorm layer. The input num_channels for the groupnorm layer is 16 and num_groups is 4. The shape of scale and bias affine param arrays in the groupnorm layer is 16. If I run a forward pass with this ONNX model, its output matches with the PyTorch output. However, I am not able to parse this model using trtexec --onnx=model.onnx command.

Using onnx-graphsurgeon, if I manually set the scale and bias array in the ONNX graph to an array with shape 4 (which is the same value as num_groups), I am able to parse and serialize the ONNX model using trtexec.

However, this means there's a mismatch in the GroupNorm definition in ONNX+PyTorch vs TensorRT
ONNX and PyTorch allow for separate affine scale and bias params per channel, whereas it seems TRT only allows for scale and bias per group. I was able to confirm this on the TensorRT operator documentation page as well - link, where it states that scale and bias must have a shape of (1,G,1,1), where G is num_groups.

Is there a workaround for this issue? I tried using GroupNormalizationPlugin provided in this repo, but I am having problems loading that plugin as well (have created a separate issue for it)

Environment

TensorRT Version: 10.7.0.23

NVIDIA GPU: Jetson AGX Orin

NVIDIA Driver Version:

CUDA Version: 12.6

CUDNN Version:

Operating System: Jetson native build

Python Version (if applicable):

Tensorflow Version (if applicable):

PyTorch Version (if applicable): 2.5.1

Baremetal or Container (if so, version):

Relevant Files

Model link:

Steps To Reproduce

Commands or scripts:

Have you tried the latest release?: Yes

Can this model run on other frameworks? For example run ONNX model with ONNXRuntime (polygraphy run <model.onnx> --onnxrt):

[kevinch-nv]

Fix for this is being tracked in #4336. Closing this as dup.