Reduction not supported with ScatterNd in opset 16

[aboubezari]

Description

The ONNX ScatterND layer in opset 16 supports a reduction argument.
In onnx2trt, ScatterND is simply lowered to iScatterLayer, which does not take a reduction argument.
We should update this with a clear error message or add reduction support.

Environment

TensorRT Version: 8.6.1
ONNX-TensorRT Version / Branch: 8.6 GA release
GPU Type: A4000
Nvidia Driver Version: 535
CUDA Version: 12.2
CUDNN Version: 8.9.5
TensorFlow + TF2ONNX Version (if applicable): 2.15, 1.17

Steps To Reproduce

Use any onnx model with ScatterND max reduction and you can see the TensorRT model does not respect reductions.

[gcunhase]

Are there any updates for this bug?

[jcdatin]

I am crashing my model generated with pytorch onnx export model in opset 17 with ORT 1.20.1 and TRT 10.7 on ScatterND operator . It used to work fine on ORT 1.18.0 and TRT 8.6, Cudnn 8.9 :

2025-02-04 12:16:36.012390098 [E:onnxruntime:ivpSelectorInference, tensorrt_execution_provider.h:88 log] [2025-02-04 12:16:36 ERROR] In node 302 with name: /model/my_model/rpn/ScatterND and operator: ScatterND (importScatterND): UNSUPPORTED_NODE_ATTR: Assertion failed: !attrs.count("reduction"): Attribute reduction is not supported.
see end of discussion of microsoft/onnxruntime#17434

Note that TRT exec accpets the same model properly

so similarly I would like to know status of ORT support of the opset 16 and beyond ScatterND operator