Expected behavior

I am trying to import NVIDIA BEVFusion ONNX models into TVM.

testcase.zip

The ONNX models contain two custom sparse operators:

• SparseConvolution
• ScatterDense

I expect the TVM ONNX frontend to either:

1. Directly import these operators into corresponding Relax operators,
   or
2. Provide a recommended extension mechanism for lowering these
   operators into TVM-compatible representations.

Actual behavior

Importing the ONNX models fails because the TVM ONNX frontend does not
recognize these operators.

Error:

OpNotImplemented:
The following operators are not supported for frontend ONNX:
SparseConvolution, ScatterDense

Environment

• TVM: TVM v0.25.dev0
• OS: Ubuntu 22.04
• Python: 3.11

Model information

Example:

testcase

testcase.zip

.onnx

Graph structure:

Input
 |
SparseConvolution x21
 |
Add
 |
Relu
 |
ScatterDense
 |
Output

The operators are custom ONNX operators generated by NVIDIA sparse
convolution implementation.

Questions

1. Should SparseConvolution and ScatterDense be added as new Relax
   operators?

2. Does TVM have an existing recommended extension mechanism for
   importing non-standard ONNX sparse operators?