Panic trying to input ONNX Model for ModernBERT

[hjarrell]

Describe the bug
Hello, I was trying to import https://huggingface.co/answerdotai/ModernBERT-base ONNX model and it panicked trying to parse Shape_4.

I am not an ML expert by any means but tried to do some basic troubleshooting:

• I did try to open it in netron and shape_4 input appears to be a tensor: float32[batch_size,1,sequence_length,sequence_length]
• I checked the onnx-ir crate and saw something about updating the model version so tried that with the python library and same issue
• Found Mismatched types when importing ONNX models #2806 which also had an issue importing the same model but its a different issue

  DEBUG onnx_ir::proto_conversion: Converting ONNX node with type "Shape"    
  DEBUG onnx_ir::from_onnx: renaming node "/model/Shape_4"    
  ERROR burn_import::logger: PANIC => panicked at /[...]/onnx-ir-0.17.0/src/util.rs:14:14:
  Only tensor input is valid    

  --- stderr

  thread 'main' panicked at /[...]/onnx-ir-0.17.0/src/util.rs:14:14:

To Reproduce

1. Download model.onnx from https://huggingface.co/answerdotai/ModernBERT-base/tree/main/onnx
2. Run:

use burn_import::onnx::ModelGen;

fn main() {
    ModelGen::new()
        .input("src/model/model.onnx")
        .out_dir("model/")
        .run_from_script();
}

3. Panic

Expected behavior
An output of rust code for the model.

Screenshots
N/A

Desktop (please complete the following information):

• OS: macOS 15.5
• Browser N/A

[laggui]

Thanks for flagging the issue! The ONNX import module is going through some changes to improve it.

A quick glance at the error tells me this might be a scalar input issue. But given that the input to the node is a tensor with shape [batch_size,1,sequence_length,sequence_length], it's probably an incorrect conversion when parsing one of the preceding graph nodes.

[antimora]

I'll investigate after the big refactoring PRs are merged.