Anomaly in exporting pre-training weights

[Monitoryang]

When I export the weights dinov2_vitg14_ade20k_m2f.pth using the code below it appears that PyTorch's linear layer (which contains the weights and bias) is broken up into separate MatMul/Add operations, e.g. onnx__Add_15356, model.backbone. blocks.0.mlp.w2.bias, model.decode_head.transformer_decoder.layers.2.norms.1.weight, and similar files like these number over a thousand, resulting in an exported ONNX of only 2MB, which I queried and found seems to be similar to PyTorch's JIT Tracer is not able to fully capture the dynamic logic, resulting in the weights not being captured correctly, how should I proceed to solve the problem?