the sparse convolution to onnx inference output does not match the torch output

After training with my own data, the sparse convolution to onnx inference output does not match the torch output. It is found that the sparse convolution weight is too large and exceeds the fp16 precision range. What may be the cause?