I have found several documentation and tools for converting ONNX model to float16, but none of them supports converting to bfloat16.

The model is originally trained using tensorflow and converted to ONNX. I do not have access to the original tensorflow code/models.

Do you think I can convert the model to tensorflow then convert back to onnx as bfloat16? Will this quantization "dilutes" the trained model weights?

Please advice.

I have looked into tools for converting onnx models, but only fp16 is found, not bfp16.