Converting GPT-2 into fp16

[liehtman]

Hey guys. I'm trying to convert pretrained gpt2 model to onnx format. The coversion gones well via script onnxruntime_tools/transformers/convert_to_onnx.py. If I use -p fp32 while conversion, the inference is fine. But if i change it to fp16, model starts to output non-sense like it would be if I use wrong tokenizer. My GPU is Tesla V100 16GB. What might be a problem?

[yufenglee]

Add @tianleiwu. @liehtman, are you using a official build or from master? Could you provide more information?