Error when extracting top-k logits

Thanks for the awesome work on BERT for Text Generation. But I am facing the following problem when trying to extract topk logits
```python dump_teacher_topk.py --bert_hidden /home/aniruddha/aniruddha-raid/pretam/Distill-BERT-Textgen/data/dump/english computing topk...: 0%| | 0/160239 [00:00
main(args)
File "dump_teacher_topk.py", line 42, in main
topk = linear(bert_hidden).topk(dim=-1, k=opts.topk)
File "/home/aniruddha/anaconda3/envs/bert-textgen/lib/python3.6/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/aniruddha/anaconda3/envs/bert-textgen/lib/python3.6/site-packages/torch/nn/modules/linear.py", line 93, in forward
return F.linear(input, self.weight, self.bias)
File "/home/aniruddha/anaconda3/envs/bert-textgen/lib/python3.6/site-packages/torch/nn/functional.py", line 1690, in linear
ret = torch.addmm(bias, input, weight.t())
RuntimeError: expected scalar type Float but found Half
```

The C-MLM fine-tuning went fine without any errors, and also the task of extracting logits from teacher hidden states. I have switched off the fp16 flag so as not to get the above error, but still getting it. How to mitigate the issue?