Kaiyu Shi

Fix for #54

### Reproduce scripts ```python for _ in range(10000): from kaldi.fstext import LatticeVectorFst ori_lat = LatticeVectorFst() for _ in range(1000): ori_lat.add_state() byte = ori_lat.to_bytes() for _ in range(100): lat = LatticeVectorFst.from_bytes(byte)...

The existing examples are so old and contains some un-expected library dependencies like in `examples/Gemm` folder.

Not implemented for type Half

4

I followed the tutorial to create my cpp_extension. But the module failed to do half-precision computation. It gives: ``` RuntimeError: "op" not implemented for 'Half' (operator() at /op_cuda_kernel.cu:146) frame #0:...

On reusing codes from gensim

2

I found that `gensim` does a good job of embedding tasks. If you could reuse the `embedding model` definition from gensim, then you can put your major focus on various...

Why use float_64 for gradient

6

Hi wang, I'm just wondering why to convert the gradient Tensor into `float64`, I thought they might be just `float32`. And it should be more accurate than SGD required. https://github.com/hwang595/ps_pytorch/blob/89a1cfa136b957073576fae827d39ef0fb09d2fc/src/distributed_worker.py#L258

Since it's supposed to deal with non-ascii characters, it's better to ensure open the text file as 'utf-8' encoding. It fixes my problem when post-processing the corpus.