Tao Lei

@liziru It is version 2 by default. But you can test v1 by passing `v1 = True` in `SRU` or `SRUCell` constructor. See https://github.com/asappresearch/sru/blob/master/sru/modules.py#L444

hi, could you give me more context? the code or the input specification for example.

It seems a lot people are having issues compiling the cpp code using `ninja`. This is only needed for inference on CPU. I'll release version that makes this optional..

Have a PR open to address the issue: https://github.com/taolei87/sru/pull/76

It seems a lot people are having issues compiling the cpp code using `ninja`. This is only needed for inference on CPU. I'll release version that makes this optional..

Have a PR open to address the issue: https://github.com/taolei87/sru/pull/76

I will rename the parameter for consistency.

@loveJasmine I haven't tested the time. A rough estimate is that each LSTM layer has eight d*d matrix multiplications, while SRU has three. So SRU should have some advantage unless...

hi @v-nhandt21 , the compilation arguments such as "--generate-dependencies-with-compile" are automatically added by `ninja` / `nvcc`. Looking at your first screenshot, ninja/nvcc attempts to build the code using `--sm_75` and...

Hi @ctlaltdefeat , Could you also provide some details: - are you trying to use the torchscript model in Python or C++? - are you trying to use the torchscript...