Sparse Tensors in pytorch 0.2.0

I was using torch.smm for cuda sparse matrix multiplication a few weeks back, after I move to the 0.2.0. the same code will complain,

RuntimeError: WARNING: Sparse Cuda Tensor op sspaddmm is not implemented at /pytorch/torch/lib/THCS/generic/

are the old sparse kernels removed from 0.2.0. Or API has been changed?


When I look at the history of torch/lib/THCS/generic/, it does not look like this kernel was ever implemented. So what is likely happening is that there was some change in how we select kernels and now we are attempting to use an unimplemented kernel. If you posted some code I might be able to say better.

Thanks. I went back and check the code, I was doing torch.smm on CPU. Any plans on implementing smm corresponding kernels on GPU?

It is not a high priority task, but we’ll definitely accept patches :slight_smile: