Unable to infer on cpu a model trained on fp16

1

Hi I have trained the transformer network from fairseq using fp16 on turing gpu. The result I get a model about half the size of original fp32 model I trained. Now I want to infer the fp16 model trained on cpu and I get the following error.
RuntimeError: _th_index_select is not implemented for type torch.HalfTensor

Is there something I missed ? Please help

2

I have the same problem. I export a FP16 model, and it works fine on GPU, but once switch to CPU I got this:

_thnn_conv2d_forward is not implemented for type CPUHalfType (_thnn_conv2d_forward at aten/src/ATen/TypeDefault.cpp:1940)