Does INC not support torch.bfloat16 during the pruning process

[tianyil1]

Hi, when I use the magnitude pruning on models with torch.bfloat16 tensor, if fails:

TypeError: Got unsupported ScalarType BFloat16

I find this error is due to the unsupported operation when converting torch.bfloat16 to numpy array. Why INC does not try to support bfloat16 operation? Is there any special consideration?

[ftian1]

thanks for raising such request. yes, currently the magnitude pruning only supports FP32. we will put this under consideration.