Closed etrommer closed 8 months ago
TA currently uses its own Quantizer implementation. It would be cleaner to provide approximate layer implementation as subclasses of torch.ao.nn.qat.modules and use the observer/quantizer API that is provided by native PyTorch.
TA currently uses its own Quantizer implementation. It would be cleaner to provide approximate layer implementation as subclasses of torch.ao.nn.qat.modules and use the observer/quantizer API that is provided by native PyTorch.