[nighthawk454]

Yeah, but it’s ’quantization aware’ during training too, which presumably is what allows the quantization at inference to work