84e32315af7b4a479450a1e65b9e277226d75576,fairseq/optim/fp16_optimizer.py,_FP16OptimizerMixin,step,#_FP16OptimizerMixin#Any#,183

Before Change

        // copy FP32 params back into FP16 model
        if self.has_flat_params:
            offset = 0
            for p in self.fp16_params:
                if not p.requires_grad:
                    continue
                numel = p.data.numel()
                p.data.copy_(self.fp32_params.data[offset:offset+numel].view_as(p.data))
                offset += numel
        else:
            for p, p32 in zip(self.fp16_params, self.fp32_params):
                if not p.requires_grad:
                    continue

After Change

        self._sync_fp16_grads_to_fp32()
        self.fp32_optimizer.step(closure)

        if self.scaler is not None:
            self.scaler.update()

        self._sync_fp32_grads_to_fp16()

    def zero_grad(self):
        Clears the gradients of all optimized parameters.

Instances