Method update_fp16_ratio

deepspeed/runtime/quantize.py:175–180

(self)

            return input_flat

    def update_fp16_ratio(self):
        if self.q_mixed_fp16:
            if self.quantize_real_ratio > 0:
                self.quantize_real_ratio -= self.q_change_ratio
            else:
                self.quantize_real_ratio = 0.000