Method mixed_fp16_quantize

deepspeed/runtime/quantize.py:123–127

(self, input, input_q, index)

        return output

    def mixed_fp16_quantize(self, input, input_q, index):
        if self.q_mixed_fp16 and self.q_start_bits[index] >= (self.q_target_bits - 1):
            input_q = input * self.quantize_real_ratio + (1 - self.quantize_real_ratio) * input_q
            return input_q
        return input_q

    def compute_quantization(self, input, index=0, factor=1):
        # fixing the quantization bits based on the training steps