性能验证

全量性能测试需要覆盖batch_size、input_seq_len、out_seq_len三个维度，默认取值范围：batch_size为1 ~ 32，input_seq_len和out_seq_len为2⁵ ~ 2¹⁰。

测试代码如下，具体可参考“pytorch/examples/xx/zhipu_test.py”：

    # 给定batch/seq_len随机构造指定范围内的输入数据
    past_key_values = None
    input_ids = torch.randint(65504, (batch, seq_len)).npu()
    position_ids = torch.arange(seq_len).reshape([1, seq_len]).npu()
    attention_mask = torch.ones([1, seq_len]).to(torch.int64).npu()
    model_inputs = {
        "input_ids": input_ids,
        "past_key_values": past_key_values,
        "position_ids": position_ids,
        "attention_mask": attention_mask
    }
    torch.npu.synchronize()
    start = time.time()
    # 首次推理
    outputs = model(
        **model_inputs,
        return_dict=True,
        output_attentions=False,
        output_hidden_states=False,
    )
    torch.npu.synchronize()
    end = time.time()
    first_time = (end - start) * 1000
    print(f"first token: {first_time}ms")
    sum_time = 0
    test_cycle -= 1
    input_ids = torch.randint(65504, (batch, 1)).npu()
    position_ids = torch.randint(seq_len, (1, 1)).npu()
    attention_mask = torch.ones([1, 1]).to(torch.int64).npu()
    for i in range(test_cycle):
        past_key_values = outputs.past_key_values
        model_inputs = {
            "input_ids": input_ids,
            "past_key_values": past_key_values,
            "position_ids": position_ids,
            "attention_mask": attention_mask
        }
        torch.npu.synchronize()
        start = time.time()
        # 增量推理
        outputs = model(
            **model_inputs,
            return_dict=True,
            output_attentions=False,
            output_hidden_states=True,
        )
        torch.npu.synchronize()
        end = time.time()
        cur_time = (end - start) * 1000
        sum_time += cur_time
        # print(f"token_{i + 1}: {cur_time}ms")
    avg_time = sum_time / test_cycle
    return first_time, avg_time