将上述代码用C++实现

由于C++没有现成的transformers库，需要使用其他第三方库实现类似的功能。这里以HuggingFace公司开发的transformers库为例，具体实现如下：

#include <iostream>
#include <vector>
#include <string>
#include <torch/torch.h>
#include <transformers/modeling_gpt2.h>
#include <transformers/tokenization_gpt2.h>

using namespace transformers;
using namespace torch;

int main() {
    // Load pre-trained model and tokenizer
    std::string model_path = "path/to/gpt2_model";
    std::string vocab_path = "path/to/gpt2_vocab";
    auto model = GPT2Model(GPT2ModelConfig(model_path));
    auto tokenizer = GPT2Tokenizer(vocab_path, true);

    // Prepare input
    std::string prompt = "Hello, how are you?";
    std::vector<std::string> history = {"I'm good, thanks. How about you?", "I'm doing well, thanks for asking."};
    std::string input_text = prompt + "\n" + history[0] + "\n" + history[1] + "\n";
    auto input_tokens = tokenizer.encode(input_text);
    auto input_ids = torch::tensor(input_tokens.get_ids()).unsqueeze(0);
    auto input_mask = torch::tensor(input_tokens.get_mask()).unsqueeze(0);
    auto input_type_ids = torch::tensor(input_tokens.get_type_ids()).unsqueeze(0);

    // Generate response
    auto outputs = model.forward<GPT2LMHeadModelOutput>(
        input_ids.to(torch::kCUDA), input_mask.to(torch::kCUDA), input_type_ids.to(torch::kCUDA));
    auto logits = outputs.logits().squeeze(0);
    auto next_token_logits = logits[input_ids.size(1)-1];
    auto next_token_probabilities = torch::softmax(next_token_logits, 0);
    auto next_token = next_token_probabilities.argmax().item<int64_t>();
    auto response = tokenizer.decode(std::vector<int64_t>{next_token});

    // Output response
    std::cout << "AI: " << response << std::endl;

    return 0;
}

其中，`model_path`和`vocab_path`分别为预训练模型和词表的路径。`prompt`为对话的开头，`history`为之前的对话历史。首先将输入文本编码成token，然后使用GPT2模型生成下一个token的概率分布，并根据分布选出最可能的下一个token，最后将它解码成文本作为回答输出。注意，由于transformers库使用了CUDA加速，因此需要将输入张量转移到GPU上。