简介

class LLMReq {
 public:
  LLMReq() = default;
  ~LLMReq() = default;
  //设置请求的ID，不同请求的ID不能重复
  void SetReqId(const uint64_t req_id) {
    req_id_ = req_id;
  }
  //获取请求ID
  uint64_t GetReqId() const {
    return req_id_;
  }
  //设置全量kv的长度，当前暂没用到
  void SetPromptLength(const uint64_t prompt_length) {
    prompt_length_ = prompt_length;
  }
  //获取kv长度
  uint64_t GetPromptLength() const {
    return prompt_length_;
  }
 //设置请求所在的全量集群ID
  void SetPromptClusterId(const uint64_t prompt_cluster_id) {
    prompt_cluster_id_ = prompt_cluster_id;
  }
 //获取请求所在的全量集群ID
  uint64_t GetPromptClusterId() const {
    return prompt_cluster_id_;
  }
  //设置请求所在的增量集群ID
  void SetDecoderClusterId(const uint64_t decoder_cluster_id) {
    decoder_cluster_id_ = decoder_cluster_id;
  }
  //获取请求所在的增量集群ID
  uint64_t GetDecoderClusterId() const {
    return decoder_cluster_id_;
  }
  //预加载公共前缀
  void SetPrefixId(const uint64_t prefix_id) {
    prefix_id_ = prefix_id;
  }
  //释放公共前缀
  uint64_t GetPrefixId() const {
    return prefix_id_;
  }

  void SetSequenceLen(const uint64_t sequence_length) {
    sequence_length_ = sequence_length;
  }

  uint64_t GetSequenceLen() const {
    return sequence_length_;
  }

 private:
  uint64_t req_id_{kInvalidReqId};
  // 请求Prompt的句子长度，做完padding的值， 用于申请Prompt的KV Cache
  uint64_t prompt_length_{0UL};
  uint64_t prompt_cluster_id_{0UL};
  uint64_t decoder_cluster_id_{0UL};
  uint64_t prefix_id_{kInvalidPrefixId};
  uint64_t sequence_length_{0UL};
  int8_t reserved_[120];
}