* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#ifndef BUILT_IN_LLM_COMMON_KV_CACHE_MANAGER_H
#define BUILT_IN_LLM_COMMON_KV_CACHE_MANAGER_H
#include <mutex>
#include <unordered_map>
#include "fsm/state_define.h"
#include "flow_func/flow_msg.h"
#include "flow_func/meta_run_context.h"
#include "llm_common/llm_common.h"
namespace FlowFunc {
constexpr uint64_t kHashShiftLeft = 32UL;
class KvCacheManager {
public:
struct ReqKey {
uint64_t req_id;
uint64_t model_id;
bool operator==(const ReqKey &req_key) const {
return ((req_id == req_key.req_id) && (model_id == req_key.model_id));
}
};
struct ReqKeyHash {
std::size_t operator()(const ReqKey &req_key) const {
std::hash<uint64_t> hash_fn;
return hash_fn(req_key.req_id) ^ hash_fn(req_key.model_id << kHashShiftLeft);
}
};
struct PrefixReqKey {
uint64_t prefix_id;
uint64_t model_id;
bool operator==(const PrefixReqKey &req_key) const {
return ((prefix_id == req_key.prefix_id) && (model_id == req_key.model_id));
}
};
struct PrefixReqKeyHash {
std::size_t operator()(const PrefixReqKey &req_key) const {
std::hash<uint64_t> hash_fn;
return hash_fn(req_key.prefix_id) ^ hash_fn(req_key.model_id << kHashShiftLeft);
}
};
static KvCacheManager &GetInstance();
virtual ~KvCacheManager();
std::shared_ptr<FlowMsg> QueryDecoderKvCache(const ReqKey &req_key);
FsmStatus SaveDecoderKvCache(const ReqKey &req_key, const std::shared_ptr<FlowMsg> &kv_cache);
FsmStatus ReleaseDecoderKvCache(const ReqKey &req_key);
FsmStatus ReleaseDecoderKvCaches();
void SetDecoderKvCache(const std::vector<std::shared_ptr<FlowMsg>> &kv_caches);
std::shared_ptr<FlowMsg> GetDecoderKvCache(uint64_t model_id) const;
void ClearDecoderKvCache();
uint64_t GetPromptKvCacheCount();
uint64_t GetDecoderKvCacheCount(const ReqKey &req_key);
std::vector<KvTensor> QueryPromptKvCache(const ReqKey &req_key);
std::vector<KvTensor> QueryPromptPrefixKvCache(const PrefixReqKey &prefix_req_key);
FsmStatus ShrinkKvCacheForPrefix(const PrefixReqKey &prefix_req_key,
const std::shared_ptr<MetaRunContext> &run_context);
FsmStatus ReleaseKvCacheForPrompt(const ReqKey &req_key);
FsmStatus ReleaseKvCacheForPrompt(uint64_t req_id);
FsmStatus ReleaseKvCacheForPrefix(const PrefixReqKey &prefix_req_key);
uint64_t AddKvCache(size_t batch_num, std::vector<std::shared_ptr<FlowMsg>> kv_tensors);
void SaveKvCacheForPrompt(const ReqKey &req_key, uint64_t cache_id, std::pair<int64_t, int64_t> offset_and_size);
void SaveKvCacheForPromptPrefix(const PrefixReqKey &prefix_req_key,
uint64_t cache_id,
std::pair<int64_t, int64_t> offset_and_size);
void InitBlockTensors(uint64_t model_id);
void SaveKvBlocksTensor(uint64_t model_id, std::shared_ptr<FlowMsg> &kv_blocks_tensor);
std::vector<std::shared_ptr<FlowMsg>> &QueryKvBlocksTensors(uint64_t model_id);
void ClearKvBlocksTensors();
void ClearPromptKvCache();
KvCacheManager(const KvCacheManager &) = delete;
KvCacheManager(const KvCacheManager &&) = delete;
KvCacheManager &operator=(const KvCacheManager &) = delete;
KvCacheManager &operator=(const KvCacheManager &&) = delete;
private:
KvCacheManager() = default;
std::vector<KvTensor> QueryKvTensor(uint64_t cache_id, uint64_t tensor_id);
FsmStatus ReleaseKvCacheForPromptByKey(const ReqKey &req_key);
private:
struct CacheEntry {
size_t batch_num;
std::vector<std::shared_ptr<FlowMsg>> kv_tensors;
std::unordered_map<uint64_t, std::pair<int64_t, int64_t>> id_to_offset_and_size;
};
std::vector<std::shared_ptr<FlowMsg>> decoder_kv_caches_;
std::mutex mutex_;
std::unordered_map<ReqKey, std::shared_ptr<FlowMsg>, ReqKeyHash> decoder_req_to_kv_cache_;
std::vector<std::vector<std::shared_ptr<FlowMsg>>> kv_blocks_tensors_;
uint64_t cache_id_gen_{0};
std::unordered_map<uint64_t, CacheEntry> cache_id_to_cache_entry_;
std::unordered_map<ReqKey, uint64_t, ReqKeyHash> req_id_to_cache_id_;
std::unordered_map<PrefixReqKey, uint64_t, PrefixReqKeyHash> prefix_id_to_cache_id_;
};
}
#endif