|
| 1 | +# 该文件用于提供在数据dp并行的推理模式下，共享kv cache trans相关的功能函数模块 |
| 2 | +import numpy as np |
| 3 | +import dataclasses |
| 4 | +import torch |
| 5 | +from typing import List |
| 6 | +from lightllm.common.mem_manager import MemoryManager |
| 7 | +from lightllm.utils.envs_utils import get_unique_server_name, get_env_start_args |
| 8 | +from lightllm.utils.dist_utils import get_dp_rank_in_node |
| 9 | +from lightllm.server.core.objs.shm_array import ShmArray |
| 10 | +from ...infer_batch import InferReq |
| 11 | + |
| 12 | + |
| 13 | +class DPKVSharedMoudle: |
| 14 | + _KV_LEN_INDEX = 0 |
| 15 | + _REQ_IDX_INDEX = 1 |
| 16 | + |
| 17 | + def __init__(self, max_req_num: int, max_req_seq_len: int, dp_size_in_node: int, backend): |
| 18 | + from .impl import DPChunkedPrefillBackend |
| 19 | + |
| 20 | + self.backend: DPChunkedPrefillBackend = backend |
| 21 | + self.max_req_num = max_req_num |
| 22 | + self.max_req_seq_len = max_req_seq_len |
| 23 | + |
| 24 | + # 0 代表 kv_len, 1 代表 radix_cache_len |
| 25 | + self.shared_req_infos = ShmArray( |
| 26 | + name=f"{get_unique_server_name()}_dp_shared_req_infos", |
| 27 | + shape=(self.max_req_num, dp_size_in_node, 2), |
| 28 | + dtype=np.int64, |
| 29 | + ) |
| 30 | + self.shared_req_infos.create_shm() |
| 31 | + self.dp_rank_in_node = get_dp_rank_in_node() |
| 32 | + assert get_env_start_args().diverse_mode is False |
| 33 | + |
| 34 | + def fill_reqs_info( |
| 35 | + self, |
| 36 | + reqs: List[InferReq], |
| 37 | + req_dp_ranks: List[int], |
| 38 | + ): |
| 39 | + """ |
| 40 | + 填充请求的 kv 信息到共享内存中 |
| 41 | + """ |
| 42 | + self.backend.node_nccl_group.barrier() |
| 43 | + self.shared_req_infos.arr[0 : len(reqs), self.dp_rank_in_node, self._KV_LEN_INDEX] = [ |
| 44 | + req.cur_kv_len for req in reqs |
| 45 | + ] |
| 46 | + self.shared_req_infos.arr[0 : len(reqs), self.dp_rank_in_node, self._REQ_IDX_INDEX] = [ |
| 47 | + req.req_idx for req in reqs |
| 48 | + ] |
| 49 | + return |
| 50 | + |
| 51 | + def build_shared_kv_trans_tasks( |
| 52 | + self, |
| 53 | + reqs: List[InferReq], |
| 54 | + req_dp_ranks: List[int], |
| 55 | + ) -> List["TransTask"]: |
| 56 | + """ |
| 57 | + 构建共享kv交换信息 |
| 58 | + """ |
| 59 | + from lightllm.server.router.model_infer.infer_batch import g_infer_context |
| 60 | + |
| 61 | + self.backend.node_nccl_group.barrier() |
| 62 | + |
| 63 | + trans_tasks: List[TransTask] = [] |
| 64 | + rank_max_radix_cache_lens = np.max( |
| 65 | + self.shared_req_infos.arr[0 : len(reqs), :, self._KV_LEN_INDEX], axis=1, keepdims=False |
| 66 | + ) |
| 67 | + # 如果发现自己是dp_rank 最小， radix_cache_len 最长的请求，则将数据写入到共享内存中。 |
| 68 | + for req_index, req, max_req_radix_cache_len, req_dp_rank in zip( |
| 69 | + list(range(len(reqs))), reqs, rank_max_radix_cache_lens, req_dp_ranks |
| 70 | + ): |
| 71 | + # 当前请求是本 dp_rank 负责的 |
| 72 | + is_current_dp_handle = req_dp_rank == self.dp_rank_in_node |
| 73 | + trans_size = max_req_radix_cache_len - req.cur_kv_len |
| 74 | + |
| 75 | + if is_current_dp_handle and trans_size > 0 and g_infer_context.get_can_alloc_token_num() > trans_size: |
| 76 | + g_infer_context.radix_cache.free_radix_cache_to_get_enough_token(trans_size) |
| 77 | + mem_indexes = self.backend.model.mem_manager.alloc(trans_size) |
| 78 | + max_kv_len_dp_rank = self.shared_req_infos.arr[req_index, :, self._KV_LEN_INDEX].argmax() |
| 79 | + max_kv_len_req_idx = int(self.shared_req_infos.arr[req_index, max_kv_len_dp_rank, self._REQ_IDX_INDEX]) |
| 80 | + max_kv_len_mem_manager_index = ( |
| 81 | + max_kv_len_dp_rank * self.backend.dp_world_size + self.backend.dp_rank_in_node |
| 82 | + ) |
| 83 | + max_kv_len_mem_manager: MemoryManager = self.backend.mem_managers[max_kv_len_mem_manager_index] |
| 84 | + max_kv_len_mem_indexes = max_kv_len_mem_manager.req_to_token_indexs[ |
| 85 | + max_kv_len_req_idx, req.cur_kv_len : max_req_radix_cache_len |
| 86 | + ] |
| 87 | + trans_tasks.append( |
| 88 | + TransTask( |
| 89 | + req=req, |
| 90 | + mem_indexes=mem_indexes, |
| 91 | + max_kv_len_dp_rank=int(max_kv_len_dp_rank), |
| 92 | + max_kv_len_mem_manager_index=int(max_kv_len_mem_manager_index), |
| 93 | + max_kv_len_mem_indexes=max_kv_len_mem_indexes, |
| 94 | + ) |
| 95 | + ) |
| 96 | + |
| 97 | + return trans_tasks |
| 98 | + |
| 99 | + def kv_trans(self, trans_tasks: List["TransTask"]): |
| 100 | + from lightllm.server.router.model_infer.infer_batch import g_infer_context |
| 101 | + |
| 102 | + self.backend.node_nccl_group.barrier() |
| 103 | + # kv 传输 |
| 104 | + |
| 105 | + # move_token_indexes = torch.tensor(move_token_indexes, dtype=torch.int64, device="cuda") |
| 106 | + # token_dp_indexes = torch.tensor(token_dp_indexes, dtype=torch.int32, device="cuda") |
| 107 | + |
| 108 | + # self.model.mem_manager.copy_kv_from_other_dp_ranks( |
| 109 | + # mem_managers=self.mem_managers, |
| 110 | + # move_token_indexes=move_token_indexes, |
| 111 | + # token_dp_indexes=token_dp_indexes, |
| 112 | + # mem_indexes=mem_indexes, |
| 113 | + # dp_size_in_node=self.dp_size_in_node, |
| 114 | + # rank_in_dp=self.rank_in_dp, |
| 115 | + # ) |
| 116 | + # self.logger.info(f"dp_i {self.dp_rank_in_node} transfer kv tokens num: {alloc_size}") |
| 117 | + |
| 118 | + self.backend.node_nccl_group.barrier() |
| 119 | + for trans_task in trans_tasks: |
| 120 | + g_infer_context.req_manager.req_to_token_indexs[ |
| 121 | + trans_task.req.req_idx, |
| 122 | + trans_task.req.cur_kv_len : (trans_task.req.cur_kv_len + len(trans_task.mem_indexes)), |
| 123 | + ] = trans_task.mem_indexes |
| 124 | + trans_task.req.cur_kv_len += len(trans_task.mem_indexes) |
| 125 | + if self.backend.is_master_in_dp: |
| 126 | + trans_task.req.shm_req.shm_cur_kv_len = trans_task.req.cur_kv_len |
| 127 | + self.backend.node_nccl_group.barrier() |
| 128 | + |
| 129 | + |
| 130 | +@dataclasses |
| 131 | +class TransTask: |
| 132 | + req: InferReq |
| 133 | + mem_indexes: torch.Tensor |
| 134 | + max_kv_len_dp_rank: int |
| 135 | + max_kv_len_mem_manager_index: int |
| 136 | + max_kv_len_mem_indexes: torch.Tensor |
0 commit comments