|
| 1 | +from typing import Set, Protocol, List |
| 2 | + |
| 3 | +import torch |
| 4 | +from sortedcontainers import SortedSet |
| 5 | + |
| 6 | +from lightllm.server.router.dynamic_prompt.radix_cache import RadixCache, TreeNode |
| 7 | +from lightllm.common.kv_cache_mem_manager.mem_manager import MemoryManager |
| 8 | +from lightllm.server.router.model_infer.infer_batch import InferReq |
| 9 | + |
| 10 | + |
| 11 | +class HybridMemManager(MemoryManager): |
| 12 | + def alloc_buffer(self, need_size): |
| 13 | + ... |
| 14 | + |
| 15 | + def free_buffer(self, free_buffer_indexes): |
| 16 | + ... |
| 17 | + |
| 18 | + def get_buffer(self, layer_index): |
| 19 | + ... |
| 20 | + |
| 21 | + def get_buffer_can_use_size(self): |
| 22 | + ... |
| 23 | + |
| 24 | + def copy_buffer(self, src_idx, tgt_idx): |
| 25 | + ... |
| 26 | + |
| 27 | + |
| 28 | +class HybridRadixCache(RadixCache): |
| 29 | + def __init__(self, unique_name, total_token_num, rank_in_node, mem_manager=None): |
| 30 | + self.mem_manager: HybridMemManager = mem_manager |
| 31 | + super().__init__(unique_name, total_token_num, rank_in_node, mem_manager) |
| 32 | + self.evict_buffer_set: Set[TreeNode] = SortedSet(key=lambda x: x.time_id) |
| 33 | + self.evict_buffer_set.add(self.root_node) |
| 34 | + |
| 35 | + def free_radix_cache_to_get_enough_buffer(self, need_buffer_num): |
| 36 | + if need_buffer_num > self.mem_manager.get_buffer_can_use_size(): |
| 37 | + need_evict_buffer_num = need_buffer_num - self.mem_manager.get_buffer_can_use_size() |
| 38 | + |
| 39 | + release_mems = [] |
| 40 | + |
| 41 | + def release_mem(mem_index): |
| 42 | + release_mems.append(mem_index) |
| 43 | + return |
| 44 | + |
| 45 | + release_buffers = [] |
| 46 | + |
| 47 | + def release_buffer(buffer_idx): |
| 48 | + release_buffers.append(buffer_idx) |
| 49 | + return |
| 50 | + |
| 51 | + self.evict_buffer(need_evict_buffer_num, release_buffer, release_mem) |
| 52 | + self.mem_manager.free_buffer(release_buffers) |
| 53 | + if len(release_mems) > 0: |
| 54 | + mem_index = torch.concat(release_mems) |
| 55 | + self.mem_manager.free(mem_index) |
| 56 | + return |
| 57 | + |
| 58 | + def evict_buffer(self, need_evict_buffer_num, evict_buffer_callback, evict_token_callback): |
| 59 | + while need_evict_buffer_num > 0: |
| 60 | + node = self.evict_buffer_set.pop() |
| 61 | + if node.buffer_idx is not None: |
| 62 | + evict_buffer_callback(node.buffer_idx) |
| 63 | + need_evict_buffer_num -= 1 |
| 64 | + else: |
| 65 | + # 在混合注意力模型的情景里，只能匹配 buffer_idx 不为 None的节点 |
| 66 | + # 假如 buffer_idx 为 None，则当做匹配失败。 |
| 67 | + # 所以可以直接把这个节点给释放掉 |
| 68 | + if node.is_leaf() and node.ref_counter == 0: |
| 69 | + self._remove_leaf_node(node) |
| 70 | + return |
| 71 | + |
| 72 | + def insert_for_hybrid_radix_cache(self, reqs: List["InferReq"]): |
| 73 | + from lightllm.server.router.model_infer.infer_batch import g_infer_context |
| 74 | + from lightllm.common.basemodel.infer_lock import g_infer_state_lock |
| 75 | + |
| 76 | + # 确保有足够的空间用于新的 buffer |
| 77 | + g_infer_state_lock.acquire() |
| 78 | + self.free_radix_cache_to_get_enough_buffer(len(reqs)) |
| 79 | + new_buffer_indexes = self.mem_manager.alloc_buffer(len(reqs)) |
| 80 | + g_infer_state_lock.release() |
| 81 | + |
| 82 | + for i, req in enumerate(reqs): |
| 83 | + input_token_ids = req.get_input_token_ids() |
| 84 | + key = torch.tensor(input_token_ids[0 : req.cur_kv_len], dtype=torch.int64, device="cpu") |
| 85 | + value = g_infer_context.req_manager.req_to_token_indexs[req.req_idx][: req.cur_kv_len].cpu() |
| 86 | + buffer_idx = req.buffer_idx |
| 87 | + |
| 88 | + # 分配新的 buffer 并复制当前 buffer 的内容 |
| 89 | + self.mem_manager.copy_buffer(buffer_idx, new_buffer_indexes[i]) |
| 90 | + req.buffer_idx = new_buffer_indexes[i] |
| 91 | + |
| 92 | + _, new_shared_kv_node = self.insert(key, value) |
| 93 | + new_shared_kv_node.buffer_idx = buffer_idx |
| 94 | + self.dec_node_ref_counter(req.shared_kv_node) |
| 95 | + self.add_node_ref_counter(new_shared_kv_node) |
| 96 | + req.shared_kv_node = new_shared_kv_node |
| 97 | + |
| 98 | + def match_prefix(self, key, update_refs=False): |
| 99 | + assert len(key) != 0 |
| 100 | + ans_value_list = [] |
| 101 | + tree_node = self._match_prefix_helper(self.root_node, key, ans_value_list, update_refs=update_refs) |
| 102 | + |
| 103 | + while tree_node != self.root_node and tree_node.buffer_idx is None: |
| 104 | + self.dec_node_ref_counter(tree_node) |
| 105 | + if tree_node.is_leaf() and tree_node.ref_counter == 0: |
| 106 | + tree_node = self._remove_leaf_node(tree_node) |
| 107 | + else: |
| 108 | + tree_node = tree_node.parent |
| 109 | + ans_value_list.pop() |
| 110 | + |
| 111 | + if tree_node == self.root_node: |
| 112 | + return None, 0, None |
| 113 | + |
| 114 | + value = torch.concat(ans_value_list) |
| 115 | + return tree_node, len(value), value |
| 116 | + |
| 117 | + def _remove_leaf_node(self, node: TreeNode): |
| 118 | + self.evict_tree_set.discard(node) |
| 119 | + self.tree_total_tokens_num.arr[0] -= len(node.token_mem_index_value) |
| 120 | + parent_node: TreeNode = node.parent |
| 121 | + parent_node.remove_child(node) |
| 122 | + if parent_node.is_leaf(): |
| 123 | + self.evict_tree_set.add(parent_node) |
| 124 | + return parent_node |
0 commit comments