""" # Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License" # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ import argparse import concurrent.futures import gc import json import queue import threading import time import traceback from typing import List import numpy as np import paddle from fastdeploy import envs from fastdeploy.cache_manager.cache_data import CacheStatus from fastdeploy.cache_manager.ops import ( cuda_host_alloc, cuda_host_free, memory_allocated, set_data_ipc, set_device, share_external_data_, swap_cache_all_layers, swap_cache_layout, unset_data_ipc, ) from fastdeploy.cache_manager.transfer_factory import MooncakeStore from fastdeploy.config import SpeculativeConfig from fastdeploy.inter_communicator import EngineCacheQueue, IPCSignal, KVCacheStatus from fastdeploy.platforms import current_platform from fastdeploy.utils import get_logger def parse_args(): """ 从命令行解析参数 """ parser = argparse.ArgumentParser("Cache transfer manager") parser.add_argument( "--splitwise_role", type=str, default="mixed", help="splitwise role, can be decode, prefill or mixed", ) parser.add_argument("--rank", type=int, default=0, help="local tp rank") parser.add_argument("--device_id", type=int, default=0, help="device id") parser.add_argument("--max_model_len", type=int, default=32768, help="max model length") parser.add_argument("--num_layers", type=int, default=1, help="model num layers") parser.add_argument("--mp_num", type=int, default=1, help="number of model parallel") parser.add_argument( "--cache_dtype", type=str, default="bfloat16", choices=["uint8", "bfloat16", "block_wise_fp8"], help="cache dtype", ) parser.add_argument( "--default_dtype", type=str, default="bfloat16", choices=["float16", "bfloat16", "uint8"], help="paddle default dtype, swap_cache_batch only support float16、bfloat16 and uint8 now", ) parser.add_argument("--key_cache_shape", type=str, default="", help="key cache shape") parser.add_argument("--value_cache_shape", type=str, default="", help="value cache shape") parser.add_argument("--cache_queue_port", type=int, default=9923, help="cache queue port") parser.add_argument("--enable_splitwise", type=int, default=0, help="enable splitwise ") parser.add_argument("--pod_ip", type=str, default="0.0.0.0", help="pod ip") parser.add_argument( "--engine_worker_queue_port", type=int, default=9923, help="engine worker queue port", ) parser.add_argument("--num_cpu_blocks", type=int, default=4, help="cpu cache block number") parser.add_argument("--ipc_suffix", type=str, default=None, help="engine pid") parser.add_argument( "--protocol", type=str, default="ipc", help="cache transfer protocol, only support ipc now", ) parser.add_argument("--local_data_parallel_id", type=int, default=0) parser.add_argument("--rdma_port", type=str, default="", help="rmda port") parser.add_argument( "--speculative_config", type=json.loads, default="{}", help="speculative config", ) parser.add_argument("--create_cache_tensor", action="store_true") parser.add_argument( "--kvcache_storage_backend", type=str, default=None, choices=["mooncake", "none"], help="The storage backend for kvcache storage. If not set, storage backend is disabled.", ) parser.add_argument( "--write_policy", type=str, choices=["write_through"], default="write_through", help="KVCache write policy", ) args = parser.parse_args() return args class CacheTransferManager: """ 管理CPU和GPU之间缓存的交换传输 """ def __init__(self, args): """ 初始化CacheTransferManager """ device = args.device_id rank = args.rank self.gpu_cache_kvs = {} self.cpu_cache_kvs = {} self.gpu_cache_k_tensors = [] self.gpu_cache_v_tensors = [] self.gpu_cache_scales_k_tensors = [] self.gpu_cache_scales_v_tensors = [] self.speculative_config = SpeculativeConfig(args.speculative_config) self.key_cache_shape = [int(i) for i in args.key_cache_shape.split(",")] self.value_cache_shape = [] if args.value_cache_shape: self.value_cache_shape = [int(i) for i in args.value_cache_shape.split(",")] self.num_gpu_blocks = self.key_cache_shape[0] self.num_extra_layers = self.speculative_config.num_extra_cache_layer self.num_extra_layer_gpu_blocks = int(self.num_gpu_blocks * self.speculative_config.num_gpu_block_expand_ratio) paddle.set_default_dtype(args.default_dtype) self.swap_to_cpu_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1) self.swap_to_gpu_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1) self.read_storage_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1) self.write_back_storage_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1) self.timeout_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=2) self.transfer_task_queue = queue.Queue() # 用来接收传输任务 self.tansfer_done_queue = queue.Queue() # 用来告知任务执行完毕 self.n_ranks = args.mp_num self.rank = rank self.device = device self.ipc_suffix = args.ipc_suffix self.cache_dtype = args.cache_dtype address = (args.pod_ip, args.cache_queue_port) self.cache_task_queue = EngineCacheQueue( address=address, is_server=False, num_client=args.mp_num, client_id=rank, local_data_parallel_id=args.local_data_parallel_id, ) cache_ready_signal_data = np.zeros(shape=[args.mp_num], dtype=np.int32) self.cache_ready_signal = IPCSignal( name="cache_ready_signal", array=cache_ready_signal_data, dtype=np.int32, suffix=args.engine_worker_queue_port, create=False, ) swap_space_ready_data = np.zeros(shape=[args.mp_num], dtype=np.int32) self.swap_space_ready_signal = IPCSignal( name="swap_space_ready_signal", array=swap_space_ready_data, dtype=np.int32, suffix=args.engine_worker_queue_port, create=False, ) self.num_cpu_blocks = args.num_cpu_blocks self._init_gpu_cache(args) if self.num_cpu_blocks > 0: self._init_cpu_cache(args) cache_task_broadcast_data = np.zeros(shape=[1], dtype=np.int32) self.cache_task_broadcast_signal = IPCSignal( name="cache_task_broadcast_signal", array=cache_task_broadcast_data, dtype=np.int32, suffix=args.engine_worker_queue_port, create=False, ) max_chips_per_node = 16 if current_platform.is_iluvatar() else 8 array_size = min(max_chips_per_node, args.mp_num) worker_healthy_live_array = np.zeros(shape=[array_size], dtype=np.int32) self.worker_healthy_live_signal = IPCSignal( name="worker_healthy_live_signal", array=worker_healthy_live_array, dtype=np.int32, suffix=args.engine_worker_queue_port, create=False, ) if args.kvcache_storage_backend is None or args.kvcache_storage_backend == "none": self.storage_backend = None elif args.kvcache_storage_backend == "mooncake": logger.info("Start initialize mooncake store...") self.storage_backend = MooncakeStore(tp_rank=self.rank) self._init_storage_buffer(args) logger.info("Initialized mooncake store successfully") else: raise NotImplementedError(f"Unsupported storage backend: {args.kvcache_storage_backend}") if args.write_policy not in ["write_through"]: raise ValueError(f"Invalid write policy: {args.write_policy}") self.write_policy = args.write_policy # Initialize update/clear signals for RL self.kv_cache_status_signal = IPCSignal( name="kv_cache_status", array=np.zeros([1], dtype=np.int32), dtype=np.int32, suffix=args.engine_worker_queue_port, create=False, ) threading.Thread(target=self.check_cache_status, args=[args], daemon=True).start() def _init_storage_buffer(self, args): """ Initialize pinned memory buffer that can hold the cache for a longest request cache layout: layer_num * [block_num, head_num, block_size, head_dim] buffer layout: [block_num, layer_num, head_num, block_size, head_dim] """ layer_num = args.num_layers + self.num_extra_layers head_num = self.key_cache_shape[1] block_size = self.key_cache_shape[2] head_dim = self.key_cache_shape[3] block_num = (args.max_model_len + block_size - 1) // block_size logger.info( f"Creating cache buffer for storage with shape: " f"[{block_num}, {layer_num}, {head_num}, {block_size}, {head_dim}]" ) self.cache_bytes = self._get_cache_bytes(self.cache_dtype) self.storage_buffer_stride_bytes = layer_num * head_num * block_size * head_dim * self.cache_bytes total_bytes = block_num * self.storage_buffer_stride_bytes * 2 # key and value logger.info(f"Creating cpu buffer cache for alllayers: {total_bytes / 1024 ** 3:.2f}GB") read_buffer = cuda_host_alloc(total_bytes) self.storage_key_read_buffer = read_buffer self.storage_value_read_buffer = read_buffer + total_bytes // 2 self.storage_backend.register_buffer(read_buffer, total_bytes) write_buffer = cuda_host_alloc(total_bytes) self.storage_key_write_buffer = write_buffer self.storage_value_write_buffer = write_buffer + total_bytes // 2 self.storage_backend.register_buffer(write_buffer, total_bytes) def _init_gpu_cache(self, args): try: assert not args.create_cache_tensor except: logger.warn( f"In current implementation, cache transfer manager do not create cache tensors at all, " f"meaning create_cache_tensor should be False, while we got {args.create_cache_tensor}. " f"Cache tensor creation will occur in: 1) model runner in case of mixed deployment; " f"or 2) cache messager in case of disaggregation deployment. " f"Please check the codes and make sure they work correctly." ) if not args.create_cache_tensor: logger.info(f"[rank {self.rank}/{self.n_ranks}] Waiting for runners or messagers to create kv cache.") while self.cache_ready_signal.value[self.rank] != 1: time.sleep(0.1) logger.info(f"[rank {self.rank}/{self.n_ranks}] OK! Stop waiting.") if args.cache_dtype == "block_wise_fp8": cache_type = "uint8" else: cache_type = args.cache_dtype logger.info(f"[rank {self.rank}/{self.n_ranks}] Initializing kv cache for all layers.") set_device(self.device) for i in range(args.num_layers + self.num_extra_layers): num_gpu_blocks = self.num_gpu_blocks if i < args.num_layers else self.num_extra_layer_gpu_blocks key_name = f"key_caches_{i}_rank{self.rank}.device{self.device}" val_name = f"value_caches_{i}_rank{self.rank}.device{self.device}" key_cache_scales_name = f"key_cache_scales_{i}_rank{self.rank}.device{self.device}" value_cache_scales_name = f"value_cache_scales_{i}_rank{self.rank}.device{self.device}" key_cache_shape = [ num_gpu_blocks, self.key_cache_shape[1], self.key_cache_shape[2], self.key_cache_shape[3], ] value_cache_shape = [] if self.value_cache_shape: value_cache_shape = [ num_gpu_blocks, self.value_cache_shape[1], self.value_cache_shape[2], self.value_cache_shape[3], ] if args.create_cache_tensor: logger.info( f"[rank {self.rank}/{self.n_ranks}] ..creating kv cache for layer {i}: {key_cache_shape} {value_cache_shape}" ) key_cache = paddle.full(shape=key_cache_shape, fill_value=0, dtype=cache_type) set_data_ipc(key_cache, key_name) if args.cache_dtype == "block_wise_fp8": key_cache_scales = paddle.full( shape=[num_gpu_blocks, self.key_cache_shape[1], self.key_cache_shape[2]], fill_value=0, dtype=paddle.get_default_dtype(), ) set_data_ipc(key_cache_scales, key_cache_scales_name) if self.value_cache_shape: val_cache = paddle.full(shape=value_cache_shape, fill_value=0, dtype=cache_type) set_data_ipc(val_cache, val_name) if args.cache_dtype == "block_wise_fp8": value_cache_scales = paddle.full( shape=[num_gpu_blocks, self.value_cache_shape[1], self.value_cache_shape[2]], fill_value=0, dtype=paddle.get_default_dtype(), ) set_data_ipc(value_cache_scales, value_cache_scales_name) else: logger.info( f"[rank {self.rank}/{self.n_ranks}] ..attaching kv cache for layer {i}: {key_cache_shape} {value_cache_shape}" ) key_cache = paddle.empty(shape=[], dtype=cache_type) val_cache = paddle.empty(shape=[], dtype=cache_type) key_cache = share_external_data_(key_cache, key_name, key_cache_shape, True) if args.cache_dtype == "block_wise_fp8": key_cache_scales = paddle.empty(shape=[], dtype=paddle.get_default_dtype()) key_cache_scales = share_external_data_( key_cache_scales, key_cache_scales_name, [num_gpu_blocks, self.key_cache_shape[1], self.key_cache_shape[2]], True, ) if self.value_cache_shape: val_cache = share_external_data_(val_cache, val_name, value_cache_shape, True) if args.cache_dtype == "block_wise_fp8": value_cache_scales = paddle.empty(shape=[], dtype=paddle.get_default_dtype()) value_cache_scales = share_external_data_( value_cache_scales, value_cache_scales_name, [num_gpu_blocks, self.value_cache_shape[1], self.value_cache_shape[2]], True, ) self.gpu_cache_kvs[key_name] = key_cache self.gpu_cache_k_tensors.append(self.gpu_cache_kvs[key_name]) if args.cache_dtype == "block_wise_fp8": self.gpu_cache_kvs[key_cache_scales_name] = key_cache_scales self.gpu_cache_scales_k_tensors.append(self.gpu_cache_kvs[key_cache_scales_name]) if args.value_cache_shape: self.gpu_cache_kvs[val_name] = val_cache self.gpu_cache_v_tensors.append(self.gpu_cache_kvs[val_name]) if args.cache_dtype == "block_wise_fp8": self.gpu_cache_kvs[value_cache_scales_name] = value_cache_scales self.gpu_cache_scales_v_tensors.append(self.gpu_cache_kvs[value_cache_scales_name]) if args.create_cache_tensor: logger.info(f"[rank {self.rank}/{self.n_ranks}] ✅ kv cache is ready!") self.cache_ready_signal.value[self.rank] = 1 cache_kv_size_byte = sum([tmp.numel() * 1 for key, tmp in self.gpu_cache_kvs.items()]) logger.info(f"[rank {self.rank}/{self.n_ranks}] device :{self.device}") logger.info(f"[rank {self.rank}/{self.n_ranks}] cache_kv_size_byte : {cache_kv_size_byte}") logger.info(f"[rank {self.rank}/{self.n_ranks}] done init cache (full) gmem alloc : {memory_allocated()}") def _init_cpu_cache(self, args): key_cache_size = self.key_cache_shape[1] * self.key_cache_shape[2] * self.key_cache_shape[3] if args.value_cache_shape: value_cache_size = self.value_cache_shape[1] * self.value_cache_shape[2] * self.value_cache_shape[3] else: value_cache_size = 0 cache_bytes = self._get_cache_bytes(self.cache_dtype) key_need_to_allocate_bytes = args.num_cpu_blocks * cache_bytes * key_cache_size value_need_to_allocate_bytes = args.num_cpu_blocks * cache_bytes * value_cache_size if args.cache_dtype == "block_wise_fp8": cache_scales = paddle.empty(shape=[], dtype=paddle.get_default_dtype()) cache_scales_size = self.key_cache_shape[1] * self.key_cache_shape[2] scales_key_need_to_allocate_bytes = args.num_cpu_blocks * cache_scales.element_size() * cache_scales_size scales_value_need_to_allocate_bytes = args.num_cpu_blocks * cache_scales.element_size() * cache_scales_size logger.info( f"[rank {self.rank}/{self.n_ranks}] ..swap space size : {(key_need_to_allocate_bytes + value_need_to_allocate_bytes) / 1024 ** 3:.2f}GB" ) if args.num_cpu_blocks == 0: logger.info(f"[rank {self.rank}/{self.n_ranks}] 💡 no swap space (cpu cache) is specified.") self.swap_space_ready_signal.value[self.rank] = 1 return logger.info(f"[rank {self.rank}/{self.n_ranks}] Initializing swap space (cpu cache) for all layers.") paddle.set_device("cpu") self.k_dst_ptrs = [] self.v_dst_ptrs = [] self.k_scales_ptrs = [] self.v_scales_ptrs = [] for i in range(args.num_layers + self.num_extra_layers): key_name = f"key_caches_{i}_rank{self.rank}" val_name = f"value_caches_{i}_rank{self.rank}" key_cache_scales_name = f"key_cache_scales_{i}_rank{self.rank}" value_cache_scales_name = f"value_cache_scales_{i}_rank{self.rank}" logger.info( f"[rank {self.rank}/{self.n_ranks}] ..creating cpu cache for layer {i}: {(key_need_to_allocate_bytes + value_need_to_allocate_bytes) / 1024 ** 3:.2f}GB" ) self.cpu_cache_kvs[key_name] = cuda_host_alloc(key_need_to_allocate_bytes) self.k_dst_ptrs.append(self.cpu_cache_kvs[key_name]) if args.cache_dtype == "block_wise_fp8": self.cpu_cache_kvs[key_cache_scales_name] = cuda_host_alloc(scales_key_need_to_allocate_bytes) self.k_scales_ptrs.append(self.cpu_cache_kvs[key_cache_scales_name]) if value_need_to_allocate_bytes > 0: self.cpu_cache_kvs[val_name] = cuda_host_alloc(value_need_to_allocate_bytes) self.v_dst_ptrs.append(self.cpu_cache_kvs[val_name]) if args.cache_dtype == "block_wise_fp8": self.cpu_cache_kvs[value_cache_scales_name] = cuda_host_alloc(scales_value_need_to_allocate_bytes) self.v_scales_ptrs.append(self.cpu_cache_kvs[value_cache_scales_name]) logger.info(f"[rank {self.rank}/{self.n_ranks}] ✅ swap space (cpu cache) is ready!") self.swap_space_ready_signal.value[self.rank] = 1 def _get_cache_bytes(self, cache_dtype): if cache_dtype == "bfloat16": cache_bytes = 2 elif cache_dtype in ["uint8", "block_wise_fp8"]: cache_bytes = 1 else: raise ValueError(f"Unsupported cache dtype: {cache_dtype}") return cache_bytes def _storage_exist_block_num(self, k_keys: List[str], v_keys: List[str]): """ Given the k_keys and v_keys, get the valid blocks number that can be prefetched from storage backend. """ assert len(k_keys) == len(v_keys), "k_keys and v_keys must have the same length." result = self.storage_backend.exists(k_keys + v_keys) # only consider the case when both key and value exist num = 0 for k, v in zip(k_keys, v_keys): if result[k] and result[v]: num += 1 return num def _run_read_storage(self, k_cache_keys, v_cache_keys, gpu_block_ids, cpu_block_ids): try: logger.debug( f"_run_read_storage, key_hash_keys_num: {len(k_cache_keys)}, " f"value_hash_keys_num: {len(v_cache_keys)}, gpu_block_ids_num: {len(gpu_block_ids)}, " f"cpu_block_ids_num: {len(cpu_block_ids)}" ) block_num = len(gpu_block_ids) keys = k_cache_keys + v_cache_keys k_cache_ptrs = [self.storage_key_read_buffer + i * self.storage_buffer_stride_bytes for i in cpu_block_ids] v_cache_ptrs = [ self.storage_value_read_buffer + i * self.storage_buffer_stride_bytes for i in cpu_block_ids ] kv_cache_ptrs = k_cache_ptrs + v_cache_ptrs kv_block_sizes = [self.storage_buffer_stride_bytes] * block_num * 2 # key and value start_time = time.time() result = self.storage_backend.batch_get(keys, target_locations=kv_cache_ptrs, target_sizes=kv_block_sizes) read_cost_time = time.time() - start_time k_result, v_result = result[:block_num], result[block_num:] success_block_num = 0 for k, v in zip(k_result, v_result): if k > 0 and v > 0: success_block_num += 1 logger.debug(f"_run_read_storage, success_block_num: {success_block_num}") valid_gpu_block_ids = gpu_block_ids[:success_block_num] valid_cpu_block_ids = cpu_block_ids[:success_block_num] mode = 1 # cpu ==> gpu start_time = time.time() swap_cache_layout( self.gpu_cache_k_tensors, self.storage_key_read_buffer, self.key_cache_shape, valid_gpu_block_ids, valid_cpu_block_ids, self.device, mode, ) swap_cache_layout( self.gpu_cache_v_tensors, self.storage_value_read_buffer, self.value_cache_shape, valid_gpu_block_ids, valid_cpu_block_ids, self.device, mode, ) swap_cost_time = time.time() - start_time logger.debug( f"_run_read_storage, swap_cost_time: {swap_cost_time:.6f}s, read_cost_time: {read_cost_time:.6f}s" ) return valid_gpu_block_ids except Exception as e: logger.error( f"[rank {self.rank}/{self.n_ranks}] An error occurred in _run_read_storage: " f"error:{e}, {traceback.format_exc()}" ) raise def read_storage_task(self, task_id, keys, gpu_block_ids, timeout=0.1): """Read cache from the storage backend to the GPU memory.""" try: logger.debug( f"read_storage_task, task id: {task_id}, hash_keys_num: {len(keys)}, " f"gpu_block_ids_num: {len(gpu_block_ids)}, timeout: {timeout}" ) k_cache_keys = [f"{key}_key_{self.rank}" for key in keys] v_cache_keys = [f"{key}_value_{self.rank}" for key in keys] match_block_num = self._storage_exist_block_num(k_cache_keys, v_cache_keys) logger.debug(f"read_storage_task, match {match_block_num} blocks from storage for task id: {task_id}") k_cache_keys = k_cache_keys[:match_block_num] v_cache_keys = v_cache_keys[:match_block_num] gpu_block_ids = gpu_block_ids[:match_block_num] cpu_block_ids = [i for i in range(match_block_num)] valid_gpu_block_ids = [] if match_block_num > 0: # TODO: support timeout with actual block count try: valid_gpu_block_ids = self._run_read_storage( k_cache_keys, v_cache_keys, gpu_block_ids, cpu_block_ids ) logger.info( f"read_storage_task, finish loading {match_block_num} blocks from storage for task {task_id}." ) except Exception as e: logger.error(f"[rank {self.rank}/{self.n_ranks}] An error occurred: {task_id} {e}") valid_gpu_block_ids = [] result = (CacheStatus.STORAGE2GPU, task_id, keys, valid_gpu_block_ids) self.cache_task_queue.swap_storage_to_gpu_barrier.wait() self.cache_task_queue.swap_storage_to_gpu_barrier.reset() self.cache_task_queue.put_transfer_done_signal(result) logger.debug(f"read_storage_task: put_transfer_done_signal {result}") logger.info( f"read_storage_task: put_transfer_done_signal for transfer_task_id {task_id}, " f"valid block num {len(valid_gpu_block_ids)}" ) except Exception as e: logger.error( f"[rank {self.rank}/{self.n_ranks}] An error occurred in read_storage_task: " f"task_id: {task_id}, error:{e}, {traceback.format_exc()}" ) def _run_write_back_storage(self, k_cache_keys, v_cache_keys, gpu_block_ids, cpu_block_ids): try: logger.debug( f"_run_write_back_storage, k_cache_keys: {k_cache_keys}, v_cache_keys: {v_cache_keys}, " f"gpu_block_ids: {gpu_block_ids}" ) key_cache_size = [ self.key_cache_shape[0], self.key_cache_shape[1], self.key_cache_shape[2], self.key_cache_shape[3], ] mode = 0 # gpu ==> cpu start_time = time.time() swap_cache_layout( self.gpu_cache_k_tensors, self.storage_key_write_buffer, key_cache_size, gpu_block_ids, cpu_block_ids, self.device, mode, ) swap_cache_layout( self.gpu_cache_v_tensors, self.storage_value_write_buffer, key_cache_size, gpu_block_ids, cpu_block_ids, self.device, mode, ) swap_cost_time = time.time() - start_time block_num = len(gpu_block_ids) keys = k_cache_keys + v_cache_keys k_cache_ptrs = [ self.storage_key_write_buffer + i * self.storage_buffer_stride_bytes for i in cpu_block_ids ] v_cache_ptrs = [ self.storage_value_write_buffer + i * self.storage_buffer_stride_bytes for i in cpu_block_ids ] kv_cache_ptrs = k_cache_ptrs + v_cache_ptrs kv_block_sizes = [self.storage_buffer_stride_bytes] * block_num * 2 # key and value start_time = time.time() self.storage_backend.batch_set(keys, target_locations=kv_cache_ptrs, target_sizes=kv_block_sizes) write_cost_time = time.time() - start_time logger.debug( f"_run_write_back_storage, swap_cost_time: {swap_cost_time:.6f}s, write_cost_time: {write_cost_time:.6f}s" ) except Exception as e: logger.error( f"[rank {self.rank}/{self.n_ranks}] An error occurred in _run_write_back_storage: " f"error:{e}, {traceback.format_exc()}" ) def write_back_storage_task(self, task_id, keys, gpu_block_ids, timeout=0.1): """ Write cache to the storage backend from the GPU memory. """ try: logger.debug( f"write cache to storage, keys: {keys}, gpu_block_ids: {gpu_block_ids}, " f"task_id: {task_id}, timeout: {timeout}" ) k_cache_keys = [f"{key}_key_{self.rank}" for key in keys] v_cache_keys = [f"{key}_value_{self.rank}" for key in keys] match_block_num = self._storage_exist_block_num(k_cache_keys, v_cache_keys) k_cache_keys = k_cache_keys[match_block_num:] v_cache_keys = v_cache_keys[match_block_num:] gpu_block_ids = gpu_block_ids[match_block_num:] cpu_block_ids = [i for i in range(len(gpu_block_ids))] if len(k_cache_keys) == 0: logger.info(f"No uncached keys found for task {task_id}") gpu_block_ids = [] else: try: # TODO: support timeout with actual block count self._run_write_back_storage(k_cache_keys, v_cache_keys, gpu_block_ids, cpu_block_ids) except Exception as e: logger.error(f"Error in write back storage task: {e}") gpu_block_ids = [] result = (CacheStatus.GPU2STORAGE, task_id, keys, gpu_block_ids) self.cache_task_queue.swap_to_storage_barrier.wait() if self.rank == 0: # 只有当rank为0时执行同步操作 self.cache_task_queue.swap_to_storage_barrier.reset() self.cache_task_queue.put_transfer_done_signal(result) # 发送传输完成信号 logger.debug(f"write_back_storage_task: put_transfer_done_signal {result}") logger.info(f"write_back_storage_task: put_transfer_done_signal for transfer_task_id {task_id}") except Exception as e: logger.error( f"[rank {self.rank}/{self.n_ranks}] An error occurred in write_back_storage_task: " f"error:{e}, {traceback.format_exc()}" ) def _do_swap_to_cpu_task( self, swap_node_ids, gpu_block_id, cpu_block_id, event_type, transfer_task_id, ): """ swap cache GPU->CPU """ self.cache_task_queue.swap_to_cpu_barrier1.wait() if self.rank == 0: self.cache_task_queue.swap_to_cpu_barrier1.reset() result = self._transfer_data( swap_node_ids, gpu_block_id, cpu_block_id, event_type, transfer_task_id, ) self.cache_task_queue.swap_to_cpu_barrier2.wait() if self.rank == 0: self.cache_task_queue.swap_to_cpu_barrier2.reset() self.cache_task_queue.put_transfer_done_signal(result) logger.debug(f"_do_swap_to_cpu_task: put_transfer_done_signal {result}") logger.info(f"_do_swap_to_cpu_task: put_transfer_done_signal for transfer_task_id {transfer_task_id}") def _do_swap_to_gpu_task( self, swap_node_ids, gpu_block_id, cpu_block_id, event_type, transfer_task_id, ): """ swap cache CPU->GPU """ self.cache_task_queue.swap_to_gpu_barrier1.wait() if self.rank == 0: self.cache_task_queue.swap_to_gpu_barrier1.reset() result = self._transfer_data( swap_node_ids, gpu_block_id, cpu_block_id, event_type, transfer_task_id, ) self.cache_task_queue.swap_to_gpu_barrier2.wait() if self.rank == 0: self.cache_task_queue.swap_to_gpu_barrier2.reset() self.cache_task_queue.put_transfer_done_signal(result) logger.debug(f"_do_swap_to_gpu_task: put_transfer_done_signal {result}") logger.info(f"_do_swap_to_gpu_task: put_transfer_done_signal for transfer_task_id {transfer_task_id}") def check_work_status(self, time_interval_threashold=envs.FD_CACHE_PROC_EXIT_TIMEOUT): """ Check the health of the model server by checking whether all workers are alive. """ if self.worker_healthy_live_signal.value[0]: elapsed_time = time.time() - self.worker_healthy_live_signal.value[0] if elapsed_time > time_interval_threashold: return False, "Worker Service Not Healthy" return True, "" def do_data_transfer(self): """ do data transfer task """ consecutive_error_count = 0 max_errors = ( envs.FD_CACHE_PROC_ERROR_COUNT ) # After this many consecutive errors, check if the worker process exists. while True: try: if self.rank == 0: if not self.cache_task_queue.empty(): self.cache_task_broadcast_signal.value[0] = 1 if self.n_ranks > 1: self.cache_task_queue.barrier1.wait() if self.rank == 0: self.cache_task_queue.barrier1.reset() if self.cache_task_broadcast_signal.value[0] == 1: data, read_finish = self.cache_task_queue.get_transfer_task() logger.debug(f"transfer data: get_transfer_task {data}") if read_finish: self.cache_task_broadcast_signal.value[0] = 0 event_type, transfer_task_id = data[0], data[1] if event_type.value == CacheStatus.SWAP2CPU.value: swap_node_ids, gpu_block_id, cpu_block_id = data[2:] self.swap_to_cpu_thread_pool.submit( self._do_swap_to_cpu_task, swap_node_ids, gpu_block_id, cpu_block_id, event_type, transfer_task_id, ) elif event_type.value == CacheStatus.SWAP2GPU.value: swap_node_ids, gpu_block_id, cpu_block_id = data[2:] self.swap_to_gpu_thread_pool.submit( self._do_swap_to_gpu_task, swap_node_ids, gpu_block_id, cpu_block_id, event_type, transfer_task_id, ) elif event_type.value == CacheStatus.STORAGE2GPU.value: hash_keys, gpu_block_ids, timeout = data[2:] self.read_storage_thread_pool.submit( self.read_storage_task, transfer_task_id, hash_keys, gpu_block_ids, timeout, ) elif event_type.value == CacheStatus.GPU2STORAGE.value: hash_keys, gpu_block_ids, timeout = data[2:] self.write_back_storage_thread_pool.submit( self.write_back_storage_task, transfer_task_id, hash_keys, gpu_block_ids, timeout, ) else: if self.n_ranks > 1: self.cache_task_queue.barrier2.wait() if self.rank == 0: self.cache_task_queue.barrier2.reset() continue if self.n_ranks > 1: self.cache_task_queue.barrier3.wait() if self.rank == 0: self.cache_task_queue.barrier3.reset() consecutive_error_count = 0 except (BrokenPipeError, EOFError, ConnectionResetError) as e: # When a cache_transfer_manager process remains, it keeps printing error logs and may exhaust disk space. # Add a check to see if the worker process is alive; if it has ended, exit the loop to stop continuous logging. logger.error(f"[CacheTransferManager] Connection broken: {e}") consecutive_error_count += 1 if consecutive_error_count > max_errors: try: status, msg = self.check_work_status() except Exception: status = True if status is False: logger.critical( f"The Worker process has been inactive for over {envs.FD_CACHE_PROC_EXIT_TIMEOUT} seconds, and the Cache process will automatically terminate (the waiting timeout can be extended via FD_CACHE_PROC_EXIT_TIMEOUT)." ) break time.sleep(1) continue except Exception as e: logger.info(f"do_data_transfer: error: {e}, {str(traceback.format_exc())}") def _transfer_data( self, swap_node_ids, task_gpu_block_id, task_cpu_block_id, event_type, transfer_task_id, ): """ transfer data task_gpu_block_id format: [[block_id0, [fold_block_id0, fold_block_id1]], [block_id1, [fold_block_id0, fold_block_id1]], ...] """ logger.debug( f"transfer data: transfer_task_id {transfer_task_id}: swap_node_ids {swap_node_ids}" + f"task_gpu_block_id {task_gpu_block_id} task_cpu_block_id {task_cpu_block_id} event_type {event_type}" ) start_time = time.time() try: # transform block id assert len(task_gpu_block_id) == len(task_cpu_block_id) gpu_block_ids = task_gpu_block_id cpu_block_ids = task_cpu_block_id if event_type.value == CacheStatus.SWAP2CPU.value: swap_cache_all_layers( self.gpu_cache_k_tensors, self.k_dst_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 0, ) swap_cache_all_layers( self.gpu_cache_v_tensors, self.v_dst_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 0, ) if self.cache_dtype == "block_wise_fp8": swap_cache_all_layers( self.gpu_cache_scales_k_tensors, self.k_scales_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 0, ) swap_cache_all_layers( self.gpu_cache_scales_v_tensors, self.v_scales_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 0, ) elif event_type.value == CacheStatus.SWAP2GPU.value: swap_cache_all_layers( self.gpu_cache_k_tensors, self.k_dst_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 1, ) swap_cache_all_layers( self.gpu_cache_v_tensors, self.v_dst_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 1, ) if self.cache_dtype == "block_wise_fp8": swap_cache_all_layers( self.gpu_cache_scales_k_tensors, self.k_scales_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 1, ) swap_cache_all_layers( self.gpu_cache_scales_v_tensors, self.v_scales_ptrs, self.num_cpu_blocks, gpu_block_ids, cpu_block_ids, self.device, 1, ) else: logger.warning( f"transfer data: Get unexpected event type {event_type}, only SWAP2CPU and SWAP2GPU supported" ) except Exception as e: logger.error(f"transfer data: error: {e}") raise e end_time = time.time() elasped_time = end_time - start_time logger.info( f"transfer data: transfer_task_id {transfer_task_id} event_type {event_type}: " + f"transfer {len(gpu_block_ids)} blocks done elapsed_time {elasped_time:.4f}" ) return ( event_type, transfer_task_id, swap_node_ids, task_gpu_block_id, task_cpu_block_id, ) def check_cache_status(self, args): # TODO XPU support RL if unset_data_ipc is None: return logger.info("Start a thread to clear/restore kv cache when model weights are cleared/updated.") while True: # handle cache clearing/restoring if self.kv_cache_status_signal.value[0] == KVCacheStatus.CLEARING: assert args.splitwise_role == "mixed", "Only mixed mode supports clearing cache." try: logger.info(f"Start clearing caches {self.cache_ready_signal.value}") # clear cpu caches if self.num_cpu_blocks > 0 and envs.FD_ENABLE_SWAP_SPACE_CLEARING: paddle.set_device("cpu") for ptrs in self.k_dst_ptrs + self.v_dst_ptrs: cuda_host_free(ptrs) self.cpu_cache_kvs.clear() self.k_dst_ptrs.clear() self.v_dst_ptrs.clear() gc.collect() # reset swap_space_ready_signal self.swap_space_ready_signal.value[self.rank] = 0 while np.sum(self.swap_space_ready_signal.value) != 0: time.sleep(0.1) # clear gpu caches set_device(self.device) for name, tensor in self.gpu_cache_kvs.items(): unset_data_ipc(tensor, name, True, False) self.gpu_cache_kvs.clear() self.gpu_cache_k_tensors.clear() self.gpu_cache_v_tensors.clear() # reset cache_ready_signal self.cache_ready_signal.value[self.rank] = 0 logger.info(f"Finish clearing caches {self.cache_ready_signal.value}") # wait for all ranks caches to be cleared if np.sum(self.cache_ready_signal.value) != 0: time.sleep(0.1) # reset kv_cache_status_signal self.kv_cache_status_signal.value[0] = KVCacheStatus.CLEARED logger.info(f"All ranks finish clearing caches {self.cache_ready_signal.value}") except Exception as e: logger.error(f"Failed to clear caches: {e}") elif self.kv_cache_status_signal.value[0] == KVCacheStatus.UPDATING: assert args.splitwise_role == "mixed", "Only mixed mode supports updating cache." try: logger.info(f"Start restoring caches {self.cache_ready_signal.value}") # restore cpu cache if self.num_cpu_blocks > 0 and envs.FD_ENABLE_SWAP_SPACE_CLEARING: self._init_cpu_cache(args) while np.sum(self.swap_space_ready_signal.value) != args.mp_num: time.sleep(0.1) # restore gpu cache and set cache_ready_signal self._init_gpu_cache(args) logger.info(f"Finish restoring caches {self.cache_ready_signal.value}") # wait for all ranks caches to be ready while np.sum(self.cache_ready_signal.value) != args.mp_num: time.sleep(0.1) # set kv_cache_status_signal logger.info(f"All ranks finish restoring caches {self.cache_ready_signal.value}") self.kv_cache_status_signal.value[0] = KVCacheStatus.NORMAL except Exception as e: logger.error(f"Failed to restore caches: {e}") time.sleep(0.1) def main(): """ 启动cache manager """ cache_manager = CacheTransferManager(args) cache_manager.do_data_transfer() if __name__ == "__main__": args = parse_args() rank_id = args.rank + args.local_data_parallel_id * args.mp_num logger = get_logger("cache_transfer_manager", f"cache_transfer_manager_tprank{args.rank}.log") logger.info(f"args: {vars(args)}") set_device(args.device_id) try: main() except Exception as e: logger.error(f"cache_transfer_manager failed with error: {e}, traceback: {traceback.format_exc()}") raise