apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 17:17:14 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[feat] support prefix cache clearing when /clear_load_weight is called (#4008)

Browse Source 
* [feat] support clearing prefix cache (cherry-picked from release/2.1)

* [fix] fix ipc suffix, use port instead

* [fix] fix prefix caching not enabled

* [fix] fix key/value_cache_scales indent

* [fix] fix ep group all-reduce

* [fix] fix clear/update lock not working when workers > 1

* [chore] add preemption triggered info log

* [fix] fix code style

* [fix] fix max_num_seqs config

* [fix] do not force enable_prefix_caching=False in dynamic loading

* [fix] fix ci

* Revert "[fix] fix ci"

This reverts commit 0bc6d55cc8.

* [fix] initialize available_gpu_block_num with max_gpu_block_num

* [fix] fix config splitwise_role

* [fix] fix clearing caches synchronization and add more logs

* [chore] print cache_ready_signal in log

* [fix] fix scheduler_config.splitwise_role

* [fix] fix cache_messager cache_ready_signal create=True

* [fix] stop cache messager from launching in mixed deployment
This commit is contained in:
李泳桦
2025-09-28 19:42:53 +08:00
committed by GitHub
parent 59313ed7f9
commit 6265f4385f
20 changed files with 697 additions and 213 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
fastdeploy/engine/common_engine.py
View File

@@ -188,6 +188,24 @@ class EngineService:
create=True,
)
cache_ready_signal_data = np.zeros(shape=[self.cfg.parallel_config.tensor_parallel_size], dtype=np.int32)
self.cache_ready_signal = IPCSignal(
name="cache_ready_signal",
array=cache_ready_signal_data,
dtype=np.int32,
suffix=current_suffix,
create=True,
)
swap_space_ready_signal_data = np.zeros(shape=[self.cfg.parallel_config.tensor_parallel_size], dtype=np.int32)
self.swap_space_ready_signal = IPCSignal(
name="swap_space_ready_signal",
array=swap_space_ready_signal_data,
dtype=np.int32,
suffix=current_suffix,
create=True,
)
model_weights_status = np.zeros([1], dtype=np.int32)
self.model_weights_status_signal = IPCSignal(
name="model_weights_status",
@@ -197,6 +215,24 @@ class EngineService:
create=True,
)
prefix_tree_status = np.zeros([1], dtype=np.int32)
self.prefix_tree_status_signal = IPCSignal(
name="prefix_tree_status",
array=prefix_tree_status,
dtype=np.int32,
suffix=current_suffix,
create=True,
)
kv_cache_status = np.zeros([1], dtype=np.int32)
self.kv_cache_status_signal = IPCSignal(
name="kv_cache_status",
array=kv_cache_status,
dtype=np.int32,
suffix=current_suffix,
create=True,
)
def start_worker_queue_service(self, start_queue):
"""
start queue service for engine worker communication
@@ -935,7 +971,7 @@ class EngineService:
threading.Thread(target=receiver_loop, daemon=True).start()
def start_cache_service(self, device_ids, ipc_signal_suffix):
def start_cache_service(self, device_ids, ipc_signal_suffix, create_cache_tensor):
return self.resource_manager.cache_manager.launch_cache_manager(
cache_config=self.cfg.cache_config,
tensor_parallel_size=self.cfg.parallel_config.tensor_parallel_size,
@@ -945,6 +981,7 @@ class EngineService:
self.cfg.parallel_config.engine_worker_queue_port[self.cfg.parallel_config.local_data_parallel_id]
),
pid_suffix=ipc_signal_suffix,
create_cache_tensor=create_cache_tensor,
)
def check_and_free_block_tables(self):
@@ -971,8 +1008,12 @@ class EngineService:
self.exist_task_signal.clear()
self.exist_swapped_task_signal.clear()
self.worker_healthy_live_signal.clear()
self.cache_ready_signal.clear()
self.swap_space_ready_signal.clear()
self.exist_prefill_task_signal.clear()
self.model_weights_status_signal.clear()
self.prefix_tree_status_signal.clear()
self.kv_cache_status_signal.clear()
if hasattr(self, "send_response_server") and self.send_response_server is not None:
self.send_response_server.close()
if hasattr(self, "recv_request_server") and self.recv_request_server is not None: