apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 21 Packages Projects Releases Wiki Activity

Revert "[Optimize] Optimize ttft for ep (#6098)" (#6402)

Browse Source 
This reverts commit 90db0bdd0d.
This commit is contained in:
chenjian
2026-02-09 19:01:23 +08:00
committed by GitHub
parent d60daca4a8
commit 35c24f3f71
10 changed files with 142 additions and 118 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/worker/worker_process.py
+29 -52
View File
@@ -283,16 +283,6 @@ class PaddleDisWorkerProc:
create=False,
)
# init engine forward signal
engine_forward_signal_data = np.zeros([1], dtype=np.int32)
self.engine_forward_signal = IPCSignal(
name="engine_forward_signal",
array=engine_forward_signal_data,
dtype=np.int32,
suffix=self.parallel_config.local_engine_worker_queue_port,
create=False,
)
def update_weights_from_tensor(self, mmap_infos):
"""
update_weights_from_tensor
@@ -450,6 +440,9 @@ class PaddleDisWorkerProc:
# TODO: Unify status variables model_weights_status (shared memory) and model_weights_signal (numpy array) to one
self.model_weights_signal = np.zeros([1], dtype=np.int32)
while True:
# run eplb
self._run_eplb(tp_rank)
if self.fd_config.load_config.dynamic_load_weight:
self.model_weights_signal[0] = int(self.model_weights_status.value[0])
if self.ranks > 1:
@@ -523,7 +516,7 @@ class PaddleDisWorkerProc:
if self.exist_task_signal.value[0] == ExistTaskStatus.EXIST or self.task_queue.read_finish_flag.get() == 1:
logger.info(f"Rank: {self.local_rank} Detected new requests.")
self.engine_forward_signal.value[0] = 1
tasks, read_finish = self.task_queue.get_tasks()
# Only one of all tp_size client will get read_finish == True.
if read_finish:
@@ -532,48 +525,35 @@ class PaddleDisWorkerProc:
self.task_queue.read_finish_flag.set(0)
else:
self.exist_task_signal.value[0] = ExistTaskStatus.EMPTY
# In EP parallel(corresponing to dp attention), we need to barrier for prefill to prevent data imbalance due to inconsistent data arrival.
# Only EP + DP prefill should barrier for data arrival.
# In mixed mode and decoder in D, we should not barrier to influence decoding.
if self.parallel_config.use_ep and self.scheduler_config.splitwise_role == "prefill":
paddle.distributed.barrier(self.parallel_config.ep_group)
req_dicts, control_reqs = [], []
# In EP + DP prefill, empty task ([]) is delived in worker to barrier. For empty task, just skip and continue.
if tasks[0][0]:
for req_dict, bsz in tasks:
if len(req_dict) > 0 and isinstance(req_dict[0], ControlRequest):
control_reqs.append(req_dict[0])
else:
max_occupied_batch_index = int(bsz)
req_dicts.extend(req_dict)
for req_dict, bsz in tasks:
if len(req_dict) > 0 and isinstance(req_dict[0], ControlRequest):
control_reqs.append(req_dict[0])
else:
max_occupied_batch_index = int(bsz)
req_dicts.extend(req_dict)
# todo: run control request async
if len(control_reqs) > 0:
logger.info(f"Rank: {self.local_rank} received {len(control_reqs)} control request.")
for control_req in control_reqs:
self.run_control_method(control_req)
self._tp_barrier_wait() if tp_size > 1 else None
# todo: run control request async
if len(control_reqs) > 0:
logger.info(f"Rank: {self.local_rank} received {len(control_reqs)} control request.")
for control_req in control_reqs:
self.run_control_method(control_req)
self._tp_barrier_wait() if tp_size > 1 else None
# Count prefill requests in current batch
num_prefill_requests = sum(1 for req in req_dicts if req.task_type == RequestType.PREFILL)
num_scheduled_requests = len(req_dicts)
scheduled_request_ids = [req.request_id for req in req_dicts]
logger.info(
f"Rank: {self.local_rank}, num_prefill_requests: {num_prefill_requests}, "
f"max_occupied_batch_index: {max_occupied_batch_index}, "
f"num_scheduled_requests: {num_scheduled_requests}, "
f"scheduled_request_ids: {scheduled_request_ids}"
)
# Count prefill requests in current batch
num_prefill_requests = sum(1 for req in req_dicts if req.task_type == RequestType.PREFILL)
num_scheduled_requests = len(req_dicts)
scheduled_request_ids = [req.request_id for req in req_dicts]
logger.info(
f"Rank: {self.local_rank}, num_prefill_requests: {num_prefill_requests}, "
f"max_occupied_batch_index: {max_occupied_batch_index}, "
f"num_scheduled_requests: {num_scheduled_requests}, "
f"scheduled_request_ids: {scheduled_request_ids}"
)
# Process prefill inputs
self.worker.preprocess_new_task(req_dicts, max_occupied_batch_index)
else:
if self.scheduler_config.splitwise_role == "prefill":
if tp_size > 1:
# Synchronize the signal for other workers
self._tp_barrier_wait()
continue
# Process prefill inputs
self.worker.preprocess_new_task(req_dicts, max_occupied_batch_index)
if (
(not self.parallel_config.use_ep)
@@ -581,7 +561,7 @@ class PaddleDisWorkerProc:
and (not self.enable_overlap_schedule)
):
self._tp_barrier_wait() if tp_size > 1 else None
self.engine_forward_signal.value[0] = 0
time.sleep(0.001)
continue
@@ -593,9 +573,6 @@ class PaddleDisWorkerProc:
if not envs.ENABLE_V1_KVCACHE_SCHEDULER:
self.exist_prefill_task_signal.value[0] = self.worker.exist_prefill()
logger.debug(f"execute model cost: {time.time()-start_execute_time:.5f} s")
# run eplb
self._run_eplb(tp_rank)
self.engine_forward_signal.value[0] = 0
def initialize_kv_cache(self) -> None:
"""Profiles the peak memory usage of the model to determine how many