apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[BugFix] Add prefill restrictions for chunked_prefill+VL (#2983)

Browse Source
This commit is contained in:
Zero Rains
2025-07-23 16:45:57 +08:00
committed by GitHub
parent a39a67334c
commit 850c9d98d4
2 changed files with 18 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
fastdeploy/worker/gpu_model_runner.py
View File

@@ -150,10 +150,17 @@ class GPUModelRunner(ModelRunnerBase):
""" """
Check whether prefill stage finished Check whether prefill stage finished
""" """
if int(paddle.max(self.share_inputs["seq_lens_encoder"])) != 0: if self.enable_mm:
return 1 # VL only support 1 batch to prefill
prefill_statue = (self.share_inputs["seq_lens_this_time"] != 0) & (
self.share_inputs["seq_lens_this_time"] != 1
)
return not paddle.any(prefill_statue).numpy()
else: else:
return 0 if int(paddle.max(self.share_inputs["seq_lens_encoder"])) != 0:
return 1
else:
return 0
def _init_speculative_proposer(self): def _init_speculative_proposer(self):
""" """

14
fastdeploy/worker/worker_process.py
View File

@@ -285,10 +285,12 @@ class PaddleDisWorkerProc:
# The first worker detects whether there are tasks in the task queue # The first worker detects whether there are tasks in the task queue
if self.local_rank % mp_num_per_node == 0: if self.local_rank % mp_num_per_node == 0:
if self.task_queue.num_tasks() > 0: if self.task_queue.num_tasks() > 0:
if self.nnode > 1: # VL only support 1 batch to prefill
self.task_queue.read_finish_flag.set(1) if not self.fd_config.model_config.enable_mm or self.worker.prefill_finished():
else: if self.nnode > 1:
self.exist_task_signal.value[self.fd_config.parallel_config.expert_parallel_rank] = 1 self.task_queue.read_finish_flag.set(1)
else:
self.exist_task_signal.value[self.fd_config.parallel_config.expert_parallel_rank] = 1
if self.parallel_config.tensor_parallel_size > 1: if self.parallel_config.tensor_parallel_size > 1:
# Synchronize the signal for other workers # Synchronize the signal for other workers
@@ -344,8 +346,8 @@ class PaddleDisWorkerProc:
# Execute model to generate token. The generated token will be written to the buffer. # Execute model to generate token. The generated token will be written to the buffer.
# These generated tokens can be obtained through get_output op. # These generated tokens can be obtained through get_output op.
self.worker.execute_model(req_dicts) self.worker.execute_model(req_dicts)
if not self.fd_config.model_config.enable_mm:
self.exist_prefill_task_signal.value[0] = self.worker.prefill_finished() self.exist_prefill_task_signal.value[0] = self.worker.prefill_finished()
def initialize_kv_cache(self) -> None: def initialize_kv_cache(self) -> None:
"""Profiles the peak memory usage of the model to determine how many """Profiles the peak memory usage of the model to determine how many