apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-08 18:11:00 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

fix num_seqs (#3396)

Browse Source
This commit is contained in:
lizexu123
2025-08-21 14:03:11 +08:00
committed by GitHub
parent 5d4d38674f
commit fcdc5c2c54
2 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/worker/gpu_model_runner.py
View File

@@ -256,7 +256,9 @@ class GPUModelRunner(ModelRunnerBase):
) )
input_ids = request.prompt_token_ids + request.output_token_ids input_ids = request.prompt_token_ids + request.output_token_ids
logger.debug(f"Handle prefill request {request} at idx {idx} prefill_start_index {prefill_start_index} prefill_end_index {prefill_end_index} need_prefilled_token_num {len(input_ids)}") logger.debug(
f"Handle prefill request {request} at idx {idx} prefill_start_index {prefill_start_index} prefill_end_index {prefill_end_index} need_prefilled_token_num {len(input_ids)}"
)
self.share_inputs["input_ids"][idx : idx + 1, :length] = np.array( self.share_inputs["input_ids"][idx : idx + 1, :length] = np.array(
input_ids[prefill_start_index:prefill_end_index] input_ids[prefill_start_index:prefill_end_index]
) )
@@ -617,7 +619,9 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["max_length"] = paddle.full( self.share_inputs["max_length"] = paddle.full(
[max_num_seqs, 1], self.model_config.max_model_len, dtype="int64" [max_num_seqs, 1], self.model_config.max_model_len, dtype="int64"
) )
self.seq_lens_this_time_buffer = paddle.full(max_num_seqs, 0, dtype="int32") self.seq_lens_this_time_buffer = paddle.full([max_num_seqs, 1], 0, dtype="int32")
if self.fd_config.parallel_config.enable_expert_parallel:
self.share_inputs["seq_lens_this_time"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
self.share_inputs["seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32") self.share_inputs["seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
self.share_inputs["seq_lens_decoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32") self.share_inputs["seq_lens_decoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")
self.share_inputs["step_seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32") self.share_inputs["step_seq_lens_encoder"] = paddle.full([max_num_seqs, 1], 0, dtype="int32")

2
fastdeploy/worker/worker_process.py
View File

@@ -244,6 +244,7 @@ class PaddleDisWorkerProc:
Tmp loop function for ep utill DP is supported Tmp loop function for ep utill DP is supported
""" """
while True: while True:
num_running_requests = 0
self.worker_healthy_live_signal.value[self.local_rank % self.max_chips_per_node] = int(time.time()) self.worker_healthy_live_signal.value[self.local_rank % self.max_chips_per_node] = int(time.time())
if self.fd_config.parallel_config.tensor_parallel_rank == 0 and self.task_queue.num_tasks() > 0: if self.fd_config.parallel_config.tensor_parallel_rank == 0 and self.task_queue.num_tasks() > 0:
@@ -272,6 +273,7 @@ class PaddleDisWorkerProc:
self.nnode = int((self.parallel_config.tensor_parallel_size + 7) // 8) self.nnode = int((self.parallel_config.tensor_parallel_size + 7) // 8)
mp_num_per_node = self.parallel_config.tensor_parallel_size // self.nnode mp_num_per_node = self.parallel_config.tensor_parallel_size // self.nnode
req_ids = [] req_ids = []
num_running_requests = 0
while True: while True:
if self.local_rank == 0: if self.local_rank == 0:
if self.model_weights_status.value[0] != 0: if self.model_weights_status.value[0] != 0: