apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 08:16:42 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[BugFix] num_seqs (#3291)

Browse Source 
* fix num_seqs

* merge develop
This commit is contained in:
lizexu123
2025-08-11 13:38:55 +08:00
committed by GitHub
parent b88537a456
commit e0aeac58e1
3 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
fastdeploy/model_executor/models/ernie4_5_moe.py
View File

@@ -450,7 +450,7 @@ class Ernie4_5_MoeForCausalLM(ModelForCasualLM):
self.fd_config.model_config.moe_layer_start_index,
self.fd_config.model_config.num_hidden_layers,
):
self.ernie.layers[i].mlp.expert(fake_hidden_states)
self.ernie.layers[i].mlp.experts(fake_hidden_states, self.ernie.layers[i].mlp.gate)
def forward(
self,

4
fastdeploy/worker/gpu_model_runner.py
View File

@@ -795,7 +795,7 @@ class GPUModelRunner(ModelRunnerBase):
output_padding_offset,
) = pre_process(
self.share_inputs["input_ids"],
self.share_inputs["seq_lens_this_time"],
getattr(self.share_inputs, "seq_lens_this_time", self.seq_lens_this_time_buffer),
self.speculative_decoding,
(self.share_inputs["draft_tokens"] if self.speculative_decoding else None),
self.share_inputs["seq_lens_encoder"],
@@ -880,7 +880,7 @@ class GPUModelRunner(ModelRunnerBase):
max_len_tensor_cpu=self.share_inputs["max_len_tensor_cpu"],
seq_lens_encoder=self.share_inputs["seq_lens_encoder"],
seq_lens_decoder=self.share_inputs["seq_lens_decoder"],
seq_lens_this_time=self.share_inputs["seq_lens_this_time"],
seq_lens_this_time=getattr(self.share_inputs, "seq_lens_this_time", self.seq_lens_this_time_buffer),
batch_id_per_token=self.share_inputs["batch_id_per_token"],
cu_seqlens_q=self.share_inputs["cu_seqlens_q"],
cu_seqlens_k=self.share_inputs["cu_seqlens_k"],

3
fastdeploy/worker/worker_process.py
View File

@@ -244,6 +244,7 @@ class PaddleDisWorkerProc:
"""
while True:
self.worker_healthy_live_signal.value[self.local_rank % self.max_chips_per_node] = int(time.time())
num_running_requests = 0
if self.fd_config.parallel_config.tensor_parallel_rank == 0 and self.task_queue.num_tasks() > 0:
tasks, read_finish = self.task_queue.get_tasks()
@@ -271,6 +272,8 @@ class PaddleDisWorkerProc:
self.nnode = int((self.parallel_config.tensor_parallel_size + 7) // 8)
mp_num_per_node = self.parallel_config.tensor_parallel_size // self.nnode
req_ids = []
num_running_requests = 0
while True:
if self.local_rank == 0:
if self.model_weights_status.value[0] != 0: