apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Support limit thinking lengths (#4244)

Browse Source 
Co-authored-by: K11OntheBoat <“ruianmaidanglao@163.com”>
This commit is contained in:
K11OntheBoat
2025-09-24 17:30:53 +08:00
committed by GitHub
parent 12043fc476
commit 05b7800d80
8 changed files with 184 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
fastdeploy/worker/worker_process.py
View File

@@ -131,6 +131,28 @@ def update_fd_config_for_mm(fd_config: FDConfig) -> None:
fd_config.model_config.sequence_parallel = fd_config.parallel_config.sequence_parallel
def update_think_end_id_for_ernie(fd_config: FDConfig) -> None:
"""
Updates the think_end_id in the model config. Uses the ID of '</think>'
if it exists, otherwise defaults to None.
"""
is_ernie = ErnieArchitectures.contains_ernie_arch(fd_config.model_config.architectures)
if current_platform.is_cuda() and is_ernie:
tokenizer = Ernie4_5Tokenizer.from_pretrained(
fd_config.model_config.model,
model_max_length=fd_config.parallel_config.max_model_len,
padding_side="right",
use_fast=False,
)
vocab = tokenizer.get_vocab()
fd_config.model_config.think_end_id = vocab.get("</think>", None)
if fd_config.model_config.think_end_id is not None:
logger.info(f"Get think_end_id {fd_config.model_config.think_end_id} from vocab.")
else:
logger.info("No </think> token found in vocabulary, the model can not do reasoning.")
class PaddleDisWorkerProc:
"""
Paddle Distributed wrapper for fastdeploy.worker.Worker,
@@ -798,6 +820,7 @@ def initialize_fd_config(args, ranks: int = 1, local_rank: int = 0) -> FDConfig:
moba_attention_config=moba_attention_config,
)
update_fd_config_for_mm(fd_config)
update_think_end_id_for_ernie(fd_config)
return fd_config