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 (#4069)

Browse Source 
Co-authored-by: K11OntheBoat <“ruianmaidanglao@163.com”>
This commit is contained in:
K11OntheBoat
2025-09-25 19:55:56 +08:00
committed by GitHub
parent 0c6f1932c5
commit 4515ad21e9
9 changed files with 194 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
fastdeploy/model_executor/pre_and_post_process.py
View File

@@ -195,8 +195,9 @@ def post_process_normal(
) -> ModelRunnerOutput:
"""Post-processing steps after completing a single token generation."""
# handle vl:
if model_output.enable_thinking:
exists_think_end = sampler_output.sampled_token_ids == model_output.think_end_id
if model_output.think_end_id != -1:
thinking_mask = model_output.enable_thinking
exists_think_end = (sampler_output.sampled_token_ids == model_output.think_end_id) & thinking_mask
paddle.assign(
paddle.where(
exists_think_end,
@@ -206,9 +207,10 @@ def post_process_normal(
model_output.need_think_end,
)
reasoning_index_update_cond = model_output.need_think_end.cast("bool") & thinking_mask
paddle.assign(
paddle.where(
model_output.need_think_end.cast("bool"),
reasoning_index_update_cond,
model_output.reasoning_index - 1,
model_output.reasoning_index,
),
@@ -219,6 +221,8 @@ def post_process_normal(
(sampler_output.sampled_token_ids == model_output.eos_token_id.T).any(axis=1, keepdim=True)
| (model_output.reasoning_index == 0)
) & (model_output.need_think_end > 0)
stop_wo_think = stop_wo_think & thinking_mask
sampler_output.sampled_token_ids = paddle.where(
stop_wo_think,
model_output.think_end_id,