apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 00:06:38 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Merge vl execution path into normal execution path (#2829)

Browse Source 
* merge vl model into gpu_model runner

Change-Id: I9f4691a3d5f135e8d72b1d58abcd15ef3aa3f2a6

* fix chinese

Change-Id: Ic7405109b984c21e076fb3b01ff6feb571d0119a

* fix the parse parameter

Change-Id: I4cd62ee87c06220af580d91e347145d4394917fe

* fix the bug in online_inference

Change-Id: Idb111bb2114e83017c4050b2a68cf039c6d3c559

* polish code

Change-Id: I7d4194102c2f1b0743b74fbd5fc284eb8ef4d17c
This commit is contained in:
Zero Rains
2025-07-15 22:20:03 +08:00
committed by GitHub
parent 5fc659b900
commit e7bcbbab52
9 changed files with 441 additions and 1732 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
fastdeploy/model_executor/pre_and_post_process.py
View File

@@ -129,6 +129,36 @@ def post_process_normal(sampler_output: SamplerOutput,
save_each_rank: bool = False,
skip_save_output: bool = False) -> 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
paddle.assign(
paddle.where(
exists_think_end,
model_output.need_think_end - 1,
model_output.need_think_end,
), model_output.need_think_end)
paddle.assign(
paddle.where(
model_output.need_think_end.cast("bool"),
model_output.reasoning_index - 1,
model_output.reasoning_index,
), model_output.reasoning_index)
stop_wo_think = (
(sampler_output.sampled_token_ids == model_output.eos_token_id) |
(model_output.reasoning_index == 0)) & (
model_output.need_think_end > 0)
sampler_output.sampled_token_ids = paddle.where(stop_wo_think,
model_output.think_end_id,
sampler_output.sampled_token_ids)
paddle.assign(
paddle.where(
stop_wo_think,
model_output.need_think_end - 1,
model_output.need_think_end,
), model_output.need_think_end)
# 1. Set stop value
paddle.assign(
paddle.where(