length_controller in bash script

[NieSYsc20]

Hello! Thank you for providing the open-source code! However, I encountered a problem when trying to reproduce laser-d. It seems that the length_controller.pt file is not saved according to the training code, which makes it impossible to resume training from a checkpoint. The reason appears to be that length_strategy: constant is fixed in the configuration file and is not modified in the bash script. Could you please check if the bash script needs to include the length_strategy parameter?

Laser/verl/trainer/config/lengthy_ppo_trainer.yaml

Line 210 in c9e7c4e

length_strategy: constant

Laser/verl/trainer/ppo/ray_trainer.py

Line 1442 in c9e7c4e

if self.length_controller.length_strategy != 'constant':

[VPeterV]

Hi! Thanks for your interest, and sorry for the missing hyperparameter. Could you try setting it to a value other than constant? Based on our initial design, that should resolve the issue. We're currently working on updating the entire codebase to support the latest Verl and Qwen3-series models, and we’ll make sure to fix and test this feature as part of that process.