OC

> WIth this async implementation, I find that an error would immediately occur at the beginning of training (step 1) right after the rollout process. see verl/verl/trainer/ppo/ray_trainer.py @chenhaiq @SwordFaith >...

> @chenhaiq I have tried to run the example via `bash examples/grpo_trainer/run_qwen2-7b_seq_balance.sh `, but I got an error as below: `Exception: sgl-kernel is installed with version 0.1.0, which is less...

> `Exception: sgl-kernel is installed with version 0.1.0, which is less than the minimum required version 0.1.1. Please reinstall the latest version with 'pip install sgl-kernel --force-reinstall'`. Note that I...

> Title: state_dict() hangs when running run_qwen2-7b_seq_balance.sh with tp=2 fixed in https://github.com/volcengine/verl/pull/2098

> @chenhaiq When I run the following command: NGINE=sglang ROLLOUT_MODE=async bash tests/special_e2e/ppo_trainer/run_function_reward.sh the program hangs during execution. I am using 2 GPUs, and all other settings and configurations are exactly...

> We have already implemented this feature, please check `reward_model.launch_reward_fn_async=True` argument yes，my picture in first post is assume reward_model.launch_reward_fn_async=True. In current implementation, reward is a batch task async with old_log_prob...

> I pulled the latest version of verl's code and when running the official gsm8k with tool, multi turn async rollout sglang example without any modifications, the model crashes for...

I hit this error when running with: 1. vlllm 0.9.1 2. rollout_mode="async" 3. VLLM_USE_V1=1 VLLM_USE_V1=0 also can not work. NotImplementedError: VLLM_USE_V1=1 is not supported with Pipeline Parallelism without Ray distributed...

> > rebase and fix format > > tests should pass this time cc @eric-haibin-lin please fix failed ci and rebase with main branch. Do we need to patch vllm...

Is it the same with this one https://github.com/volcengine/verl/issues/1642 ?