NameError: name 'full_response_texts' is not defined

[iNARI777]

Hi, I’ve set up the environment and was able to start the training process, but I ran into the following error partway through:

ray.exceptions.RayTaskError(NameError): ray::WorkerDict.actor_rollout_generate_sequences() (pid=102168, actor_id=fcf54f15226dc61eb2850b9701000000, repr=<verl.single_controller.ray.base.WorkerDict object at 0x7f82ca9dad50>)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hadoop-aipnlp/project/SkyRL/verl/single_controller/ray/base.py", line 440, in func
    return getattr(self.worker_dict[key], name)(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hadoop-aipnlp/project/SkyRL/verl/single_controller/base/decorator.py", line 404, in inner
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/hadoop-aipnlp/project/SkyRL/verl/workers/fsdp_workers.py", line 560, in generate_sequences
    output = self.rollout.generate_sequences(prompts=prompts)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hadoop-aipnlp/project/SkyRL/verl/workers/agentic/async_rollout.py", line 114, in generate_sequences
    results = codeact_agent_group.run()
              ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hadoop-aipnlp/project/SkyRL/verl/workers/agentic/codeact.py", line 1371, in run
    return loop.run_until_complete(self.generate_trajectories_pipeline())
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
  File "/home/hadoop-aipnlp/project/SkyRL/verl/workers/agentic/codeact.py", line 1351, in generate_trajectories_pipeline
    results_dataproto = self._convert_results_to_dataproto()
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hadoop-aipnlp/project/SkyRL/verl/workers/agentic/codeact.py", line 774, in _convert_results_to_dataproto
    "response_texts": full_response_texts,
                      ^^^^^^^^^^^^^^^^^^^
NameError: name 'full_response_texts' is not defined

I checked the code and it looks like the variable full_response_texts is indeed not defined anywhere. I also did a global search and it seems that response_texts from this dictionary isn’t used elsewhere in the code.

So, is it safe to simply remove this line (line 774) from _convert_results_to_dataproto? Or is there something I might be missing?

Thanks!

[caoshiyi]

Hi @iNARI777, it's great to see you can start the training process! Yes, this line can just be removed safely. The full_response_texts entry is used for us to debug; we forgot to clean the line here.