Hi, I'm trying to use the Accelerate integration, because otherwise with NLPO I cannot run a small model (200M parameter) with 512 tokens length, not even in a 80GB A100. That makes NLPO impractical for almost any problem, unless you can use Accelerate / Deepspeed or any other integration for splitting models among GPUs and CPUs. However, when trying to do so, I receive the following error:
Traceback (most recent call last):
File "/home/alejandro.vaca/nlp_rl/RL4LMs/scripts/training/train_text_generation.py", line 95, in <module>
main(
File "/home/alejandro.vaca/nlp_rl/RL4LMs/scripts/training/train_text_generation.py", line 56, in main
trainer = OnPolicyTrainer(
File "/home/alejandro.vaca/nlp_rl/RL4LMs/rl4lms/envs/text_generation/training_utils.py", line 190, in __init__
self._setup()
File "/home/alejandro.vaca/nlp_rl/RL4LMs/rl4lms/envs/text_generation/training_utils.py", line 207, in _setup
self._alg = build_alg(
File "/home/alejandro.vaca/nlp_rl/RL4LMs/rl4lms/envs/text_generation/training_utils.py", line 151, in build_alg
alg = wrapper(
File "/home/alejandro.vaca/nlp_rl/RL4LMs/rl4lms/envs/text_generation/alg_wrappers.py", line 451, in wrap_onpolicy_alg
alg = OnPolicyAlgText(alg_kwargs, kl_coeff, tracker, accelerator, target_kl, norm_reward)
File "/home/alejandro.vaca/nlp_rl/RL4LMs/rl4lms/envs/text_generation/alg_wrappers.py", line 126, in __init__
super().__init__(**alg_kwargs)
TypeError: __init__() got an unexpected keyword argument 'accelerator'
Also, I'm curious about how you were able to carry out the benchmarks in your paper, as I don't know any single GPU bigger than 80GB and even there I cannot run NLPO with sentences longer than 128 tokens. How did you do it? Is there maybe something I'm missing? @rajcscw @jmhessel @rajammanabrolu @JulesGM @akifumi-wachi-4
Thank you very much for this amazing work!! :)
Hi, I'm trying to use the Accelerate integration, because otherwise with NLPO I cannot run a small model (200M parameter) with 512 tokens length, not even in a 80GB A100. That makes NLPO impractical for almost any problem, unless you can use Accelerate / Deepspeed or any other integration for splitting models among GPUs and CPUs. However, when trying to do so, I receive the following error:
Also, I'm curious about how you were able to carry out the benchmarks in your paper, as I don't know any single GPU bigger than 80GB and even there I cannot run NLPO with sentences longer than 128 tokens. How did you do it? Is there maybe something I'm missing? @rajcscw @jmhessel @rajammanabrolu @JulesGM @akifumi-wachi-4
Thank you very much for this amazing work!! :)