[rllib]REPRODUCIBILITY: Is current implementation is enough?

Shanchao_Yang · September 15, 2021, 1:49am

Hi,

According to the cublas (sec 2.4) official document, we should set the variable as:
os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"

For cuda with version > 10.2, the current strategy for reproducible results is not enough.
It only sets os.environ["CUBLAS_WORKSPACE_CONFIG"] = "4096:8", and does not call
torch.use_deterministic_algorithms(True).

I cannot understand what is the intention behind this implementation. I think it is not complete.

github.com

ray-project/ray/blob/b8c361d3fb45ee717748df29a18d3db4a426eeb3/rllib/utils/debug.py#L92-L108

    
      
          if framework == "torch":
              torch, _ = try_import_torch()
              torch.manual_seed(seed)
              # See https://github.com/pytorch/pytorch/issues/47672.
              cuda_version = torch.version.cuda
              if cuda_version is not None and float(torch.version.cuda) >= 10.2:
                  os.environ["CUBLAS_WORKSPACE_CONFIG"] = "4096:8"
              else:
                  from distutils.version import LooseVersion
          
          
        if LooseVersion(torch.__version__) >= LooseVersion("1.8.0"):
                      # Not all Operations support this.
                      torch.use_deterministic_algorithms(True)
                  else:
                      torch.set_deterministic(True)
              # This is only for Convolution no problem.
              torch.backends.cudnn.deterministic = True

Topic		Replies	Views
Issue Reproducing results RLlib	4	900	June 14, 2021
Issues after upgrading from 1.6.0 fro 1.7.0 RLlib	3	348	October 17, 2021
Reference runtime environment? python + tensorflow/pytorch RLlib	0	174	October 14, 2021
Evolution strategies - make reproducible RLlib	1	455	July 14, 2021
Questions getting started with newer versions of rllib with tensorflow RLlib	0	20	April 4, 2024

[rllib]REPRODUCIBILITY: Is current implementation is enough?

Related Topics