# \[rllib\]REPRODUCIBILITY: Is current implementation is enough?

**URL:** <https://discuss.ray.io/t/rllib-reproducibility-is-current-implementation-is-enough/3526>\
**Category:** RLlib\
**Created:** [September 15, 2021, 1:49am UTC](https://discuss.ray.io/t/rllib-reproducibility-is-current-implementation-is-enough/3526 "2021-09-15T01:49:28Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![Shanchao\_Yang](https://sea2.discourse-cdn.com/flex020/user_avatar/discuss.ray.io/shanchao_yang/32/803_2.png) [@Shanchao\_Yang](https://discuss.ray.io/u/Shanchao_Yang)\
**Post date:** [September 15, 2021, 1:49am UTC](https://discuss.ray.io/t/rllib-reproducibility-is-current-implementation-is-enough/3526/1 "2021-09-15T01:49:28Z")

</div>

Hi,

According to the cublas (sec 2.4) [official document](https://docs.nvidia.com/cuda/cublas/index.html), 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.

> <https://github.com/ray-project/ray/blob/b8c361d3fb45ee717748df29a18d3db4a426eeb3/rllib/utils/debug.py#L92-L108>