1.. warning:: 2 There are known non-determinism issues for RNN functions on some versions of cuDNN and CUDA. 3 You can enforce deterministic behavior by setting the following environment variables: 4 5 On CUDA 10.1, set environment variable ``CUDA_LAUNCH_BLOCKING=1``. 6 This may affect performance. 7 8 On CUDA 10.2 or later, set environment variable 9 (note the leading colon symbol) 10 ``CUBLAS_WORKSPACE_CONFIG=:16:8`` 11 or 12 ``CUBLAS_WORKSPACE_CONFIG=:4096:2``. 13 14 See the `cuDNN 8 Release Notes`_ for more information. 15 16.. _cuDNN 8 Release Notes: https://docs.nvidia.com/deeplearning/sdk/cudnn-release-notes/rel_8.html 17