Skip to content

deps/cuda: Set the XLA_FLAGS env var

Fabian Tërnava requested to merge fix/set_cuda_XLA_FLAGS into main

Tensorflow only detects the graka if active_cuda was called. But even if cuda detects the graphics card, it is not possible to use it in training, because libdevice.so requires to know where to find nvvm. The XLA_FLAGS tells libdevice.so where to find the nvvm folder.

see https://stackoverflow.com/questions/68614547/tensorflow-libdevice-not-found-why-is-it-not-found-in-the-searched-path

PS: Probaly the same issue appears on other cuda versions

Edited by Fabian Tërnava

Merge request reports