poolINTERNAL ASSERT FAILED at "../aten/src/ATen/ParallelOpenMP.cpp":65, please report a bug to PyTorch. Invalid thread pool

pytorch version: 1.11.0
device: 96 CPU cores, used by many people, often slow.(Actually the problem goes away if I use a better CPU)
OS: Centos 7
The error is as follows:


The error occurs when the model is run for the first time. Then it disappears when run the second time and so on. So it is easy to work around the problem by warming up. But I’m curious about why it fails to create caffe2 threadpool.

Could you check if you are still seeing the same issue using the latest nightly binary? If so, could you create a GitHub issue with instructions to reproduce it, please?