Open MR-Jarble opened 1 month ago
I have additional data in that this also occurs on systems with dual sockets as well. When using a dual 64 core cpu system the transformers section will only engage one cpu leaving the other unused. So I am not as sure that this is a numa issue but maybe something else.
Describe the bug
When working with the transformers section I have found that the system will only engage half my cpu cores/threads. It stops exactly at numa node two and will not use nodes three or four. I have a feeling this is an extreme edge case with my hardware as most systems only have one or two nodes but mine has four. As best I can tell this is not a thread issue as systems with lower numa counts and higher threads seem fine.
Is there an existing issue for this?
Reproduction
Load any transformer model and see half of the cores at 100%
Screenshot
Logs
System Info