-
|
Hello, I hope you are well. Upon building the entire framework: PyTorch On CPU only, is there any issue in your view to running the LAMMPS+Allegro simulation in distributed across multiple nodes? Thank you! |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 6 replies
-
|
Fixed the Issue |
Beta Was this translation helpful? Give feedback.
-
|
I understand, thank you for the reply. I can work on cpu improvement. What areas of the code would you advise to focus on? |
Beta Was this translation helpful? Give feedback.
-
|
It is possible that using PyTorch 2.0 with the new Intel Extension for PyTorch might get you somewhere. I'll also be sure to post again if we have relevant efforts in this direction. Otherwise, having built PyTorch from scratch against MKL, I would spend some time playing with the balance between LAMMPS ranks and OpenMP/MKL threads to see what gives optimal performance, maybe? |
Beta Was this translation helpful? Give feedback.
-
|
Agreed, perfect, thank you. Building PyTorch from source and using ipex helped quite a bit. Will focus on tuning the LAMMPS settings relating to OMP, etc. Will be in touch. |
Beta Was this translation helpful? Give feedback.
-
|
Sounds good--- @anjohan mentioned that he thought the extreme of one rank per core and one OpenMP thread might be best, so might be worth trying that as well. No harm probably in trying the other extreme either. |
Beta Was this translation helpful? Give feedback.
-
|
Excellent. Yeah, Anders seems to be correct actually. Scaling across many cpu nodes would require a different approach of course. I doubt we could do 1 rank per core across many CPUs and maintain performance. |
Beta Was this translation helpful? Give feedback.
Fixed the Issue