LM transformer benchmark with MLX

apple / pfl-research

Simulation framework for accelerating research in Private Federated Learning

http://apple.github.io/pfl-research/

Apache License 2.0

281 stars 27 forks source link

LM transformer benchmark with MLX #82

Closed grananqvist closed 1 month ago

grananqvist commented 1 month ago

Matches the perplexity with pytorch setup (60.87) (or even better):

                num_processes         hours to converge                 PPL
M1 Pro        5                            8.18                          59.53
M3 Max        5                            3.11                          59.25