minghaosdocker/transformer-a repository overview

version with 1 are cpu, with 2 are gpu, version 1.18 is the first to add communication optimizations. 1.20 no print. 1.24 with cache optimization no prints; 1.25 with cache optimization and prints. After 1.31, models are built with opt3 (no traffic both directions) . model size, tag 125m, 1.30 1.3b, 1.32 350m, 1.33 125m(opt4),1.35 350m(opt4),1.36 1.3b(opt4),1.37 2.7b(opt4),1.38

Starting 2.0, these are distributed model running on MicroK8S instances . model size, tag, split 125m, 2.1, 11- 2.7b, 2.3, 11- 2.7b, 2.6, 5- 1.3b, 2.8, 12- 6.7b, 2.9, 16-

6.7b, 2.10, 16- (no print) <- used for 6.7b tersting

2.7b, 2.11, 16 <- used for 2.7b tersting

1.3b, 2.12, 12- <- used for 1.3b tersting