Commit Graph

30 Commits

Author SHA1 Message Date
zzhhjjj
ec1e1e5ccf support bf16, all reduce loss 2024-10-22 23:38:44 +00:00
zzhhjjj
a6d79b07b5 add cuda kernels 2024-10-22 22:38:29 +00:00
zzhhjjj
9a7904d5d6 revert some change 2024-10-22 19:50:23 +00:00
ferdinand.mom
2b2781a374 made Tensor Parallel API compliant 2024-10-18 15:51:26 +00:00
ferdinand.mom
abd1edf9f9 all_reduce loss across pp/dp ranks + base_parallel 2024-10-18 15:51:17 +00:00
ferdinand.mom
1ebd3de5be Merge DDP + TP from @zzhhjjj 2024-10-18 15:05:01 +00:00
ferdinand.mom
83ddda2ce8 leave out CP integration at the very end 2024-10-18 14:59:39 +00:00
ferdinand.mom
0b1d02a402 various fix (modeling, dataloader, cpu load) 2024-10-18 14:33:46 +00:00
zzhhjjj
7377238741 tesnsor parallel, will clean later 2024-10-18 05:13:44 +00:00
zzhhjjj
54ad77e055 Merge branch 'main' into ddp-merge 2024-10-16 19:13:48 +00:00
zzhhjjj
24ff8d05fd add DDP 2024-10-16 16:48:55 +00:00
zzhhjjj
6080c2f26b dataloader 2024-10-16 15:58:35 +00:00
ferdinand.mom
81726dfffe accelerate dataset mapping 2024-10-15 13:32:44 +00:00
ferdinand.mom
1ca7365506 stitch cp dp cp together 2024-10-15 13:06:17 +00:00
ferdinand.mom
ffea3d2ad1 add context parallel for training 2024-10-15 12:43:28 +00:00
ferdinand.mom
1e229cae88 renaming 2024-10-14 09:26:31 +00:00
ferdinand.mom
3095ff4d4f refactor organisation 2024-10-10 15:12:14 +00:00
ferdinand.mom
47581d29e9 make new modeling compatible with training 2024-10-10 15:08:23 +00:00
ferdinand.mom
8c155f47ce all reduce gradient across DP & CP ranks 2024-09-26 14:00:06 +00:00
ferdinand.mom
31b5fb9efc ugly ass display of grid (to be changed) 2024-09-26 13:45:53 +00:00
ferdinand.mom
b8065de7aa support CPU training through gloo backend 2024-09-26 10:27:20 +00:00
ferdinand.mom
6f6bc1945a add wandb support 2024-09-25 14:19:16 +00:00
ferdinand.mom
cfbf6c170e every rank has now the loss 2024-09-25 14:12:31 +00:00
ferdinand.mom
b2e276d3b8 rename parallel_context to process_group_manager 2024-09-25 13:33:20 +00:00
ferdinand.mom
9e9ef8236e refactor to decouple pp training with normal training 2024-09-25 13:17:05 +00:00
ferdinand.mom
e2c0747fe3 add naive DP 2024-09-25 12:36:22 +00:00
ferdinand.mom
7ba1383ebb fixing socket bug by using dist.new_subgroups_by_enumeration instead 2024-09-24 13:43:22 +00:00
ferdinand.mom
7a57407c54 breaking: socketStartConnect: Connect to <ip address> failed : Software caused connection abort 2024-09-23 21:14:48 +00:00
ferdinand.mom
bce75fd508 enhance parallel context to handle 3D 2024-09-23 10:28:01 +00:00
ferdinand.mom
c36d415b47 add training and generate for pp 2024-09-19 14:06:46 +00:00