Skip to content

ETOgaosion/Megatron-LM-AutoTuner

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

46 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AutoTuner for Megatron + TransformerEngine

Scenarios

This is a practical auto-tuner on Megatron targeted at post-training frameworks like verl project.

Our performance tuning target is MFU in MCore training process on both forward-only models and forward-backward-update models, which leads to high performance on training side in RLHF.

Speaking of tuning dimensions, currently include:

Dense Layer TP CP DP PP VPP
MoE Parallel Folding ETP EP EDP
Pipeline layout
Seqlen max_token_len
Recompute recompute_granularity recompute_method recompute_num_layers recompute_modules

Target shapes:

  • thd (mainly)
  • bshd

Docs

Doc in Lark

Go to docs directory

About

No description, website, or topics provided.

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 5