Thanks for open sourcing ForgeTrain! Could you share the benchmark harness used for the reported ~10% MFU improvement over Megatron-LM?
The env details (Megatron-LM commit, configs and launch scripts), hardware topology, and measurement methodology would help enable with validation when using this new library.
Thanks for open sourcing ForgeTrain! Could you share the benchmark harness used for the reported ~10% MFU improvement over Megatron-LM?
The env details (Megatron-LM commit, configs and launch scripts), hardware topology, and measurement methodology would help enable with validation when using this new library.