Request for Megatron benchmark harness

[sbhavani]

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.

[sbhavani]

Also the results supporting “outperforms FA3 / on par with FA4” doesn't seem to match with the current benchmark code. It appears to compare TE, FA2, and fa4_cute, rather than the custom implementation against FA3 and FA4.

A same-shape comparison with pinned versions, correctness checks, and timings would be helpful.