We need to tune LLAMA3 405B FP8-CS performance on GB300, some of the optimization to enable:

1. Gradient Accumulation Fusion - @dingqingy-nv
2. Low SM NVLink communication - @dingqingy-nv (Please take help from Youngeun on this)
3. Try TP1 by using more memory on GB300 GPU - @malay-nagda
4. If TP1 doesn't work the way we want it above, try Hybrid-FSDP - @malay-nagda

cc: @erhoo82