Replace mamba1 with mamba2 and training becomes very slow!

[YQ-097]

@torch.compile(options={"triton.cudagraphs": True}, fullgraph=True) generates an error. Is there any other way?

[tridao]

If you use a large model the triton overhead will be neglibile.

[dragonBrother1]

@torch.compile（options={“triton.cudagraphs”： True}， fullgraph=True） 生成错误。还有其他方法吗？

I encounted some questions when i chose to value mamba2 instead of mamba.

Dose it mean that i should vary the MambaConfig?