On: Mamba: The Easy Way

[espadrine]

Another element is that Mamba required a very custom implementation down to custom fused kernels which I expect would need to be implemented in deepspeed or the equivalent library for a larger training run spanning thousands of GPUs.

[cs702 → espadrine]

Not necessarily:

https://www.reddit.com/r/MachineLearning/comments/1amb3xu/d_...