Experience with Triton

[sarvghotra]

I see that DeepSpeed used Triton to implement block-sparse attention.

1. Was wondering if anyone from DeepSpeed could comment on the experience they had with Triton.
2. Will DeepSpeed continue to use it in the future in their implementations?

Context: I am trying to decide if I should invest in Triton for custom Kernel implementations. From their GitHub page, it looks like it may not be the right time yet to depend on it. (I may be totally wrong here, hence asking).