Which attention is enough for optimal performance

[Rogala]

Currently, when working with ComfyUI, we have 4 accelerators:

1. Torch_attention (as Gemini says it is part of the FlashAttention code)
2. xFormers_attention (as Gemini says it contains part of the FlashAttention2 code)
3. FlashAttention 2 (independent version)
4. SageAttention

• Therefore, the question has arisen after some tests, xFormers_attention is clearly faster than the standard Torch_attention
• FlashAttention 2 cannot work with xFormers as it already has its code, cannot work with SageAttention
• xFormers_attention can work together with SageAttention and this pair has the highest performance

Question:
So is the ability in a separate FlashAttention 2 when it is part of xFormers

Do I need xFormers itself, or is Torch_attention+SageAttention enough?

What's the point of installing pure FlashAttention 2 where it can show itself?