Use flash attention for qwen3 models

[guoqingbao]

No description provided.

[DrJesseGlass]

I'll review this PR to make sure that I built along the same lines however I had integrated GPU flash attention and CPU flash attention into Qwen3 (but not MoE) but as I was doing it found substantial reason to upgrade the CPU flash attention and made an early simple submission #3254 to get an improved CPU flash structure before I started serious optimizations.

However, someone else was simultaneously working on #3250 for integrating varlen and the agreement at the time was once varlen was integrated that I would tweak mine to incorporate but it's been sitting unfinished.

[guoqingbao]

I'll review this PR to make sure that I built along the same lines however I had integrated GPU flash attention and CPU flash attention into Qwen3 (but not MoE) but as I was doing it found substantial reason to upgrade the CPU flash attention and made an early simple submission #3254 to get an improved CPU flash structure before I started serious optimizations.

This uses the previously defined flash attention interface and remains compatible with existing implementations.

[DrJesseGlass]

Yes. It makes sense to get this integrated sooner if possible. I just meant to call out the redundancy.

[guoqingbao]

Yes. It makes sense to get this integrated sooner if possible. I just meant to call out the redundancy.

It would be better to have a unified entry point for varlen attention (CPU and GPU). I think I can help with GPU varlen attention, but it seems we are missing corresponding use cases in candle. Varlen attention requires parallel requests (though it can work with a single request) to meaningfully demonstrate performance.