Any comparison between mega-triton-kernel & sglang/vllm?

https://github.com/ByteDance-Seed/Triton-distributed/blob/main/docs/mega_triton_kernel.md

Does this mega triton kernel implementation provide critical inference engine caps, say:
- paged attention
- continuous batching
- zero scheduling overheads

What's the goal of mega-triton-kernel? Is it aimed to replace the existing vllm/sglang op+graph based impl?

Start with docs/mega_triton_kernel.md and review the documented scope of mega-triton-kernel. Compare its stated capabilities with paged attention, continuous batching, and scheduling overhead, then document whether those capabilities are provided and clarify its relationship to vLLM and SGLang.

Written by the indexing model from the issue text.