SGLang int8 kernels

[vadimkantorov]

I wonder if these Triton kernels are any relevant for wider torchao / pytorch usage (and if this Triton impl is also any portable for CPU):

• [Feature] DeepSeek V3/R1 INT8 Quantization (channel-wise)  sgl-project/sglang#3888
• Feature DeepSeek V3/R1 INT8 Quantization (block-wise) sgl-project/sglang#3730

and if not - I wonder why sglang does not use the quant triton kernels/bindings from ao?

(Also similar question on liger / unsloth kernels - including the notorious rmsnorm kernels - any plans to upstream their main components like linear + chunked cross entropy some place upstream?)