[Enhancement] MLAPreprocess not support bf16/fp16

[alanhe151220037]

I find out that MLAPreprocess accelerated very well on w8a8_int8. Will this operator support bf16/fp16 in the future? Because some moe models may not quantize their weights in attention layer.

https://github.com/sgl-project/sglang/blob/a2423052f6673256e3f7e2d8a946893f3653cb5d/python/sglang/srt/layers/attention/npu_ops/mla_preprocess.py#L386-L393

[iforgetmyname]

In fact we have a fused op mla_prolog that supports unquanted attention layer, refer to the docs here: https://www.hiascend.com/document/detail/zh/Pytorch/720/apiref/torchnpuCustomsapi/context/torch_npu-npu_mla_prolog_v2.md

we are still working with this op, and we'll be glad if you can help to plug it in to the model

[alanhe151220037]

In fact we have a fused op mla_prolog that supports unquanted attention layer, refer to the docs here: https://www.hiascend.com/document/detail/zh/Pytorch/720/apiref/torchnpuCustomsapi/context/torch_npu-npu_mla_prolog_v2.md

we are still working with this op, and we'll be glad if you can help to plug it in to the model

@iforgetmyname I'm interested to finish this job if the fused op mla_prolog is ready.