add deepseek mhc pre

[qescccczmr]

N=8192 H=4096 M=4 BH=1024 NH=4 n_splits=1 sinkhorn=10
── TileLang vs Ref ── ── Triton-TLX vs Ref ──
Output max_abs mean_abs max_rel max_abs mean_abs max_rel
────────────── ────────── ────────── ──────────
post_mix 5.96e-08 1.35e-10 1.37e-07 1.19e-07 1.02e-08 2.38e-07
comb_mix 8.94e-08 1.23e-08 4.21e-07 1.19e-07 1.88e-08 4.96e-07
layer_input 1.56e-02 2.56e-08 4.45e-01 1.56e-02 3.15e-08 6.60e-01

TileLang vs Triton-TLX:
Output max_abs mean_abs max_rel
────────────── ────────── ────────── ──────────
post_mix 1.19e-07 1.02e-08 2.38e-07
comb_mix 1.04e-07 1.79e-08 4.49e-07
layer_input 1.56e-02 1.78e-08 7.05e-01

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

chengqinxiu seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[kateyijian]

mo变量命名改下，以及下面存在同样问题