DeepGEMM Mega-mHC beats the TileLang fused path 1.14x to 1.51x.

PR #56962, merged September 15, reopens a closed PR rebased onto the DeepGEMM fork pin and retargets the model path to deepseek_v41 (GitHub). CUPTI spans on a single GB200 at H=5120/7168 with hc_mult=4 show Mega-mHC ahead by 1.14x at one decode token and 1.51x at 64, widening as H grows. End to end on GB200 TP4 with official DeepSeek-V4.1-Flash and fp8_ds_mla, all 504 requests completed: at concurrency 32 and 32k input it delivers 2.9% lower mean TPOT and 4.2% higher output tok/s than TileLang, and 3.7% to 5.0% more output tok/s than the competing Overlap approach.