Optimize SM90 MegaMoE split path

[liz-badada]

Summary

• Enable optimized SM90 MegaMoE split L1/L2 path.
• Centralize retained SM90 split heuristics.
• Add split-vs-one correctness and benchmark support.

Performance

M	Final us	vs DeepEP	vs Original	vs Prev two-kernel
8	745.5	2.04x	1.26x	1.26x
16	769.1	2.11x	1.44–1.60x	1.50x
32	760.6	2.17x	1.50x	1.70x
64	791.0	2.08x	1.45–1.51x	1.53x
128	818.5	2.04x	1.28x	1.50x
256	1121.3	1.55x	1.43x	1.45x
260	1222.0	1.42x	1.30x	1.27x
512	1921.7	1.32–1.33x	1.26–1.43x	1.34x
819	2798.4	1.20x	1.38–1.42x	1.41x
1024	3278.0	1.29x	1.34–1.52x	0.99x
2048	6098.0	1.20x	1.34–1.37x	0.98x
4096	11430.0	1.16x	1.37–1.38x	1.00x
8192	22286.0	1.14x	1.39x	1.02x

H20

M	latency us	achieved TFLOPS	TC SOL	roofline bound us	roofline SOL	bound
8	745.5	7.6	2.6%	255.5	34.3%	HBM/overhead
16	769.1	14.7	5.0%	289.3	37.6%	HBM/overhead
32	760.6	29.6	10.0%	295.0	38.8%	HBM/overhead
64	791.0	57.0	19.3%	296.5	37.5%	HBM/overhead
128	818.5	110.2	37.2%	304.7	37.2%	compute
256	1121.3	160.9	54.3%	609.4	54.3%	compute
260	1222.0	149.9	50.7%	618.9	50.7%	compute
512	1921.7	187.7	63.4%	1218.8	63.4%	compute
819	2798.4	206.2	69.7%	1949.7	69.7%	compute
1024	3278.0	220.1	74.4%	2437.7	74.4%	compute
2048	6098.0	236.7	80.0%	4875.4	80.0%	compute
3072	8746.0	247.5	83.6%	7313.1	83.6%	compute
4096	11430.0	252.5	85.3%	9750.7	85.3%	compute
8192	22286.0	259.0	87.5%	19501.5	87.5%	compute

Correctness

Case	Result
Main random M8/M16/M32/M64/M128/M256/M260/M512/M819/M1024/M2048/M3072/M4096/M8192	max_abs=0
Hopper random M512/M1024/M2048/M3072/M4096	max_abs=0
Main power-law M1024 / masked M1024 / EPLB hash M2048	max_abs=0
Hopper power-law M2048	max_abs=0