[Perf] Tune DeepSeek V4 fused MoE for C1/C2 decode

Select the kw4 stage-1 route for the E385/top-k7 C1 and C2 rows while keeping the existing stage-2 kernels unchanged. Paired same-GPU dispatch measurements improve total stage time by 29.38% at C1 and 22.53% at C2. Fixed-seed strict-reference checks pass without NaNs, and matched FSE-on serving improves output throughput by 8.09% at C1 and 0.85% at C2.

AI assistance was used to tune, validate, benchmark, and document this change.

Co-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Fangzhou Ai <fangzhouai@gmail.com>

Author: Fangzhou-Ai

aiter/configs/model_configs/dsv4_fp8fp4_tuned_fmoe.csv

@@ -39,8 +39,8 @@ gfx950,256,16,7168,1536,384,6,ActivationType.Silu,torch.bfloat16,torch.float8_e4
 gfx950,256,32,7168,1536,384,6,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,336.3454,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_w4_gui_xcd4_fp8,0.0%,173.2903,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic_bnt2_xcd4,0.0%,509.6357,0,0,0,24.89,24888.88,
 gfx950,256,64,7168,1536,384,6,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,673.3172,flydsl_moe1_afp8_wfp4_bf16_t32x64x256_w4_gui_fp8,0.0%,335.4652,flydsl_moe2_afp8_wfp4_bf16_t32x128x256_atomic_bnt2,0.0%,1008.7824,0,0,0,25.15,12574.52,
 gfx950,256,128,7168,1536,384,6,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,690.1582,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_w3_gui_xcd4_fp8,0.0%,342.4631,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic_bnt2_xcd4_persist,0.0%,1032.6213,0,0,0,49.13,12285.56,
-gfx950,256,1,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,18.3042,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_w4_gui_fp8,0.0%,8.7352,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic,0.0%,27.0394,0,0,0,5.7,156767.16,
-gfx950,256,2,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,20.2251,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_w2_gui_fp8,0.0%,12.4074,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic_xcd4,0.0%,32.6325,0,0,0,9.45,129898.46,
+gfx950,256,1,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,12.3290,flydsl_moe1_afp8_wfp4_bf16_t32x64x256_w4_gui_kw4_fp8,0.0%,11.1626,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic,0.0%,23.4916,0,0,0,6.56,180442.80,
+gfx950,256,2,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,13.8011,flydsl_moe1_afp8_wfp4_bf16_t32x64x256_w4_gui_kw4_fp8,0.0%,13.2991,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic_xcd4,0.0%,27.1002,0,0,0,11.38,156416.24,
 gfx950,256,4,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,21.9888,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_gui_fp8,0.0%,17.2479,flydsl_moe2_afp8_wfp4_bf16_t32x256x256_atomic_bnt2_xcd4,0.0%,39.2367,0,0,0,15.71,108035.45,
 gfx950,256,8,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,37.0281,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_w2_gui_fp8,0.0%,25.8309,flydsl_moe2_afp8_wfp4_bf16_t32x128x256_atomic_bnt2_persist,0.0%,62.859,0,0,0,19.62,67437.29,
 gfx950,256,16,7168,512,385,7,ActivationType.Silu,torch.bfloat16,torch.float8_e4m3fn,torch.float4_e2m1fn_x2,QuantType.per_1x32,1,0,32,0,71.2231,flydsl_moe1_afp8_wfp4_bf16_t32x128x256_w2_gui_fp8,0.0%,41.8664,flydsl_moe2_afp8_wfp4_bf16_t32x128x256_atomic_bnt2,0.0%,113.0895,0,0,0,21.81,37485.47,