update moe ffn tune config (#72583)

Author: chang-wenbin

paddle/phi/kernels/fusion/cutlass/cutlass_extensions/ft_gemm_configs.h

@@ -48,16 +48,18 @@ enum class CutlassTileConfig {
   // Warp configs for M=16
   CtaShape16x128x64_WarpShape16x32x64,
   CtaShape16x256x64_WarpShape16x64x64,
-
+  CtaShape16x256x64_WarpShape64x16x128,
   // Warp configs for M=32
   CtaShape32x128x64_WarpShape32x32x64,
 
   // Warp configs for M=64
+  CtaShape64x64x64_WarpShape32x32x64,
   CtaShape64x128x64_WarpShape32x64x64,
   CtaShape64x128x64_WarpShape64x32x64,
   CtaShape64x128x64_WarpShape64x64x64,
 
   // Warp configs for M=128
+  CtaShape128x64x64_WarpShape64x32x64,
   CtaShape128x128x64_WarpShape64x32x64,
   CtaShape128x128x64_WarpShape64x64x64,
   CtaShape128x128x64_WarpShape128x32x64,

paddle/phi/kernels/fusion/cutlass/cutlass_kernels/cutlass_heuristic.h

@@ -60,9 +60,14 @@ static std::vector<CutlassTileConfig> get_candidate_tiles(
   };
   std::vector<CutlassTileConfig> quant_B_configs_sm80{
       CutlassTileConfig::CtaShape16x128x64_WarpShape16x32x64,
+      CutlassTileConfig::CtaShape16x256x64_WarpShape16x64x64,
       CutlassTileConfig::CtaShape32x128x64_WarpShape32x32x64,
+      CutlassTileConfig::CtaShape64x64x64_WarpShape32x32x64,
+      CutlassTileConfig::CtaShape64x128x64_WarpShape64x32x64,
       CutlassTileConfig::CtaShape64x128x64_WarpShape64x64x64,
+      CutlassTileConfig::CtaShape128x64x64_WarpShape64x32x64,
       CutlassTileConfig::CtaShape128x128x64_WarpShape64x64x64,
+      CutlassTileConfig::CtaShape128x128x64_WarpShape128x32x64,
       CutlassTileConfig::CtaShape128x256x64_WarpShape64x64x64,
   };
   if (is_moe) {