Bump Cutlass version to 2.11.0 (#50073)

cmake/external/cutlass.cmake

@@ -17,7 +17,7 @@ include(ExternalProject)
 set(CUTLASS_PREFIX_DIR ${THIRD_PARTY_PATH}/cutlass)
 
 set(CUTLASS_REPOSITORY https://github.com/NVIDIA/cutlass.git)
-set(CUTLASS_TAG v2.10.0)
+set(CUTLASS_TAG v2.11.0)
 
 include_directories("${THIRD_PARTY_PATH}/cutlass/src/extern_cutlass/")
 include_directories("${THIRD_PARTY_PATH}/cutlass/src/extern_cutlass/include/")

paddle/phi/kernels/fusion/cutlass/moe_cutlass_kernel.h → paddle/phi/kernels/fusion/cutlass/moe/moe_cutlass_kernel.h

@@ -42,6 +42,7 @@
 #include "cutlass/gemm/kernel/grouped_problem_visitor.h"
 #include "cutlass/layout/matrix.h"
 #include "cutlass/trace.h"
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
 namespace cutlass {
@@ -350,14 +351,16 @@ template <typename ThreadblockShape,
           int ThreadCount,
           bool Transposed = false>
 struct GemmMoeProblemVisitor
-    : public MoeProblemVisitor<detail::GemmGroupedProblemSizeHelper<Transposed>,
+    : public MoeProblemVisitor<
+          detail::GemmGroupedProblemSizeHelper<ThreadblockShape, Transposed>,
           ThreadblockShape,
           GroupScheduleMode_,
           PrefetchTileCount,
           ThreadCount> {
   static bool const kTransposed = Transposed;
 
-  using ProblemSizeHelper = detail::GemmGroupedProblemSizeHelper<Transposed>;
+  using ProblemSizeHelper =
+      detail::GemmGroupedProblemSizeHelper<ThreadblockShape, Transposed>;
   using Base = MoeProblemVisitor<ProblemSizeHelper,
                                  ThreadblockShape,
                                  GroupScheduleMode_,

paddle/phi/kernels/fusion/cutlass/moe_kernel.cu

@@ -17,7 +17,7 @@
 #include "paddle/phi/core/dense_tensor.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/funcs/elementwise_base.h"
-#include "paddle/phi/kernels/fusion/cutlass/moe_kernel_impl.h"
+#include "paddle/phi/kernels/fusion/cutlass/moe/moe_kernel_impl.h"
 
 // Ignore CUTLASS warnings about type punning
 #pragma GCC diagnostic push
@@ -32,13 +32,15 @@
 #include "cutlass/gemm/kernel/default_gemm_grouped.h"
 #include "cutlass/numeric_conversion.h"
 #include "paddle/phi/backends/gpu/gpu_info.h"
-#include "paddle/phi/kernels/fusion/cutlass/default_moe_fc_traits.h"
-#include "paddle/phi/kernels/fusion/cutlass/linear_combination_ft_gelu.h"
-#include "paddle/phi/kernels/fusion/cutlass/moe_cutlass_kernel.h"
+#include "paddle/phi/kernels/fusion/cutlass/moe/default_moe_fc_traits.h"
+#include "paddle/phi/kernels/fusion/cutlass/moe/linear_combination_ft_gelu.h"
+#include "paddle/phi/kernels/fusion/cutlass/moe/moe_cutlass_kernel.h"
 #pragma GCC diagnostic pop
+
 namespace phi {
 
 namespace {
+
 inline int getSMVersion() {
   const int device = phi::backends::gpu::GetCurrentDeviceId();
   const phi::gpuDeviceProp prop =