update 3.8 v2 (#2112)

* update 3.8 v2

* update 3.8

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>

tools/library/include/cutlass/library/descriptions.h

@@ -35,6 +35,8 @@
 #include <cutlass/blas3_types.h>
 #include <cutlass/gemm_coord.h>
 
+#include <optional>
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
 namespace cutlass {
@@ -300,13 +302,34 @@ struct GemmDescription : public OperationDescription {
     transform_B(transform_B) {}
 };
 
+struct BlockScaleDescription {
+  /// Describes the SFA operand
+  TensorDescription SFA;
+
+  /// Describes the SFB operand
+  TensorDescription SFB;
+
+  /// Describes the SFD operand
+  TensorDescription SFD;
+
+  /// Describes the input ScaleFactor VectorSize
+  int SFVecSize;
+
+  /// Describes the Output ScaleFactor VectorSize
+  int EpilogueSFVecSize;
+};
+
+struct GroupedGemmDescription : public OperationDescription {
+  GemmDescription gemm;
+  std::optional<BlockScaleDescription> block_scales;
+};
 
 /// Description of all GEMM computations
 struct BlockScaledGemmDescription : public OperationDescription {
 
   /// Indicates the kind of GEMM performed
   GemmKind gemm_kind;
-  
+
   /// Describes the A operand
   TensorDescription A;
 

tools/library/include/cutlass/library/library.h

@@ -336,11 +336,10 @@ struct GemmUniversalArguments {
   void *packed_Scale{nullptr};          // Packed scale for int4 * fp8
 
   int device_index{0};
-  
+
   bool use_pdl{false};
 };
 
-
 /// Block Scaled GEMM
 //
 // OperationKind: kBlockScaledGemm
@@ -495,29 +494,31 @@ struct GemmGroupedConfiguration {
   int64_t* lda;
   int64_t* ldb;
   int64_t* ldc;
+
+  cute::Shape<int, int, int>* problem_sizes_3x_host;
 };
 
 struct GemmGroupedArguments {
   int problem_count{};
   gemm::GemmCoord* problem_sizes{nullptr};
 
-  void * ptr_A{nullptr};
-  void * ptr_B{nullptr};
-  void * ptr_C{nullptr};
-  void * ptr_D{nullptr};
+  void* ptr_A{nullptr};
+  void* ptr_B{nullptr};
+  void* ptr_C{nullptr};
+  void* ptr_D{nullptr};
 
-  int64_t *lda{nullptr};
-  int64_t *ldb{nullptr};
-  int64_t *ldc{nullptr};
-  int64_t *ldd{nullptr};
+  int64_t* lda{nullptr};
+  int64_t* ldb{nullptr};
+  int64_t* ldc{nullptr};
+  int64_t* ldd{nullptr};
 
   void const *alpha{nullptr};
   void const *beta{nullptr};
   ScalarPointerMode pointer_mode{};
   bool use_pdl{false};
 
-  gemm::GemmCoord cluster_shape{};          
-  gemm::GemmCoord cluster_shape_fallback{}; 
+  gemm::GemmCoord cluster_shape{};
+  gemm::GemmCoord cluster_shape_fallback{};
 
   // these should really be in the configuration but staying consistent with GEMM
   int sm_count{0};
@@ -529,6 +530,13 @@ struct GemmGroupedArguments {
   cute::Shape<int, int, int>* problem_sizes_3x_host;
 };
 
+struct GroupedGemmBlockScaledArguments : GemmGroupedArguments {
+  void* SFA{nullptr};
+  void* SFB{nullptr};
+  void* SFD{nullptr};
+  void* norm_constant{nullptr};
+};
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
 //
 // OperationKind: kSparseGemm