CUTLASS 3.6.0 (#1850)

* v3.6

* update changelog

* update readme

* fix typo

* fixing typos

* hopper gemm with weight prefetch

---------

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

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

@@ -113,6 +113,12 @@ template <> struct ArchMap<arch::Sm90, arch::OpClassTensorOp> {
   static int const kMax = 90;
 };
 
+// Arch conditional sparse WGMMA
+template <> struct ArchMap<arch::Sm90, arch::OpClassSparseTensorOp> {
+  static int const kMin = 90;
+  static int const kMax = 90;
+};
+
 /////////////////////////////////////////////////////////////////////////////////////////////////
 
 } // namespace library

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

@@ -103,6 +103,17 @@ public:
     void *device_workspace = nullptr,
     cudaStream_t stream = nullptr) const = 0;
 
+  // Originally designed for metadata, but should be useful for FP8/6/4 too.
+  virtual Status initialize_with_profiler_workspace(
+    void const *configuration,
+    void *host_workspace,
+    void *device_workspace,
+    uint8_t **profiler_workspace_ptrs,
+    int problem_count,
+    cudaStream_t stream = nullptr) {
+    return Status::kErrorNotSupported;
+  }
+
   virtual Status run(
     void const *arguments,
     void *host_workspace,
@@ -290,7 +301,6 @@ struct GemmUniversalArguments {
 
   // Needed for some 3.x kernels
   int sm_count{0};
-
   library::RasterOrder raster_order{};
   int swizzle_size{1};
 };