[feat]: CUTLASS block scaled group gemm for SM100 (vllm-project#19757)

Signed-off-by: Duncan Moss <djm.moss@gmail.com>
Co-authored-by: Duncan Moss <dmoss@nvidia.com>

Author: djmmoss

CMakeLists.txt

@@ -259,7 +259,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v3.9.2" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v4.0.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -615,6 +615,26 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                      "in CUDA target architectures.")
     endif()
   endif()
+  
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
+    message(STATUS "Building blockwise_scaled_group_mm_sm100 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 kernels as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or later "
+                     "if you intend on running FP8 quantized MoE models on Blackwell.")
+    else()
+      message(STATUS "Not building blockwise_scaled_group_mm_sm100 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
 
   #
   # Machete kernels

csrc/cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp

@@ -45,7 +45,6 @@
 #include "cute/algorithm/functional.hpp"
 #include "cute/atom/mma_atom.hpp"
 #include "cute/algorithm/gemm.hpp"
-#include "cute/tensor_predicate.hpp"
 #include "cute/numeric/arithmetic_tuple.hpp"
 
 #include "cutlass_extensions/gemm/dispatch_policy.hpp"

csrc/ops.h

@@ -239,6 +239,11 @@ void cutlass_moe_mm(
     torch::Tensor const& b_strides, torch::Tensor const& c_strides,
     bool per_act_token, bool per_out_ch);
 
+void cutlass_blockwise_scaled_grouped_mm(
+    torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
+    const torch::Tensor& scales_a, const torch::Tensor& scales_b,
+    const torch::Tensor& problem_sizes, const torch::Tensor& expert_offsets);
+
 void cutlass_fp4_group_mm(
     torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
     const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,

csrc/quantization/cutlass_w8a8/c3x/scaled_mm.cuh

@@ -51,7 +51,8 @@ struct cutlass_3x_gemm {
   // These are the minimum alignments needed for the kernels to compile
   static constexpr int AlignmentAB =
       128 / cutlass::sizeof_bits<ElementAB>::value;
-  static constexpr int AlignmentCD = 4;
+  static constexpr int AlignmentCD =
+      128 / cutlass::sizeof_bits<ElementD>::value;
 
   using CollectiveEpilogue =
       typename cutlass::epilogue::collective::CollectiveBuilder<