rebase fix?

neuralmagic · Dec 7, 2024 · 96e4a7a · 96e4a7a
1 parent e5c6d10
commit 96e4a7a
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 2 deletions.
diff --git a/csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp b/csrc/cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp
@@ -36,13 +36,13 @@ struct ScaledEpilogueBase {
   // Don't want to support nullptr by default
   template <typename T, bool EnableNullPtr = false>
   using ColLoad = cutlass::epilogue::fusion::Sm90ColBroadcast<
-      0 /*Stages*/, typename EpilogueDescriptor::TileShape, T,
+      0 /*Stages*/, typename EpilogueDescriptor::TileShape, T, T,
       Stride<Int<1>, Int<0>, Int<0>>, 128 / sizeof_bits_v<T>, EnableNullPtr>;
 
   // Don't want to support nullptr by default
   template <typename T, bool EnableNullPtr = false>
   using RowLoad = cutlass::epilogue::fusion::Sm90RowBroadcast<
-      0 /*Stages*/, typename EpilogueDescriptor::TileShape, T,
+      0 /*Stages*/, typename EpilogueDescriptor::TileShape, T, T,
       Stride<Int<0>, Int<1>, Int<0>>, 128 / sizeof_bits_v<T>, EnableNullPtr>;
 
   // This utility function constructs the arguments for the load descriptors

diff --git a/csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu b/csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu
@@ -1,5 +1,17 @@
 #include <stddef.h>
+
+#include <cudaTypedefs.h>
+
+#if defined CUDA_VERSION && CUDA_VERSION >= 12000
+
 #include <torch/all.h>
+
+#include <ATen/cuda/CUDAContext.h>
+
+#include <iostream>
+#include <sstream>
+#include <vector>
+
 #include "cutlass/cutlass.h"
 #include "scaled_mm_c3x.cuh"