Enable mxfp8-mxfp4 group gemm on cutlass

Signed-off-by: Faraz Khoubsirat <58580514+farazkh80@users.noreply.github.com>

include/cutlass/gemm/collective/builders/sm1xx_common.inl

@@ -566,6 +566,8 @@ check_input_datatypes() {
         ((SfVectorSizeA == 32 && cute::is_same_v<KernelScheduleAuto, BuilderScheduleTag>)
       || (SfVectorSizeA == 32 && cute::is_same_v<KernelTmaWarpSpecializedPingpong, BuilderScheduleTag>)
       || (SfVectorSizeA == 32 && cute::is_same_v<KernelTmaWarpSpecializedCooperative, BuilderScheduleTag>)
+      || (SfVectorSizeA == 32 && cute::is_same_v<KernelPtrArrayTmaWarpSpecializedPingpong, BuilderScheduleTag>)
+      || (SfVectorSizeA == 32 && cute::is_same_v<KernelPtrArrayTmaWarpSpecializedCooperative, BuilderScheduleTag>)
       || (SfVectorSizeA == 32 && cute::is_base_of_v<KernelScheduleBlockScaledGemmSm100, BuilderScheduleTag>)
       || (SfVectorSizeA == 32 && cute::is_base_of_v<KernelSchedulePtrArrayBlockScaledGemmSm100, BuilderScheduleTag>)
       || (SfVectorSizeA == 64 && cute::is_base_of_v<KernelScheduleBlockScaledSparseGemmSm100, BuilderScheduleTag>)

test/unit/gemm/device/gemm_testbed_3x_ptr_array.hpp

@@ -2256,12 +2256,16 @@ bool TestSmall(double alpha = 1.0, double beta = 1.0,
   using ElementA = typename Gemm::GemmKernel::ElementA;
   using ElementB = typename Gemm::GemmKernel::ElementB;
   using TiledMma = typename Gemm::GemmKernel::TiledMma;
-  int alignment_bits = 128;
   
   static constexpr bool IsF8F6F4 = cutlass::gemm::collective::detail::is_sm100_mma_f8f6f4<TiledMma, ElementA, ElementB>();
-  alignment_bits = cutlass::detail::get_input_alignment_bits<ElementA, IsF8F6F4>();
-  // For fp4 and fp6 kernels, the min alignment_input is 128 elements, so we don't need to add alignment_input in test problem sizes.
-  int alignment_input = (alignment_bits / cute::sizeof_bits<ElementA>::value == 128) ? 0 : (alignment_bits / cute::sizeof_bits<ElementA>::value);
+  // For fp4 and fp6 kernels, the min alignment_input is 128 elements, so we don't need to add alignment_input in test problem sizes.  
+  int alignment_bits_a = cutlass::detail::get_input_alignment_bits<ElementA, IsF8F6F4>();
+  int alignment_input_a = (alignment_bits_a / cute::sizeof_bits<ElementA>::value == 128) ? 0 : (alignment_bits_a / cute::sizeof_bits<ElementA>::value);
+  
+  int alignment_bits_b = cutlass::detail::get_input_alignment_bits<ElementB, IsF8F6F4>();
+  int alignment_input_b = (alignment_bits_b / cute::sizeof_bits<ElementB>::value == 128) ? 0 : (alignment_bits_b / cute::sizeof_bits<ElementB>::value);
+  
+  int alignment_input = (alignment_input_a == 0 || alignment_input_b == 0) ? 0 : std::max(alignment_input_a, alignment_input_b);
 
   
   if constexpr (apply_alignment_offset) {

test/unit/gemm/device/sm120_blockscaled_tensorop_gemm/CMakeLists.txt

@@ -71,6 +71,7 @@ cutlass_test_unit_gemm_device_add_executable(
 cutlass_test_unit_gemm_device_add_executable(
   cutlass_test_unit_bs_grouped_gemm_device_tensorop_sm120
   sm120_bs_gemm_nvf4_nvf4_f32_nvf4_group_gemm_fusion.cu
+  sm120_bs_gemm_mxf8_mxf4_f32_group_gemm_fusion.cu
 )
 
 endif()

test/unit/gemm/device/sm120_blockscaled_tensorop_gemm/sm120_bs_gemm_mxf8_mxf4_f32_group_gemm_fusion.cu

@@ -0,0 +1,362 @@
+/***************************************************************************************************
+ * Copyright (c) 2025 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+ * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+
+/*! \file
+    \brief Tests for device-wide grouped GEMM interface
+*/
+
+#include <iostream>
+
+#include "cutlass/cutlass.h"
+#include "cute/tensor.hpp"
+#include "cute/atom/mma_atom.hpp"
+
+#include "cutlass/numeric_types.h"
+
+#include "cutlass/gemm/device/gemm_universal_adapter.h"
+#include "cutlass/gemm/kernel/gemm_universal.hpp"
+#include "cutlass/epilogue/collective/collective_builder.hpp"
+#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass/epilogue/collective/default_epilogue.hpp"
+#include "cutlass/epilogue/thread/linear_combination.h"
+#include "cutlass/gemm/dispatch_policy.hpp"
+#include "cutlass/epilogue/thread/activation.h"
+
+#include "../../../common/cutlass_unit_test.h"
+#include "../gemm_testbed_3x_ptr_array.hpp"
+
+
+using namespace cute;
+
+#if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)
+
+// Pingpong kernel schedule
+TEST(SM120_Device_Gemm_e5m2t_e2m1n_e2m1t_tensorop_f32_epilogue_VS32_group_pingpong, row_sf) {
+  using ElementInputA = float_e5m2_t;
+  using ElementInputB = float_e2m1_t;
+  using ElementA = cutlass::mx_float8_t<ElementInputA>;
+  using ElementB = cutlass::mx_float4_t<ElementInputB>;
+  using ElementC = cutlass::half_t;
+  using ElementD = cutlass::float_e2m1_t;
+  using ElementCompute = float;
+  using ElementAccumulator = float;
+  using ElementSF = cutlass::float_ue8m0_t;
+  using ElementSFD  = ElementSF;
+  using ElementAccumulator = float;
+  using GmemLayoutA = cutlass::layout::RowMajor;
+  using GmemLayoutB = cutlass::layout::ColumnMajor;
+  using GmemLayoutC = cutlass::layout::RowMajor;
+  constexpr int SFVectorSize = 32;
+  using TileShape_MNK = Shape<_128,_128,_128>;
+  using ClusterShape_MNK = Shape<_1,_1,_1>;
+
+  constexpr int AlignmentA  = 128 / cutlass::sizeof_bits<ElementInputA>::value;
+  constexpr int AlignmentB  = 128;
+  constexpr int AlignmentC  = 128 / cutlass::sizeof_bits<ElementC>::value;
+  constexpr int AlignmentD  = 128 / cutlass::sizeof_bits<ElementD>::value;  
+
+  //
+  // Construct CollectiveEpilogue
+  //
+
+  constexpr int OutputSFVectorSize = SFVectorSize;
+  // D = alpha * acc + beta * C
+  // With Row-major BlockScaleFactor generation.
+  using FusionOperation = cutlass::epilogue::fusion::LinCombBlockScaleFactor<
+      OutputSFVectorSize,
+      ElementD, 
+      ElementCompute, 
+      ElementSFD, GmemLayoutC,
+      ElementC>;
+
+  using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::epilogue::collective::EpilogueTileAuto,
+      ElementAccumulator, ElementCompute,
+      ElementC, GmemLayoutC *, AlignmentC,
+      ElementD, GmemLayoutC *, AlignmentD,
+      cutlass::epilogue::collective::EpilogueScheduleAuto,
+      FusionOperation
+    >::CollectiveOp;
+
+  //
+  // Construct CollectiveMainloop
+  //
+  using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      ElementA, GmemLayoutA *, AlignmentA,
+      ElementB, GmemLayoutB *, AlignmentB,
+      ElementAccumulator,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpong
+    >::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      cutlass::gemm::GroupProblemShape<Shape<int,int,int>>,
+      CollectiveMainloop,
+      CollectiveEpilogue
+    >;
+  
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  auto pass = test::gemm::device::TestSmallFusion<Gemm>(1.0, 0.5);
+  EXPECT_TRUE(pass);
+}
+
+
+
+TEST(SM120_Device_Gemm_e5m2t_e2m1n_e2m1t_tensorop_f32_epilogue_VS32_group_pingpong, silu_row_sf) {
+  using ElementInputA = float_e5m2_t;
+  using ElementInputB = float_e2m1_t;
+  using ElementA = cutlass::mx_float8_t<ElementInputA>;
+  using ElementB = cutlass::mx_float4_t<ElementInputB>;
+  using ElementC = cutlass::half_t;
+  using ElementD = cutlass::float_e2m1_t;
+  using ElementCompute = float;
+  using ElementAccumulator = float;
+  using ElementSF = cutlass::float_ue4m3_t;
+  using ElementSFD  = ElementSF;
+  using ElementAccumulator = float;
+  using GmemLayoutA = cutlass::layout::RowMajor;
+  using GmemLayoutB = cutlass::layout::ColumnMajor;
+  using GmemLayoutC = cutlass::layout::RowMajor;
+  constexpr int SFVectorSize = 32;
+  using TileShape_MNK = Shape<_128,_128,_128>;
+  using ClusterShape_MNK = Shape<_1,_1,_1>;
+
+  constexpr int AlignmentA  = 128 / cutlass::sizeof_bits<ElementInputA>::value;
+  constexpr int AlignmentB  = 128;
+  constexpr int AlignmentC  = 128 / cutlass::sizeof_bits<ElementC>::value;
+  constexpr int AlignmentD  = 128 / cutlass::sizeof_bits<ElementD>::value;  
+
+  //
+  // Construct CollectiveEpilogue
+  //
+
+  constexpr int OutputSFVectorSize = SFVectorSize;
+  // D = SiLu(alpha * acc + beta * C)
+  // With Row-major BlockScaleFactor generation.
+  using FusionOperation = cutlass::epilogue::fusion::LinCombEltActBlockScaleFactor<
+      cutlass::epilogue::thread::SiLu,
+      OutputSFVectorSize,
+      ElementD, 
+      ElementCompute, 
+      ElementSFD, GmemLayoutC,
+      ElementC>;
+
+  using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::epilogue::collective::EpilogueTileAuto,
+      ElementAccumulator, ElementCompute,
+      ElementC, GmemLayoutC *, AlignmentC,
+      ElementD, GmemLayoutC *, AlignmentD,
+      cutlass::epilogue::collective::EpilogueScheduleAuto,
+      FusionOperation
+    >::CollectiveOp;
+
+  //
+  // Construct CollectiveMainloop
+  //
+  using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      ElementA, GmemLayoutA *, AlignmentA,
+      ElementB, GmemLayoutB *, AlignmentB,
+      ElementAccumulator,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
+      cutlass::gemm::KernelPtrArrayTmaWarpSpecializedPingpong
+    >::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      cutlass::gemm::GroupProblemShape<Shape<int,int,int>>,
+      CollectiveMainloop,
+      CollectiveEpilogue
+    >;
+  
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  auto pass = test::gemm::device::TestSmallFusion<Gemm>(1.0, 0.5);
+  EXPECT_TRUE(pass);
+}
+
+
+// Cooperative kenel schedule
+TEST(SM120_Device_Gemm_e5m2t_e2m1n_e2m1t_tensorop_f32_epilogue_VS32_group_cooperative, row_sf) {
+  using ElementInputA = float_e5m2_t;
+  using ElementInputB = float_e2m1_t;
+  using ElementA = cutlass::mx_float8_t<ElementInputA>;
+  using ElementB = cutlass::mx_float4_t<ElementInputB>;
+  using ElementC = cutlass::half_t;
+  using ElementD = cutlass::float_e2m1_t;
+  using ElementCompute = float;
+  using ElementAccumulator = float;
+  using ElementSF = cutlass::float_ue4m3_t;
+  using ElementSFD  = ElementSF;
+  using ElementAccumulator = float;
+  using GmemLayoutA = cutlass::layout::RowMajor;
+  using GmemLayoutB = cutlass::layout::ColumnMajor;
+  using GmemLayoutC = cutlass::layout::RowMajor;
+  constexpr int SFVectorSize = 32;
+  using TileShape_MNK = Shape<_128,_128,_128>;
+  using ClusterShape_MNK = Shape<_1,_1,_1>;
+
+  constexpr int AlignmentA  = 128 / cutlass::sizeof_bits<ElementInputA>::value;
+  constexpr int AlignmentB  = 128;
+  constexpr int AlignmentC  = 128 / cutlass::sizeof_bits<ElementC>::value;
+  constexpr int AlignmentD  = 128 / cutlass::sizeof_bits<ElementD>::value;  
+
+  //
+  // Construct CollectiveEpilogue
+  //
+
+  constexpr int OutputSFVectorSize = SFVectorSize;
+  // D = alpha * acc + beta * C
+  // With Row-major BlockScaleFactor generation.
+  using FusionOperation = cutlass::epilogue::fusion::LinCombBlockScaleFactor<
+      OutputSFVectorSize,
+      ElementD, 
+      ElementCompute, 
+      ElementSFD, GmemLayoutC,
+      ElementC>;
+
+  using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::epilogue::collective::EpilogueTileAuto,
+      ElementAccumulator, ElementCompute,
+      ElementC, GmemLayoutC *, AlignmentC,
+      ElementD, GmemLayoutC *, AlignmentD,
+      cutlass::epilogue::collective::EpilogueScheduleAuto,
+      FusionOperation
+    >::CollectiveOp;
+
+  //
+  // Construct CollectiveMainloop
+  //
+  using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      ElementA, GmemLayoutA *, AlignmentA,
+      ElementB, GmemLayoutB *, AlignmentB,
+      ElementAccumulator,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
+      cutlass::gemm::collective::KernelScheduleAuto
+    >::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      cutlass::gemm::GroupProblemShape<Shape<int,int,int>>,
+      CollectiveMainloop,
+      CollectiveEpilogue
+    >;
+  
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  auto pass = test::gemm::device::TestSmallFusion<Gemm>(1.0, 0.5);
+  EXPECT_TRUE(pass);
+}
+
+
+
+TEST(SM120_Device_Gemm_e5m2t_e2m1n_e2m1t_tensorop_f32_epilogue_VS32_group_cooperative, silu_row_sf) {
+  using ElementInputA = float_e5m2_t;
+  using ElementInputB = float_e2m1_t;
+  using ElementA = cutlass::mx_float8_t<ElementInputA>;
+  using ElementB = cutlass::mx_float4_t<ElementInputB>;
+  using ElementC = cutlass::half_t;
+  using ElementD = cutlass::float_e2m1_t;
+  using ElementCompute = float;
+  using ElementAccumulator = float;
+  using ElementSF = cutlass::float_ue4m3_t;
+  using ElementSFD  = ElementSF;
+  using ElementAccumulator = float;
+  using GmemLayoutA = cutlass::layout::RowMajor;
+  using GmemLayoutB = cutlass::layout::ColumnMajor;
+  using GmemLayoutC = cutlass::layout::RowMajor;
+  constexpr int SFVectorSize = 32;
+  using TileShape_MNK = Shape<_128,_128,_128>;
+  using ClusterShape_MNK = Shape<_1,_1,_1>;
+
+  constexpr int AlignmentA  = 128 / cutlass::sizeof_bits<ElementInputA>::value;
+  constexpr int AlignmentB  = 128;
+  constexpr int AlignmentC  = 128 / cutlass::sizeof_bits<ElementC>::value;
+  constexpr int AlignmentD  = 128 / cutlass::sizeof_bits<ElementD>::value;  
+
+  //
+  // Construct CollectiveEpilogue
+  //
+
+  constexpr int OutputSFVectorSize = SFVectorSize;
+  // D = SiLu(alpha * acc + beta * C)
+  // With Row-major BlockScaleFactor generation.
+  using FusionOperation = cutlass::epilogue::fusion::LinCombEltActBlockScaleFactor<
+      cutlass::epilogue::thread::SiLu,
+      OutputSFVectorSize,
+      ElementD, 
+      ElementCompute, 
+      ElementSFD, GmemLayoutC,
+      ElementC>;
+
+  using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::epilogue::collective::EpilogueTileAuto,
+      ElementAccumulator, ElementCompute,
+      ElementC, GmemLayoutC *, AlignmentC,
+      ElementD, GmemLayoutC *, AlignmentD,
+      cutlass::epilogue::collective::EpilogueScheduleAuto,
+      FusionOperation
+    >::CollectiveOp;
+
+  //
+  // Construct CollectiveMainloop
+  //
+  using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
+      cutlass::arch::Sm120, cutlass::arch::OpClassBlockScaledTensorOp,
+      ElementA, GmemLayoutA *, AlignmentA,
+      ElementB, GmemLayoutB *, AlignmentB,
+      ElementAccumulator,
+      TileShape_MNK, ClusterShape_MNK,
+      cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
+      cutlass::gemm::collective::KernelScheduleAuto
+    >::CollectiveOp;
+
+  using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
+      cutlass::gemm::GroupProblemShape<Shape<int,int,int>>,
+      CollectiveMainloop,
+      CollectiveEpilogue
+    >;
+  
+  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
+  auto pass = test::gemm::device::TestSmallFusion<Gemm>(1.0, 0.5);
+  EXPECT_TRUE(pass);
+}
+#endif // #if defined(CUTLASS_ARCH_MMA_SM120_SUPPORTED)