submission 154601

from typing import Type, Tuple, Union

import cutlass
import cutlass.cute as cute
from cutlass.cute.nvgpu import cpasync, tcgen05
import cutlass.utils as utils
import cutlass.pipeline as pipeline
import cutlass.utils.blackwell_helpers as sm100_utils
import cutlass.utils.blockscaled_layout as blockscaled_utils
from cutlass.cute.runtime import from_dlpack, make_ptr
from task import input_t, output_t

class Sm100BlockScaledPersistentDenseGemmKernel:
    """This class implements batched matrix multiplication (C = A x SFA x B x SFB) with support for various data types
    and architectural features specific to Blackwell GPUs with persistent tile scheduling and warp specialization.

    :param sf_vec_size: Scalefactor vector size.
    :type sf_vec_size: int
    :param mma_tiler_mn: Shape of the Matrix Multiply-Accumulate (MMA) tile (M,N)
    :type mma_tiler_mn: Tuple[int, int]
    :param cluster_shape_mn: Cluster dimensions (M,N) for parallel processing
    :type cluster_shape_mn: Tuple[int, int]

    :note: In current version, A and B tensor must have the same data type

    :note: Supported combinations of A/B data types, SF data typs and SF vector size:
        - NVF4: A/B: Float4E2M1FN + SF: Float8E4M3FN + sf_vec_size: 16

    :note: Supported accumulator data types:
        - Float32

    :note: Supported C data types:
        - Float32
        - Float16/BFloat16

    :note: Constraints:
        - MMA tiler M must be 128 or 256 (use_2cta_instrs)
        - MMA tiler N must be 64/128/192/256
        - Cluster shape M must be multiple of 2 if Mma tiler M is 256
        - Cluster shape M/N must be positive and power of 2, total cluster size <= 16
        - Also, Cluster shape M/N must be <= 4 for scale factor multicasts due to limited size of scale factors

    Example:
        >>> gemm = Sm100BlockScaledPersistentDenseGemmKernel(
        ...     sf_vec_size=16,
        ...     mma_tiler_mn=(256, 128),
        ...     cluster_shape_mn=(2, 1)
        ... )
    """

    def __init__(
        self,
        sf_vec_size: int,
        mma_tiler_mn: Tuple[int, int],
        cluster_shape_mn: Tuple[int, int],
    ):
        """Initializes the configuration for a Blackwell dense GEMM kernel.

        This configuration includes several key aspects:

        1.  MMA Instruction Settings (tcgen05):
            - acc_dtype: Data types for MMA accumulator, always set to Float32
            - sf_vec_size: Scalefactor A/B vector size.
            - mma_tiler_mn: The (M, N) shape of the MMA instruction tiler.

        2.  Cluster Shape:
            - cluster_shape_mn: The (ClusterM, ClusterN) shape of the CTA cluster.

        :param sf_vec_size: Scalefactor vector size.
        :type sf_vec_size: int
        :param mma_tiler_mn: Tuple (M, N) shape of the MMA instruction.
        :type mma_tiler_mn: Tuple[int, int]
        :param cluster_shape_mn: Tuple (ClusterM, ClusterN) shape of the cluster.
        :type cluster_shape_mn: Tuple[int, int]
        """

        self.acc_dtype = cutlass.Float32
        self.sf_vec_size = sf_vec_size
        self.use_2cta_instrs = mma_tiler_mn[0] == 256
        self.cluster_shape_mn = cluster_shape_mn
        # K dimension is deferred in _setup_attributes
        self.mma_tiler = (*mma_tiler_mn, 1)

        self.cta_group = (
            tcgen05.CtaGroup.TWO if self.use_2cta_instrs else tcgen05.CtaGroup.ONE
        )

        self.occupancy = 1
        # Set specialized warp ids
        self.epilog_warp_id = (
            0,
            1,
            2,
            3,
        )
        self.mma_warp_id = 4
        self.tma_warp_id = 5
        self.threads_per_cta = 32 * len(
            (self.mma_warp_id, self.tma_warp_id, *self.epilog_warp_id)
        )
        # Set barrier id for cta sync, epilogue sync and tmem ptr sync
        self.cta_sync_barrier = pipeline.NamedBarrier(
            barrier_id=1,
            num_threads=self.threads_per_cta,
        )
        self.epilog_sync_barrier = pipeline.NamedBarrier(
            barrier_id=2,
            num_threads=32 * len(self.epilog_warp_id),
        )
        self.tmem_alloc_barrier = pipeline.NamedBarrier(
            barrier_id=3,
            num_threads=32 * len((self.mma_warp_id, *self.epilog_warp_id)),
        )
        self.smem_capacity = utils.get_smem_capacity_in_bytes("sm_100")
        SM100_TMEM_CAPACITY_COLUMNS = 512
        self.num_tmem_alloc_cols = SM100_TMEM_CAPACITY_COLUMNS

    def _setup_attributes(self):
        """Set up configurations that are dependent on GEMM inputs

        This method configures various attributes based on the input tensor properties
        (data types, leading dimensions) and kernel settings:
        - Configuring tiled MMA
        - Computing MMA/cluster/tile shapes
        - Computing cluster layout
        - Computing multicast CTAs for A/B/SFA/SFB
        - Computing epilogue subtile
        - Setting up A/B/SFA/SFB/C stage counts in shared memory
        - Computing A/B/SFA/SFB/C shared memory layout
        """
        # Compute mma instruction shapes
        # (MMA_Tile_Shape_M, MMA_Tile_Shape_N, MMA_Inst_Shape_K)
        self.mma_inst_shape_mn = (
            self.mma_tiler[0],
            self.mma_tiler[1],
        )
        # (CTA_Tile_Shape_M, Round_Up(MMA_Tile_Shape_N, 128), MMA_Inst_Shape_K)
        self.mma_inst_shape_mn_sfb = (
            self.mma_inst_shape_mn[0] // (2 if self.use_2cta_instrs else 1),
            cute.round_up(self.mma_inst_shape_mn[1], 128),
        )

        tiled_mma = sm100_utils.make_blockscaled_trivial_tiled_mma(
            self.a_dtype,
            self.a_major_mode,
            self.b_major_mode,
            self.sf_dtype,
            self.sf_vec_size,
            self.cta_group,
            self.mma_inst_shape_mn,
        )

        tiled_mma_sfb = sm100_utils.make_blockscaled_trivial_tiled_mma(
            self.a_dtype,
            self.a_major_mode,
            self.b_major_mode,
            self.sf_dtype,
            self.sf_vec_size,
            cute.nvgpu.tcgen05.CtaGroup.ONE,
            self.mma_inst_shape_mn_sfb,
        )

        # Compute mma/cluster/tile shapes
        mma_inst_shape_k = cute.size(tiled_mma.shape_mnk, mode=[2])
        mma_inst_tile_k = 4
        self.mma_tiler = (
            self.mma_inst_shape_mn[0],
            self.mma_inst_shape_mn[1],
            mma_inst_shape_k * mma_inst_tile_k,
        )
        self.mma_tiler_sfb = (
            self.mma_inst_shape_mn_sfb[0],
            self.mma_inst_shape_mn_sfb[1],
            mma_inst_shape_k * mma_inst_tile_k,
        )
        self.cta_tile_shape_mnk = (
            self.mma_tiler[0] // cute.size(tiled_mma.thr_id.shape),
            self.mma_tiler[1],
            self.mma_tiler[2],
        )

        # Compute cluster layout
        self.cluster_layout_vmnk = cute.tiled_divide(
            cute.make_layout((*self.cluster_shape_mn, 1)),
            (tiled_mma.thr_id.shape,),
        )
        self.cluster_layout_sfb_vmnk = cute.tiled_divide(
            cute.make_layout((*self.cluster_shape_mn, 1)),
            (tiled_mma_sfb.thr_id.shape,),
        )

        # Compute number of multicast CTAs for A/B
        self.num_mcast_ctas_a = cute.size(self.cluster_layout_vmnk.shape[2])
        self.num_mcast_ctas_b = cute.size(self.cluster_layout_vmnk.shape[1])
        self.num_mcast_ctas_sfb = cute.size(self.cluster_layout_sfb_vmnk.shape[1])
        self.is_a_mcast = self.num_mcast_ctas_a > 1
        self.is_b_mcast = self.num_mcast_ctas_b > 1
        self.is_sfb_mcast = self.num_mcast_ctas_sfb > 1

        # Compute epilogue subtile
        self.epi_tile = sm100_utils.compute_epilogue_tile_shape(
            self.cta_tile_shape_mnk,
            self.use_2cta_instrs,
            self.c_layout,
            self.c_dtype,
        )

        # Setup A/B/C stage count in shared memory and ACC stage count in tensor memory
        self.num_acc_stage, self.num_ab_stage, self.num_c_stage = self._compute_stages(
            tiled_mma,
            self.mma_tiler,
            self.a_dtype,
            self.b_dtype,
            self.epi_tile,
            self.c_dtype,
            self.c_layout,
            self.sf_dtype,
            self.sf_vec_size,
            self.smem_capacity,
            self.occupancy,
        )

        # Compute A/B/SFA/SFB/C shared memory layout
        self.a_smem_layout_staged = sm100_utils.make_smem_layout_a(
            tiled_mma,
            self.mma_tiler,
            self.a_dtype,
            self.num_ab_stage,
        )
        self.b_smem_layout_staged = sm100_utils.make_smem_layout_b(
            tiled_mma,
            self.mma_tiler,
            self.b_dtype,
            self.num_ab_stage,
        )
        self.sfa_smem_layout_staged = blockscaled_utils.make_smem_layout_sfa(
            tiled_mma,
            self.mma_tiler,
            self.sf_vec_size,
            self.num_ab_stage,
        )
        self.sfb_smem_layout_staged = blockscaled_utils.make_smem_layout_sfb(
            tiled_mma,
            self.mma_tiler,
            self.sf_vec_size,
            self.num_ab_stage,
        )
        self.c_smem_layout_staged = sm100_utils.make_smem_layout_epi(
            self.c_dtype,
            self.c_layout,
            self.epi_tile,
            self.num_c_stage,
        )

    @cute.jit
    def __call__(
        self,
        a_ptr: cute.Pointer,
        b_ptr: cute.Pointer,
        sfa_ptr: cute.Pointer,
        sfb_ptr: cute.Pointer,
        c_ptr: cute.Pointer,
        layouts: cutlass.Constexpr[
            Tuple[tcgen05.OperandMajorMode, tcgen05.OperandMajorMode, utils.LayoutEnum]
        ],
        problem_mnkl: Tuple[int, int, int, int],
        max_active_clusters: cutlass.Constexpr,
        epilogue_op: cutlass.Constexpr = lambda x: x,
    ):
        """Execute the GEMM operation in steps:
        - Setup static attributes before smem/grid/tma computation
        - Setup TMA load/store atoms and tensors
        - Compute grid size with regard to hardware constraints
        - Define shared storage for kernel
        - Launch the kernel synchronously

        :param a_tensor: Input tensor A
        :type a_tensor: cute.Tensor
        :param b_tensor: Input tensor B
        :type b_tensor: cute.Tensor
        :param sfa_tensor: Scale factor tensor A
        :type sfa_tensor: cute.Tensor
        :param sfb_tensor: Scale factor tensor B
        :type sfb_tensor: cute.Tensor
        :param c_tensor: Output tensor C
        :type c_tensor: cute.Tensor
        :param max_active_clusters: Maximum number of active clusters
        :type max_active_clusters: cutlass.Constexpr
        :param epilogue_op: Optional elementwise lambda function to apply to the output tensor
        :type epilogue_op: cutlass.Constexpr
        :raises TypeError: If input data types are incompatible with the MMA instruction.
        """
        # Setup static attributes before smem/grid/tma computation
        self.a_dtype: Type[cutlass.Numeric] = a_ptr.value_type
        self.b_dtype: Type[cutlass.Numeric] = b_ptr.value_type
        self.sf_dtype: Type[cutlass.Numeric] = sfa_ptr.value_type
        self.c_dtype: Type[cutlass.Numeric] = c_ptr.value_type

        m, n, k, l = problem_mnkl
        self.a_major_mode, self.b_major_mode, self.c_layout = layouts

        # Check if input data types are compatible with MMA instruction
        if cutlass.const_expr(self.a_dtype != self.b_dtype):
            raise TypeError(f"Type must match: {self.a_dtype} != {self.b_dtype}")

        # Setup attributes that dependent on gemm inputs
        self._setup_attributes()

        a_tensor = cute.make_tensor(
            a_ptr,
            cute.make_layout(
                (m, cute.assume(k, 32), l),
                stride=(cute.assume(k, 32), 1, cute.assume(m * k, 32)),
            ),
        )
        b_tensor = cute.make_tensor(
            b_ptr,
            cute.make_layout(
                (n, cute.assume(k, 32), l),
                stride=(cute.assume(k, 32), 1, cute.assume(n * k, 32)),
            ),
        )
        c_tensor = cute.make_tensor(
            c_ptr, cute.make_layout((cute.assume(m, 32), n, l), stride=(n, 1, m * n))
        )
        # Setup sfa/sfb tensor by filling A/B tensor to scale factor atom layout
        # ((Atom_M, Rest_M),(Atom_K, Rest_K),RestL)
        sfa_layout = blockscaled_utils.tile_atom_to_shape_SF(
            a_tensor.shape, self.sf_vec_size
        )
        sfa_tensor = cute.make_tensor(sfa_ptr, sfa_layout)

        # ((Atom_N, Rest_N),(Atom_K, Rest_K),RestL)
        sfb_layout = blockscaled_utils.tile_atom_to_shape_SF(
            b_tensor.shape, self.sf_vec_size
        )
        sfb_tensor = cute.make_tensor(sfb_ptr, sfb_layout)

        tiled_mma = sm100_utils.make_blockscaled_trivial_tiled_mma(
            self.a_dtype,
            self.a_major_mode,
            self.b_major_mode,
            self.sf_dtype,
            self.sf_vec_size,
            self.cta_group,
            self.mma_inst_shape_mn,
        )

        # For 2CTA blockscaled kernels, SFB needs to be replicated across peer CTAs. # {$nv-internal-release}
        tiled_mma_sfb = sm100_utils.make_blockscaled_trivial_tiled_mma(
            self.a_dtype,
            self.a_major_mode,
            self.b_major_mode,
            self.sf_dtype,
            self.sf_vec_size,
            cute.nvgpu.tcgen05.CtaGroup.ONE,
            self.mma_inst_shape_mn_sfb,
        )
        atom_thr_size = cute.size(tiled_mma.thr_id.shape)

        # Setup TMA load for A
        a_op = sm100_utils.cluster_shape_to_tma_atom_A(
            self.cluster_shape_mn, tiled_mma.thr_id
        )
        a_smem_layout = cute.slice_(self.a_smem_layout_staged, (None, None, None, 0))
        tma_atom_a, tma_tensor_a = cute.nvgpu.make_tiled_tma_atom_A(
            a_op,
            a_tensor,
            a_smem_layout,
            self.mma_tiler,
            tiled_mma,
            self.cluster_layout_vmnk.shape,
        )

        # Setup TMA load for B
        b_op = sm100_utils.cluster_shape_to_tma_atom_B(
            self.cluster_shape_mn, tiled_mma.thr_id
        )
        b_smem_layout = cute.slice_(self.b_smem_layout_staged, (None, None, None, 0))
        tma_atom_b, tma_tensor_b = cute.nvgpu.make_tiled_tma_atom_B(
            b_op,
            b_tensor,
            b_smem_layout,
            self.mma_tiler,
            tiled_mma,
            self.cluster_layout_vmnk.shape,
        )

        # Setup TMA load for SFA
        sfa_op = sm100_utils.cluster_shape_to_tma_atom_A(
            self.cluster_shape_mn, tiled_mma.thr_id
        )
        sfa_smem_layout = cute.slice_(
            self.sfa_smem_layout_staged, (None, None, None, 0)
        )
        tma_atom_sfa, tma_tensor_sfa = cute.nvgpu.make_tiled_tma_atom_A(
            sfa_op,
            sfa_tensor,
            sfa_smem_layout,
            self.mma_tiler,
            tiled_mma,
            self.cluster_layout_vmnk.shape,
            internal_type=cutlass.Int16,
        )

        # Setup TMA load for SFB
        sfb_op = sm100_utils.cluster_shape_to_tma_atom_SFB(
            self.cluster_shape_mn, tiled_mma.thr_id
        )
        sfb_smem_layout = cute.slice_(
            self.sfb_smem_layout_staged, (None, None, None, 0)
        )
        tma_atom_sfb, tma_tensor_sfb = cute.nvgpu.make_tiled_tma_atom_B(
            sfb_op,
            sfb_tensor,
            sfb_smem_layout,
            self.mma_tiler_sfb,
            tiled_mma_sfb,
            self.cluster_layout_sfb_vmnk.shape,
            internal_type=cutlass.Int16,
        )
        # This modifies the layout to handle overlapping 256x(# of scale factors for a single column of B (nNSF)) logical blocks for SFB when cta_tile_shape_n=192
        if cutlass.const_expr(self.cta_tile_shape_mnk[1] == 192):
            x = tma_tensor_sfb.stride[0][1]
            y = cute.ceil_div(tma_tensor_sfb.shape[0][1], 4)

            new_shape = (
                (tma_tensor_sfb.shape[0][0], ((2, 2), y)),
                tma_tensor_sfb.shape[1],
                tma_tensor_sfb.shape[2],
            )
            # Use right multiplication for ScaledBasis (3 * x instead of x * 3)
            x_times_3 = 3 * x
            new_stride = (
                (tma_tensor_sfb.stride[0][0], ((x, x), x_times_3)),
                tma_tensor_sfb.stride[1],
                tma_tensor_sfb.stride[2],
            )
            tma_tensor_sfb_new_layout = cute.make_layout(new_shape, stride=new_stride)
            tma_tensor_sfb = cute.make_tensor(
                tma_tensor_sfb.iterator, tma_tensor_sfb_new_layout
            )

        a_copy_size = cute.size_in_bytes(self.a_dtype, a_smem_layout)
        b_copy_size = cute.size_in_bytes(self.b_dtype, b_smem_layout)
        sfa_copy_size = cute.size_in_bytes(self.sf_dtype, sfa_smem_layout)
        sfb_copy_size = cute.size_in_bytes(self.sf_dtype, sfb_smem_layout)
        self.num_tma_load_bytes = (
            a_copy_size + b_copy_size + sfa_copy_size + sfb_copy_size
        ) * atom_thr_size

        # Setup TMA store for C
        epi_smem_layout = cute.slice_(self.c_smem_layout_staged, (None, None, 0))
        tma_atom_c, tma_tensor_c = cpasync.make_tiled_tma_atom(
            cpasync.CopyBulkTensorTileS2GOp(),
            c_tensor,
            epi_smem_layout,
            self.epi_tile,
        )

        # Compute grid size
        self.tile_sched_params, grid = self._compute_grid(
            c_tensor,
            self.cta_tile_shape_mnk,
            self.cluster_shape_mn,
            max_active_clusters,
        )

        self.buffer_align_bytes = 1024

        # Define shared storage for kernel
        @cute.struct
        class SharedStorage:
            ab_full_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage]
            ab_empty_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_ab_stage]
            acc_full_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage]
            acc_empty_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_acc_stage]
            tmem_dealloc_mbar_ptr: cutlass.Int64
            tmem_holding_buf: cutlass.Int32
            # (EPI_TILE_M, EPI_TILE_N, STAGE)
            sC: cute.struct.Align[
                cute.struct.MemRange[
                    self.c_dtype,
                    cute.cosize(self.c_smem_layout_staged.outer),
                ],
                self.buffer_align_bytes,
            ]
            # (MMA, MMA_M, MMA_K, STAGE)
            sA: cute.struct.Align[
                cute.struct.MemRange[
                    self.a_dtype, cute.cosize(self.a_smem_layout_staged.outer)
                ],
                self.buffer_align_bytes,
            ]
            # (MMA, MMA_N, MMA_K, STAGE)
            sB: cute.struct.Align[
                cute.struct.MemRange[
                    self.b_dtype, cute.cosize(self.b_smem_layout_staged.outer)
                ],
                self.buffer_align_bytes,
            ]
            # (MMA, MMA_M, MMA_K, STAGE)
            sSFA: cute.struct.Align[
                cute.struct.MemRange[
                    self.sf_dtype, cute.cosize(self.sfa_smem_layout_staged)
                ],
                self.buffer_align_bytes,
            ]
            # (MMA, MMA_N, MMA_K, STAGE)
            sSFB: cute.struct.Align[
                cute.struct.MemRange[
                    self.sf_dtype, cute.cosize(self.sfb_smem_layout_staged)
                ],
                self.buffer_align_bytes,
            ]

        self.shared_storage = SharedStorage

        # Launch the kernel synchronously
        self.kernel(
            tiled_mma,
            tiled_mma_sfb,
            tma_atom_a,
            tma_tensor_a,
            tma_atom_b,
            tma_tensor_b,
            tma_atom_sfa,
            tma_tensor_sfa,
            tma_atom_sfb,
            tma_tensor_sfb,
            tma_atom_c,
            tma_tensor_c,
            self.cluster_layout_vmnk,
            self.cluster_layout_sfb_vmnk,
            self.a_smem_layout_staged,
            self.b_smem_layout_staged,
            self.sfa_smem_layout_staged,
            self.sfb_smem_layout_staged,
            self.c_smem_layout_staged,
            self.epi_tile,
            self.tile_sched_params,
            epilogue_op,
        ).launch(
            grid=grid,
            block=[self.threads_per_cta, 1, 1],
            cluster=(*self.cluster_shape_mn, 1),
            min_blocks_per_mp=1,
        )
        return

    # GPU device kernel
    @cute.kernel
    def kernel(
        self,
        tiled_mma: cute.TiledMma,
        tiled_mma_sfb: cute.TiledMma,
        tma_atom_a: cute.CopyAtom,
        mA_mkl: cute.Tensor,
        tma_atom_b: cute.CopyAtom,
        mB_nkl: cute.Tensor,
        tma_atom_sfa: cute.CopyAtom,
        mSFA_mkl: cute.Tensor,
        tma_atom_sfb: cute.CopyAtom,
        mSFB_nkl: cute.Tensor,
        tma_atom_c: cute.CopyAtom,
        mC_mnl: cute.Tensor,
        cluster_layout_vmnk: cute.Layout,
        cluster_layout_sfb_vmnk: cute.Layout,
        a_smem_layout_staged: cute.ComposedLayout,
        b_smem_layout_staged: cute.ComposedLayout,
        sfa_smem_layout_staged: cute.Layout,
        sfb_smem_layout_staged: cute.Layout,
        c_smem_layout_staged: Union[cute.Layout, cute.ComposedLayout],
        epi_tile: cute.Tile,
        tile_sched_params: utils.PersistentTileSchedulerParams,
        epilogue_op: cutlass.Constexpr,
    ):
        """
        GPU device kernel performing the Persistent batched GEMM computation.
        """
        warp_idx = cute.arch.warp_idx()
        warp_idx = cute.arch.make_warp_uniform(warp_idx)

        #
        # Prefetch tma desc
        #
        if warp_idx == self.tma_warp_id:
            cpasync.prefetch_descriptor(tma_atom_a)
            cpasync.prefetch_descriptor(tma_atom_b)
            cpasync.prefetch_descriptor(tma_atom_sfa)
            cpasync.prefetch_descriptor(tma_atom_sfb)
            cpasync.prefetch_descriptor(tma_atom_c)

        use_2cta_instrs = cute.size(tiled_mma.thr_id.shape) == 2

        #
        # Setup cta/thread coordinates
        #
        # Coords inside cluster
        bidx, bidy, bidz = cute.arch.block_idx()
        mma_tile_coord_v = bidx % cute.size(tiled_mma.thr_id.shape)
        is_leader_cta = mma_tile_coord_v == 0
        cta_rank_in_cluster = cute.arch.make_warp_uniform(
            cute.arch.block_idx_in_cluster()
        )
        block_in_cluster_coord_vmnk = cluster_layout_vmnk.get_flat_coord(
            cta_rank_in_cluster
        )
        block_in_cluster_coord_sfb_vmnk = cluster_layout_sfb_vmnk.get_flat_coord(
            cta_rank_in_cluster
        )
        # Coord inside cta
        tidx, _, _ = cute.arch.thread_idx()

        #
        # Alloc and init: a+b full/empty, accumulator full/empty, tensor memory dealloc barrier
        #
        smem = utils.SmemAllocator()
        storage = smem.allocate(self.shared_storage)

        # Initialize mainloop ab_pipeline (barrier) and states
        ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
        num_tma_producer = self.num_mcast_ctas_a + self.num_mcast_ctas_b - 1
        ab_pipeline_consumer_group = pipeline.CooperativeGroup(
            pipeline.Agent.Thread, num_tma_producer
        )
        ab_pipeline = pipeline.PipelineTmaUmma.create(
            barrier_storage=storage.ab_full_mbar_ptr.data_ptr(),
            num_stages=self.num_ab_stage,
            producer_group=ab_pipeline_producer_group,
            consumer_group=ab_pipeline_consumer_group,
            tx_count=self.num_tma_load_bytes,
            cta_layout_vmnk=cluster_layout_vmnk,
        )

        # Initialize acc_pipeline (barrier) and states
        acc_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
        num_acc_consumer_threads = len(self.epilog_warp_id) * (
            2 if use_2cta_instrs else 1
        )
        acc_pipeline_consumer_group = pipeline.CooperativeGroup(
            pipeline.Agent.Thread, num_acc_consumer_threads
        )
        acc_pipeline = pipeline.PipelineUmmaAsync.create(
            barrier_storage=storage.acc_full_mbar_ptr.data_ptr(),
            num_stages=self.num_acc_stage,
            producer_group=acc_pipeline_producer_group,
            consumer_group=acc_pipeline_consumer_group,
            cta_layout_vmnk=cluster_layout_vmnk,
        )

        # Tensor memory dealloc barrier init
        tmem = utils.TmemAllocator(
            storage.tmem_holding_buf,
            barrier_for_retrieve=self.tmem_alloc_barrier,
            allocator_warp_id=self.epilog_warp_id[0],
            is_two_cta=use_2cta_instrs,
            two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar_ptr,
        )

        # Cluster arrive after barrier init
        if cute.size(self.cluster_shape_mn) > 1:
            cute.arch.cluster_arrive_relaxed()

        #
        # Setup smem tensor A/B/SFA/SFB/C
        #
        # (EPI_TILE_M, EPI_TILE_N, STAGE)
        sC = storage.sC.get_tensor(
            c_smem_layout_staged.outer, swizzle=c_smem_layout_staged.inner
        )
        # (MMA, MMA_M, MMA_K, STAGE)
        sA = storage.sA.get_tensor(
            a_smem_layout_staged.outer, swizzle=a_smem_layout_staged.inner
        )
        # (MMA, MMA_N, MMA_K, STAGE)
        sB = storage.sB.get_tensor(
            b_smem_layout_staged.outer, swizzle=b_smem_layout_staged.inner
        )
        # (MMA, MMA_M, MMA_K, STAGE)
        sSFA = storage.sSFA.get_tensor(sfa_smem_layout_staged)
        # (MMA, MMA_N, MMA_K, STAGE)
        sSFB = storage.sSFB.get_tensor(sfb_smem_layout_staged)

        #
        # Compute multicast mask for A/B/SFA/SFB buffer full
        #
        a_full_mcast_mask = None
        b_full_mcast_mask = None
        sfa_full_mcast_mask = None
        sfb_full_mcast_mask = None
        if cutlass.const_expr(self.is_a_mcast or self.is_b_mcast or use_2cta_instrs):
            a_full_mcast_mask = cpasync.create_tma_multicast_mask(
                cluster_layout_vmnk, block_in_cluster_coord_vmnk, mcast_mode=2
            )
            b_full_mcast_mask = cpasync.create_tma_multicast_mask(
                cluster_layout_vmnk, block_in_cluster_coord_vmnk, mcast_mode=1
            )
            sfa_full_mcast_mask = cpasync.create_tma_multicast_mask(
                cluster_layout_vmnk, block_in_cluster_coord_vmnk, mcast_mode=2
            )
            sfb_full_mcast_mask = cpasync.create_tma_multicast_mask(
                cluster_layout_sfb_vmnk, block_in_cluster_coord_sfb_vmnk, mcast_mode=1
            )

        #
        # Local_tile partition global tensors
        #
        # (bM, bK, RestM, RestK, RestL)
        gA_mkl = cute.local_tile(
            mA_mkl, cute.slice_(self.mma_tiler, (None, 0, None)), (None, None, None)
        )
        # (bN, bK, RestN, RestK, RestL)
        gB_nkl = cute.local_tile(
            mB_nkl, cute.slice_(self.mma_tiler, (0, None, None)), (None, None, None)
        )
        # (bM, bK, RestM, RestK, RestL)
        gSFA_mkl = cute.local_tile(
            mSFA_mkl, cute.slice_(self.mma_tiler, (None, 0, None)), (None, None, None)
        )
        # (bN, bK, RestN, RestK, RestL)
        gSFB_nkl = cute.local_tile(
            mSFB_nkl,
            cute.slice_(self.mma_tiler_sfb, (0, None, None)),
            (None, None, None),
        )
        # (bM, bN, RestM, RestN, RestL)
        gC_mnl = cute.local_tile(
            mC_mnl, cute.slice_(self.mma_tiler, (None, None, 0)), (None, None, None)
        )
        k_tile_cnt = cute.size(gA_mkl, mode=[3])

        #
        # Partition global tensor for TiledMMA_A/B/C
        #
        thr_mma = tiled_mma.get_slice(mma_tile_coord_v)
        thr_mma_sfb = tiled_mma_sfb.get_slice(mma_tile_coord_v)
        # (MMA, MMA_M, MMA_K, RestM, RestK, RestL)
        tCgA = thr_mma.partition_A(gA_mkl)
        # (MMA, MMA_N, MMA_K, RestN, RestK, RestL)
        tCgB = thr_mma.partition_B(gB_nkl)
        # (MMA, MMA_M, MMA_K, RestM, RestK, RestL)
        tCgSFA = thr_mma.partition_A(gSFA_mkl)
        # (MMA, MMA_N, MMA_K, RestN, RestK, RestL)
        tCgSFB = thr_mma_sfb.partition_B(gSFB_nkl)
        # (MMA, MMA_M, MMA_N, RestM, RestN, RestL)
        tCgC = thr_mma.partition_C(gC_mnl)

        #
        # Partition global/shared tensor for TMA load A/B
        #
        # TMA load A partition_S/D
        a_cta_layout = cute.make_layout(
            cute.slice_(cluster_layout_vmnk, (0, 0, None, 0)).shape
        )
        # ((atom_v, rest_v), STAGE)
        # ((atom_v, rest_v), RestM, RestK, RestL)
        tAsA, tAgA = cpasync.tma_partition(
            tma_atom_a,
            block_in_cluster_coord_vmnk[2],
            a_cta_layout,
            cute.group_modes(sA, 0, 3),
            cute.group_modes(tCgA, 0, 3),
        )
        # TMA load B partition_S/D
        b_cta_layout = cute.make_layout(
            cute.slice_(cluster_layout_vmnk, (0, None, 0, 0)).shape
        )
        # ((atom_v, rest_v), STAGE)
        # ((atom_v, rest_v), RestN, RestK, RestL)
        tBsB, tBgB = cpasync.tma_partition(
            tma_atom_b,
            block_in_cluster_coord_vmnk[1],
            b_cta_layout,
            cute.group_modes(sB, 0, 3),
            cute.group_modes(tCgB, 0, 3),
        )

        #  TMALDG_SFA partition_S/D
        sfa_cta_layout = a_cta_layout
        # ((atom_v, rest_v), STAGE)
        # ((atom_v, rest_v), RestM, RestK, RestL)
        tAsSFA, tAgSFA = cute.nvgpu.cpasync.tma_partition(
            tma_atom_sfa,
            block_in_cluster_coord_vmnk[2],
            sfa_cta_layout,
            cute.group_modes(sSFA, 0, 3),
            cute.group_modes(tCgSFA, 0, 3),
        )
        tAsSFA = cute.filter_zeros(tAsSFA)
        tAgSFA = cute.filter_zeros(tAgSFA)

        # TMALDG_SFB partition_S/D
        sfb_cta_layout = cute.make_layout(
            cute.slice_(cluster_layout_sfb_vmnk, (0, None, 0, 0)).shape
        )
        # ((atom_v, rest_v), STAGE)
        # ((atom_v, rest_v), RestN, RestK, RestL)
        tBsSFB, tBgSFB = cute.nvgpu.cpasync.tma_partition(
            tma_atom_sfb,
            block_in_cluster_coord_sfb_vmnk[1],
            sfb_cta_layout,
            cute.group_modes(sSFB, 0, 3),
            cute.group_modes(tCgSFB, 0, 3),
        )
        tBsSFB = cute.filter_zeros(tBsSFB)
        tBgSFB = cute.filter_zeros(tBgSFB)

        #
        # Partition shared/tensor memory tensor for TiledMMA_A/B/C
        #
        # (MMA, MMA_M, MMA_K, STAGE)
        tCrA = tiled_mma.make_fragment_A(sA)
        # (MMA, MMA_N, MMA_K, STAGE)
        tCrB = tiled_mma.make_fragment_B(sB)
        # (MMA, MMA_M, MMA_N)
        acc_shape = tiled_mma.partition_shape_C(self.mma_tiler[:2])
        # (MMA, MMA_M, MMA_N, STAGE)
        tCtAcc_fake = tiled_mma.make_fragment_C(
            cute.append(acc_shape, self.num_acc_stage)
        )

        #
        # Cluster wait before tensor memory alloc
        #
        if cute.size(self.cluster_shape_mn) > 1:
            cute.arch.cluster_wait()
        else:
            self.cta_sync_barrier.arrive_and_wait()

        #
        # Specialized TMA load warp
        #
        if warp_idx == self.tma_warp_id:
            #
            # Persistent tile scheduling loop
            #
            tile_sched = utils.StaticPersistentTileScheduler.create(
                tile_sched_params, cute.arch.block_idx(), cute.arch.grid_dim()
            )
            work_tile = tile_sched.initial_work_tile_info()

            ab_producer_state = pipeline.make_pipeline_state(
                pipeline.PipelineUserType.Producer, self.num_ab_stage
            )

            while work_tile.is_valid_tile:
                # Get tile coord from tile scheduler
                cur_tile_coord = work_tile.tile_idx
                mma_tile_coord_mnl = (
                    cur_tile_coord[0] // cute.size(tiled_mma.thr_id.shape),
                    cur_tile_coord[1],
                    cur_tile_coord[2],
                )

                #
                # Slice to per mma tile index
                #
                # ((atom_v, rest_v), RestK)
                tAgA_slice = tAgA[
                    (None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])
                ]
                # ((atom_v, rest_v), RestK)
                tBgB_slice = tBgB[
                    (None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])
                ]

                # ((atom_v, rest_v), RestK)
                tAgSFA_slice = tAgSFA[
                    (None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])
                ]

                # Apply SFB slicing hack when cta_tile_shape_n=64 # {$nv-internal-release}
                slice_n = mma_tile_coord_mnl[1]
                if cutlass.const_expr(self.cta_tile_shape_mnk[1] == 64):
                    slice_n = mma_tile_coord_mnl[1] // 2
                # ((atom_v, rest_v), RestK)
                tBgSFB_slice = tBgSFB[(None, slice_n, None, mma_tile_coord_mnl[2])]

                # Peek (try_wait) AB buffer empty for k_tile = prefetch_k_tile_cnt
                ab_producer_state.reset_count()
                peek_ab_empty_status = cutlass.Boolean(1)
                if ab_producer_state.count < k_tile_cnt:
                    peek_ab_empty_status = ab_pipeline.producer_try_acquire(
                        ab_producer_state
                    )
                #
                # Tma load loop
                #
                for k_tile in cutlass.range(0, k_tile_cnt, 1, unroll=1):
                    # Conditionally wait for AB buffer empty
                    ab_pipeline.producer_acquire(
                        ab_producer_state, peek_ab_empty_status
                    )

                    # TMA load A/B/SFA/SFB
                    cute.copy(
                        tma_atom_a,
                        tAgA_slice[(None, ab_producer_state.count)],
                        tAsA[(None, ab_producer_state.index)],
                        tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
                        mcast_mask=a_full_mcast_mask,
                    )
                    cute.copy(
                        tma_atom_b,
                        tBgB_slice[(None, ab_producer_state.count)],
                        tBsB[(None, ab_producer_state.index)],
                        tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
                        mcast_mask=b_full_mcast_mask,
                    )
                    cute.copy(
                        tma_atom_sfa,
                        tAgSFA_slice[(None, ab_producer_state.count)],
                        tAsSFA[(None, ab_producer_state.index)],
                        tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
                        mcast_mask=sfa_full_mcast_mask,
                    )
                    cute.copy(
                        tma_atom_sfb,
                        tBgSFB_slice[(None, ab_producer_state.count)],
                        tBsSFB[(None, ab_producer_state.index)],
                        tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
                        mcast_mask=sfb_full_mcast_mask,
                    )

                    # Peek (try_wait) AB buffer empty for k_tile = prefetch_k_tile_cnt + k_tile + 1
                    ab_producer_state.advance()
                    peek_ab_empty_status = cutlass.Boolean(1)
                    if ab_producer_state.count < k_tile_cnt:
                        peek_ab_empty_status = ab_pipeline.producer_try_acquire(
                            ab_producer_state
                        )

                #
                # Advance to next tile
                #
                tile_sched.advance_to_next_work()
                work_tile = tile_sched.get_current_work()

            #
            # Wait A/B buffer empty
            #
            ab_pipeline.producer_tail(ab_producer_state)

        #
        # Specialized MMA warp
        #
        if warp_idx == self.mma_warp_id:
            #
            # Bar sync for retrieve tensor memory ptr from shared mem
            #
            tmem.wait_for_alloc()

            #
            # Retrieving tensor memory ptr and make accumulator/SFA/SFB tensor
            #
            acc_tmem_ptr = tmem.retrieve_ptr(self.acc_dtype)
            # Make accumulator tmem tensor
            # (MMA, MMA_M, MMA_N, STAGE)
            tCtAcc_base = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)

            # Make SFA tmem tensor
            sfa_tmem_ptr = cute.recast_ptr(
                acc_tmem_ptr + tcgen05.find_tmem_tensor_col_offset(tCtAcc_base),
                dtype=self.sf_dtype,
            )
            # (MMA, MMA_M, MMA_K)
            tCtSFA_layout = blockscaled_utils.make_tmem_layout_sfa(
                tiled_mma,
                self.mma_tiler,
                self.sf_vec_size,
                cute.slice_(sfa_smem_layout_staged, (None, None, None, 0)),
            )
            tCtSFA = cute.make_tensor(sfa_tmem_ptr, tCtSFA_layout)

            # Make SFB tmem tensor
            sfb_tmem_ptr = cute.recast_ptr(
                acc_tmem_ptr
                + tcgen05.find_tmem_tensor_col_offset(tCtAcc_base)
                + tcgen05.find_tmem_tensor_col_offset(tCtSFA),
                dtype=self.sf_dtype,
            )
            # (MMA, MMA_N, MMA_K)
            tCtSFB_layout = blockscaled_utils.make_tmem_layout_sfb(
                tiled_mma,
                self.mma_tiler,
                self.sf_vec_size,
                cute.slice_(sfb_smem_layout_staged, (None, None, None, 0)),
            )
            tCtSFB = cute.make_tensor(sfb_tmem_ptr, tCtSFB_layout)
            #
            # Partition for S2T copy of SFA/SFB
            #
            (
                tiled_copy_s2t_sfa,
                tCsSFA_compact_s2t,
                tCtSFA_compact_s2t,
            ) = self.mainloop_s2t_copy_and_partition(sSFA, tCtSFA)
            (
                tiled_copy_s2t_sfb,
                tCsSFB_compact_s2t,
                tCtSFB_compact_s2t,
            ) = self.mainloop_s2t_copy_and_partition(sSFB, tCtSFB)

            #
            # Persistent tile scheduling loop
            #
            tile_sched = utils.StaticPersistentTileScheduler.create(
                tile_sched_params, cute.arch.block_idx(), cute.arch.grid_dim()
            )
            work_tile = tile_sched.initial_work_tile_info()

            ab_consumer_state = pipeline.make_pipeline_state(
                pipeline.PipelineUserType.Consumer, self.num_ab_stage
            )
            acc_producer_state = pipeline.make_pipeline_state(
                pipeline.PipelineUserType.Producer, self.num_acc_stage
            )

            while work_tile.is_valid_tile:
                # Get tile coord from tile scheduler
                cur_tile_coord = work_tile.tile_idx
                mma_tile_coord_mnl = (
                    cur_tile_coord[0] // cute.size(tiled_mma.thr_id.shape),
                    cur_tile_coord[1],
                    cur_tile_coord[2],
                )

                # Set tensor memory buffer for current tile
                # (MMA, MMA_M, MMA_N)
                tCtAcc = tCtAcc_base[(None, None, None, acc_producer_state.index)]

                # Peek (try_wait) AB buffer full for k_tile = 0
                ab_consumer_state.reset_count()
                peek_ab_full_status = cutlass.Boolean(1)
                if ab_consumer_state.count < k_tile_cnt and is_leader_cta:
                    peek_ab_full_status = ab_pipeline.consumer_try_wait(
                        ab_consumer_state
                    )

                #
                # Wait for accumulator buffer empty
                #
                if is_leader_cta:
                    acc_pipeline.producer_acquire(acc_producer_state)

                # Apply TMEM pointer offset hack when cta_tile_shape_n=192 or cta_tile_shape_n=64 # {$nv-internal-release}
                tCtSFB_mma = tCtSFB
                if cutlass.const_expr(self.cta_tile_shape_mnk[1] == 192):
                    # If this is an ODD tile, shift the TMEM start address for cta_tile_shape_n=192 case by two words (ignores first 64 columns of SFB)
                    offset = (
                        cutlass.Int32(2)
                        if mma_tile_coord_mnl[1] % 2 == 1
                        else cutlass.Int32(0)
                    )
                    shifted_ptr = cute.recast_ptr(
                        acc_tmem_ptr
                        + tcgen05.find_tmem_tensor_col_offset(tCtAcc_base)
                        + tcgen05.find_tmem_tensor_col_offset(tCtSFA)
                        + offset,
                        dtype=self.sf_dtype,
                    )
                    tCtSFB_mma = cute.make_tensor(shifted_ptr, tCtSFB_layout)
                elif cutlass.const_expr(self.cta_tile_shape_mnk[1] == 64):
                    # Move in increments of 64 columns of SFB
                    offset = cutlass.Int32((mma_tile_coord_mnl[1] % 2) * 2)
                    shifted_ptr = cute.recast_ptr(
                        acc_tmem_ptr
                        + tcgen05.find_tmem_tensor_col_offset(tCtAcc_base)
                        + tcgen05.find_tmem_tensor_col_offset(tCtSFA)
                        + offset,
                        dtype=self.sf_dtype,
                    )
                    tCtSFB_mma = cute.make_tensor(shifted_ptr, tCtSFB_layout)

                #
                # Reset the ACCUMULATE field for each tile
                #
                tiled_mma.set(tcgen05.Field.ACCUMULATE, False)

                #
                # Mma mainloop
                #
                for k_tile in range(k_tile_cnt):
                    if is_leader_cta:
                        # Conditionally wait for AB buffer full
                        ab_pipeline.consumer_wait(
                            ab_consumer_state, peek_ab_full_status
                        )

                        #  Copy SFA/SFB from smem to tmem
                        s2t_stage_coord = (
                            None,
                            None,
                            None,
                            None,
                            ab_consumer_state.index,
                        )
                        tCsSFA_compact_s2t_staged = tCsSFA_compact_s2t[s2t_stage_coord]
                        tCsSFB_compact_s2t_staged = tCsSFB_compact_s2t[s2t_stage_coord]
                        cute.copy(
                            tiled_copy_s2t_sfa,
                            tCsSFA_compact_s2t_staged,
                            tCtSFA_compact_s2t,
                        )
                        cute.copy(
                            tiled_copy_s2t_sfb,
                            tCsSFB_compact_s2t_staged,
                            tCtSFB_compact_s2t,
                        )

                        # tCtAcc += tCrA * tCrSFA * tCrB * tCrSFB
                        num_kblocks = cute.size(tCrA, mode=[2])
                        for kblock_idx in cutlass.range(num_kblocks, unroll_full=True):
                            kblock_coord = (
                                None,
                                None,
                                kblock_idx,
                                ab_consumer_state.index,
                            )

                            # Set SFA/SFB tensor to tiled_mma
                            sf_kblock_coord = (None, None, kblock_idx)
                            tiled_mma.set(
                                tcgen05.Field.SFA,
                                tCtSFA[sf_kblock_coord].iterator,
                            )
                            tiled_mma.set(
                                tcgen05.Field.SFB,
                                tCtSFB_mma[sf_kblock_coord].iterator,
                            )

                            cute.gemm(
                                tiled_mma,
                                tCtAcc,
                                tCrA[kblock_coord],
                                tCrB[kblock_coord],
                                tCtAcc,
                            )

                            # Enable accumulate on tCtAcc after first kblock
                            tiled_mma.set(tcgen05.Field.ACCUMULATE, True)

                        # Async arrive AB buffer empty
                        ab_pipeline.consumer_release(ab_consumer_state)

                    # Peek (try_wait) AB buffer full for k_tile = k_tile + 1
                    ab_consumer_state.advance()
                    peek_ab_full_status = cutlass.Boolean(1)
                    if ab_consumer_state.count < k_tile_cnt:
                        if is_leader_cta:
                            peek_ab_full_status = ab_pipeline.consumer_try_wait(
                                ab_consumer_state
                            )

                #
                # Async arrive accumulator buffer full
                #
                if is_leader_cta:
                    acc_pipeline.producer_commit(acc_producer_state)
                acc_producer_state.advance()

                #
                # Advance to next tile
                #
                tile_sched.advance_to_next_work()
                work_tile = tile_sched.get_current_work()

            #
            # Wait for accumulator buffer empty
            #
            acc_pipeline.producer_tail(acc_producer_state)
        #
        # Specialized epilogue warps
        #
        if warp_idx < self.mma_warp_id:
            #
            # Alloc tensor memory buffer
            #
            tmem.allocate(self.num_tmem_alloc_cols)

            #
            # Bar sync for retrieve tensor memory ptr from shared memory
            #
            tmem.wait_for_alloc()

            #
            # Retrieving tensor memory ptr and make accumulator tensor
            #
            acc_tmem_ptr = tmem.retrieve_ptr(self.acc_dtype)
            # (MMA, MMA_M, MMA_N, STAGE)
            tCtAcc_base = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)

            #
            # Partition for epilogue
            #
            epi_tidx = tidx
            (
                tiled_copy_t2r,
                tTR_tAcc_base,
                tTR_rAcc,
            ) = self.epilog_tmem_copy_and_partition(
                epi_tidx, tCtAcc_base, tCgC, epi_tile, use_2cta_instrs
            )

            tTR_rC = cute.make_rmem_tensor(tTR_rAcc.shape, self.c_dtype)
            tiled_copy_r2s, tRS_rC, tRS_sC = self.epilog_smem_copy_and_partition(
                tiled_copy_t2r, tTR_rC, epi_tidx, sC
            )
            (
                tma_atom_c,
                bSG_sC,
                bSG_gC_partitioned,
            ) = self.epilog_gmem_copy_and_partition(
                epi_tidx, tma_atom_c, tCgC, epi_tile, sC
            )

            #
            # Persistent tile scheduling loop
            #
            tile_sched = utils.StaticPersistentTileScheduler.create(
                tile_sched_params, cute.arch.block_idx(), cute.arch.grid_dim()
            )
            work_tile = tile_sched.initial_work_tile_info()

            acc_consumer_state = pipeline.make_pipeline_state(
                pipeline.PipelineUserType.Consumer, self.num_acc_stage
            )

            # Threads/warps participating in tma store pipeline
            c_producer_group = pipeline.CooperativeGroup(
                pipeline.Agent.Thread,
                32 * len(self.epilog_warp_id),
            )
            c_pipeline = pipeline.PipelineTmaStore.create(
                num_stages=self.num_c_stage,
                producer_group=c_producer_group,
            )

            while work_tile.is_valid_tile:
                # Get tile coord from tile scheduler
                cur_tile_coord = work_tile.tile_idx
                mma_tile_coord_mnl = (
                    cur_tile_coord[0] // cute.size(tiled_mma.thr_id.shape),
                    cur_tile_coord[1],
                    cur_tile_coord[2],
                )

                #
                # Slice to per mma tile index
                #
                # ((ATOM_V, REST_V), EPI_M, EPI_N)
                bSG_gC = bSG_gC_partitioned[
                    (
                        None,
                        None,
                        None,
                        *mma_tile_coord_mnl,
                    )
                ]

                # Set tensor memory buffer for current tile
                # (T2R, T2R_M, T2R_N, EPI_M, EPI_M)
                tTR_tAcc = tTR_tAcc_base[
                    (None, None, None, None, None, acc_consumer_state.index)
                ]

                #
                # Wait for accumulator buffer full
                #
                acc_pipeline.consumer_wait(acc_consumer_state)

                tTR_tAcc = cute.group_modes(tTR_tAcc, 3, cute.rank(tTR_tAcc))
                bSG_gC = cute.group_modes(bSG_gC, 1, cute.rank(bSG_gC))

                #
                # Store accumulator to global memory in subtiles
                #
                subtile_cnt = cute.size(tTR_tAcc.shape, mode=[3])
                num_prev_subtiles = tile_sched.num_tiles_executed * subtile_cnt
                for subtile_idx in cutlass.range(subtile_cnt):
                    #
                    # Load accumulator from tensor memory buffer to register
                    #
                    tTR_tAcc_mn = tTR_tAcc[(None, None, None, subtile_idx)]
                    cute.copy(tiled_copy_t2r, tTR_tAcc_mn, tTR_rAcc)

                    #
                    # Convert to C type
                    #
                    acc_vec = tiled_copy_r2s.retile(tTR_rAcc).load()
                    acc_vec = epilogue_op(acc_vec.to(self.c_dtype))
                    tRS_rC.store(acc_vec)

                    #
                    # Store C to shared memory
                    #
                    c_buffer = (num_prev_subtiles + subtile_idx) % self.num_c_stage
                    cute.copy(
                        tiled_copy_r2s,
                        tRS_rC,
                        tRS_sC[(None, None, None, c_buffer)],
                    )
                    # Fence and barrier to make sure shared memory store is visible to TMA store
                    cute.arch.fence_proxy(
                        cute.arch.ProxyKind.async_shared,
                        space=cute.arch.SharedSpace.shared_cta,
                    )
                    self.epilog_sync_barrier.arrive_and_wait()

                    #
                    # TMA store C to global memory
                    #
                    if warp_idx == self.epilog_warp_id[0]:
                        cute.copy(
                            tma_atom_c,
                            bSG_sC[(None, c_buffer)],
                            bSG_gC[(None, subtile_idx)],
                        )
                        # Fence and barrier to make sure shared memory store is visible to TMA store
                        c_pipeline.producer_commit()
                        c_pipeline.producer_acquire()
                    self.epilog_sync_barrier.arrive_and_wait()

                #
                # Async arrive accumulator buffer empty
                #
                with cute.arch.elect_one():
                    acc_pipeline.consumer_release(acc_consumer_state)
                acc_consumer_state.advance()

                #
                # Advance to next tile
                #
                tile_sched.advance_to_next_work()
                work_tile = tile_sched.get_current_work()

            #
            # Dealloc the tensor memory buffer
            #
            tmem.relinquish_alloc_permit()
            self.epilog_sync_barrier.arrive_and_wait()
            tmem.free(acc_tmem_ptr)
            #
            # Wait for C store complete
            #
            c_pipeline.producer_tail()

    def mainloop_s2t_copy_and_partition(
        self,
        sSF: cute.Tensor,
        tSF: cute.Tensor,
    ) -> Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]:
        """
        Make tiledCopy for smem to tmem load for scale factor tensor, then use it to partition smem memory (source) and tensor memory (destination).

        :param sSF: The scale factor tensor in smem
        :type sSF: cute.Tensor
        :param tSF: The scale factor tensor in tmem
        :type tSF: cute.Tensor

        :return: A tuple containing (tiled_copy_s2t, tCsSF_compact_s2t, tCtSF_compact_s2t) where:
            - tiled_copy_s2t: The tiled copy operation for smem to tmem load for scale factor tensor(s2t)
            - tCsSF_compact_s2t: The partitioned scale factor tensor in smem
            - tSF_compact_s2t: The partitioned scale factor tensor in tmem
        :rtype: Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]
        """
        # (MMA, MMA_MN, MMA_K, STAGE)
        tCsSF_compact = cute.filter_zeros(sSF)
        # (MMA, MMA_MN, MMA_K)
        tCtSF_compact = cute.filter_zeros(tSF)

        # Make S2T CopyAtom and tiledCopy
        copy_atom_s2t = cute.make_copy_atom(
            tcgen05.Cp4x32x128bOp(self.cta_group),
            self.sf_dtype,
        )
        tiled_copy_s2t = tcgen05.make_s2t_copy(copy_atom_s2t, tCtSF_compact)
        thr_copy_s2t = tiled_copy_s2t.get_slice(0)

        # ((ATOM_V, REST_V), Rest_Tiler, MMA_MN, MMA_K, STAGE)
        tCsSF_compact_s2t_ = thr_copy_s2t.partition_S(tCsSF_compact)
        # ((ATOM_V, REST_V), Rest_Tiler, MMA_MN, MMA_K, STAGE)
        tCsSF_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(
            tiled_copy_s2t, tCsSF_compact_s2t_
        )
        # ((ATOM_V, REST_V), Rest_Tiler, MMA_MN, MMA_K)
        tCtSF_compact_s2t = thr_copy_s2t.partition_D(tCtSF_compact)

        return tiled_copy_s2t, tCsSF_compact_s2t, tCtSF_compact_s2t

    def epilog_tmem_copy_and_partition(
        self,
        tidx: cutlass.Int32,
        tAcc: cute.Tensor,
        gC_mnl: cute.Tensor,
        epi_tile: cute.Tile,
        use_2cta_instrs: Union[cutlass.Boolean, bool],
    ) -> Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]:
        """
        Make tiledCopy for tensor memory load, then use it to partition tensor memory (source) and register array (destination).

        :param tidx: The thread index in epilogue warp groups
        :type tidx: cutlass.Int32
        :param tAcc: The accumulator tensor to be copied and partitioned
        :type tAcc: cute.Tensor
        :param gC_mnl: The global tensor C
        :type gC_mnl: cute.Tensor
        :param epi_tile: The epilogue tiler
        :type epi_tile: cute.Tile
        :param use_2cta_instrs: Whether use_2cta_instrs is enabled
        :type use_2cta_instrs: bool

        :return: A tuple containing (tiled_copy_t2r, tTR_tAcc, tTR_rAcc) where:
            - tiled_copy_t2r: The tiled copy operation for tmem to register copy(t2r)
            - tTR_tAcc: The partitioned accumulator tensor
            - tTR_rAcc: The accumulated tensor in register used to hold t2r results
        :rtype: Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]
        """
        # Make tiledCopy for tensor memory load
        copy_atom_t2r = sm100_utils.get_tmem_load_op(
            self.cta_tile_shape_mnk,
            self.c_layout,
            self.c_dtype,
            self.acc_dtype,
            epi_tile,
            use_2cta_instrs,
        )
        # (EPI_TILE_M, EPI_TILE_N, EPI_M, EPI_N, STAGE)
        tAcc_epi = cute.flat_divide(
            tAcc[((None, None), 0, 0, None)],
            epi_tile,
        )
        # (EPI_TILE_M, EPI_TILE_N)
        tiled_copy_t2r = tcgen05.make_tmem_copy(
            copy_atom_t2r, tAcc_epi[(None, None, 0, 0, 0)]
        )

        thr_copy_t2r = tiled_copy_t2r.get_slice(tidx)
        # (T2R, T2R_M, T2R_N, EPI_M, EPI_M, STAGE)
        tTR_tAcc = thr_copy_t2r.partition_S(tAcc_epi)

        # (EPI_TILE_M, EPI_TILE_N, EPI_M, EPI_N, RestM, RestN, RestL)
        gC_mnl_epi = cute.flat_divide(
            gC_mnl[((None, None), 0, 0, None, None, None)], epi_tile
        )
        # (T2R, T2R_M, T2R_N, EPI_M, EPI_N, RestM, RestN, RestL)
        tTR_gC = thr_copy_t2r.partition_D(gC_mnl_epi)
        # (T2R, T2R_M, T2R_N)
        tTR_rAcc = cute.make_rmem_tensor(
            tTR_gC[(None, None, None, 0, 0, 0, 0, 0)].shape, self.acc_dtype
        )
        return tiled_copy_t2r, tTR_tAcc, tTR_rAcc

    def epilog_smem_copy_and_partition(
        self,
        tiled_copy_t2r: cute.TiledCopy,
        tTR_rC: cute.Tensor,
        tidx: cutlass.Int32,
        sC: cute.Tensor,
    ) -> Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]:
        """
        Make tiledCopy for shared memory store, then use it to partition register array (source) and shared memory (destination).

        :param tiled_copy_t2r: The tiled copy operation for tmem to register copy(t2r)
        :type tiled_copy_t2r: cute.TiledCopy
        :param tTR_rC: The partitioned accumulator tensor
        :type tTR_rC: cute.Tensor
        :param tidx: The thread index in epilogue warp groups
        :type tidx: cutlass.Int32
        :param sC: The shared memory tensor to be copied and partitioned
        :type sC: cute.Tensor
        :type sepi: cute.Tensor

        :return: A tuple containing (tiled_copy_r2s, tRS_rC, tRS_sC) where:
            - tiled_copy_r2s: The tiled copy operation for register to smem copy(r2s)
            - tRS_rC: The partitioned tensor C (register source)
            - tRS_sC: The partitioned tensor C (smem destination)
        :rtype: Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]
        """
        copy_atom_r2s = sm100_utils.get_smem_store_op(
            self.c_layout, self.c_dtype, self.acc_dtype, tiled_copy_t2r
        )
        tiled_copy_r2s = cute.make_tiled_copy_D(copy_atom_r2s, tiled_copy_t2r)
        # (R2S, R2S_M, R2S_N, PIPE_D)
        thr_copy_r2s = tiled_copy_r2s.get_slice(tidx)
        tRS_sC = thr_copy_r2s.partition_D(sC)
        # (R2S, R2S_M, R2S_N)
        tRS_rC = tiled_copy_r2s.retile(tTR_rC)
        return tiled_copy_r2s, tRS_rC, tRS_sC

    def epilog_gmem_copy_and_partition(
        self,
        tidx: cutlass.Int32,
        atom: Union[cute.CopyAtom, cute.TiledCopy],
        gC_mnl: cute.Tensor,
        epi_tile: cute.Tile,
        sC: cute.Tensor,
    ) -> Tuple[cute.CopyAtom, cute.Tensor, cute.Tensor]:
        """Make tiledCopy for global memory store, then use it to:
        partition shared memory (source) and global memory (destination) for TMA store version.

        :param tidx: The thread index in epilogue warp groups
        :type tidx: cutlass.Int32
        :param atom: The copy_atom_c to be used for TMA store version, or tiled_copy_t2r for none TMA store version
        :type atom: cute.CopyAtom or cute.TiledCopy
        :param gC_mnl: The global tensor C
        :type gC_mnl: cute.Tensor
        :param epi_tile: The epilogue tiler
        :type epi_tile: cute.Tile
        :param sC: The shared memory tensor to be copied and partitioned
        :type sC: cute.Tensor

        :return: A tuple containing (tma_atom_c, bSG_sC, bSG_gC) where:
            - tma_atom_c: The TMA copy atom
            - bSG_sC: The partitioned shared memory tensor C
            - bSG_gC: The partitioned global tensor C
        :rtype: Tuple[cute.CopyAtom, cute.Tensor, cute.Tensor]
        """
        # (EPI_TILE_M, EPI_TILE_N, EPI_M, EPI_N, RestM, RestN, RestL)
        gC_epi = cute.flat_divide(
            gC_mnl[((None, None), 0, 0, None, None, None)], epi_tile
        )

        tma_atom_c = atom
        sC_for_tma_partition = cute.group_modes(sC, 0, 2)
        gC_for_tma_partition = cute.group_modes(gC_epi, 0, 2)
        # ((ATOM_V, REST_V), EPI_M, EPI_N)
        # ((ATOM_V, REST_V), EPI_M, EPI_N, RestM, RestN, RestL)
        bSG_sC, bSG_gC = cpasync.tma_partition(
            tma_atom_c,
            0,
            cute.make_layout(1),
            sC_for_tma_partition,
            gC_for_tma_partition,
        )
        return tma_atom_c, bSG_sC, bSG_gC

    @staticmethod
    def _compute_stages(
        tiled_mma: cute.TiledMma,
        mma_tiler_mnk: Tuple[int, int, int],
        a_dtype: Type[cutlass.Numeric],
        b_dtype: Type[cutlass.Numeric],
        epi_tile: cute.Tile,
        c_dtype: Type[cutlass.Numeric],
        c_layout: utils.LayoutEnum,
        sf_dtype: Type[cutlass.Numeric],
        sf_vec_size: int,
        smem_capacity: int,
        occupancy: int,
    ) -> Tuple[int, int, int]:
        """Computes the number of stages for A/B/C operands based on heuristics.

        :param tiled_mma: The tiled MMA object defining the core computation.
        :type tiled_mma: cute.TiledMma
        :param mma_tiler_mnk: The shape (M, N, K) of the MMA tiler.
        :type mma_tiler_mnk: tuple[int, int, int]
        :param a_dtype: Data type of operand A.
        :type a_dtype: type[cutlass.Numeric]
        :param b_dtype: Data type of operand B.
        :type b_dtype: type[cutlass.Numeric]
        :param epi_tile: The epilogue tile shape.
        :type epi_tile: cute.Tile
        :param c_dtype: Data type of operand C (output).
        :type c_dtype: type[cutlass.Numeric]
        :param c_layout: Layout enum of operand C.
        :type c_layout: utils.LayoutEnum
        :param sf_dtype: Data type of Scale factor.
        :type sf_dtype: type[cutlass.Numeric]
        :param sf_vec_size: Scale factor vector size.
        :type sf_vec_size: int
        :param smem_capacity: Total available shared memory capacity in bytes.
        :type smem_capacity: int
        :param occupancy: Target number of CTAs per SM (occupancy).
        :type occupancy: int

        :return: A tuple containing the computed number of stages for:
                 (ACC stages, A/B operand stages, C stages)
        :rtype: tuple[int, int, int]
        """
        # ACC stages
        num_acc_stage = 1 if mma_tiler_mnk[1] == 256 else 2

        # Default C stages
        num_c_stage = 2

        # Calculate smem layout and size for one stage of A, B, SFA, SFB and C
        a_smem_layout_stage_one = sm100_utils.make_smem_layout_a(
            tiled_mma,
            mma_tiler_mnk,
            a_dtype,
            1,  # a tmp 1 stage is provided
        )
        b_smem_layout_staged_one = sm100_utils.make_smem_layout_b(
            tiled_mma,
            mma_tiler_mnk,
            b_dtype,
            1,  # a tmp 1 stage is provided
        )
        sfa_smem_layout_staged_one = blockscaled_utils.make_smem_layout_sfa(
            tiled_mma,
            mma_tiler_mnk,
            sf_vec_size,
            1,  # a tmp 1 stage is provided
        )
        sfb_smem_layout_staged_one = blockscaled_utils.make_smem_layout_sfb(
            tiled_mma,
            mma_tiler_mnk,
            sf_vec_size,
            1,  # a tmp 1 stage is provided
        )

        c_smem_layout_staged_one = sm100_utils.make_smem_layout_epi(
            c_dtype,
            c_layout,
            epi_tile,
            1,
        )

        ab_bytes_per_stage = (
            cute.size_in_bytes(a_dtype, a_smem_layout_stage_one)
            + cute.size_in_bytes(b_dtype, b_smem_layout_staged_one)
            + cute.size_in_bytes(sf_dtype, sfa_smem_layout_staged_one)
            + cute.size_in_bytes(sf_dtype, sfb_smem_layout_staged_one)
        )
        mbar_helpers_bytes = 1024
        c_bytes_per_stage = cute.size_in_bytes(c_dtype, c_smem_layout_staged_one)
        c_bytes = c_bytes_per_stage * num_c_stage

        # Calculate A/B/SFA/SFB stages:
        # Start with total smem per CTA (capacity / occupancy)
        # Subtract reserved bytes and initial C stages bytes
        # Divide remaining by bytes needed per A/B/SFA/SFB stage
        num_ab_stage = (
            smem_capacity // occupancy - (mbar_helpers_bytes + c_bytes)
        ) // ab_bytes_per_stage

        # Refine epilogue stages:
        # Calculate remaining smem after allocating for A/B/SFA/SFB stages and reserved bytes
        # Add remaining unused smem to epilogue
        num_c_stage += (
            smem_capacity
            - occupancy * ab_bytes_per_stage * num_ab_stage
            - occupancy * (mbar_helpers_bytes + c_bytes)
        ) // (occupancy * c_bytes_per_stage)

        return num_acc_stage, num_ab_stage, num_c_stage

    @staticmethod
    def _compute_grid(
        c: cute.Tensor,
        cta_tile_shape_mnk: Tuple[int, int, int],
        cluster_shape_mn: Tuple[int, int],
        max_active_clusters: cutlass.Constexpr,
    ) -> Tuple[utils.PersistentTileSchedulerParams, Tuple[int, int, int]]:
        """Use persistent tile scheduler to compute the grid size for the output tensor C.

        :param c: The output tensor C
        :type c: cute.Tensor
        :param cta_tile_shape_mnk: The shape (M, N, K) of the CTA tile.
        :type cta_tile_shape_mnk: tuple[int, int, int]
        :param cluster_shape_mn: Shape of each cluster in M, N dimensions.
        :type cluster_shape_mn: tuple[int, int]
        :param max_active_clusters: Maximum number of active clusters.
        :type max_active_clusters: cutlass.Constexpr

        :return: A tuple containing:
            - tile_sched_params: Parameters for the persistent tile scheduler.
            - grid: Grid shape for kernel launch.
        :rtype: Tuple[utils.PersistentTileSchedulerParams, tuple[int, int, int]]
        """
        c_shape = cute.slice_(cta_tile_shape_mnk, (None, None, 0))
        gc = cute.zipped_divide(c, tiler=c_shape)
        num_ctas_mnl = gc[(0, (None, None, None))].shape
        cluster_shape_mnl = (*cluster_shape_mn, 1)

        tile_sched_params = utils.PersistentTileSchedulerParams(
            num_ctas_mnl, cluster_shape_mnl
        )
        grid = utils.StaticPersistentTileScheduler.get_grid_shape(
            tile_sched_params, max_active_clusters
        )

        return tile_sched_params, grid

    @staticmethod
    def is_valid_dtypes_and_scale_factor_vec_size(
        ab_dtype: Type[cutlass.Numeric],
        sf_dtype: Type[cutlass.Numeric],
        sf_vec_size: int,
        c_dtype: Type[cutlass.Numeric],
    ) -> bool:
        """
        Check if the dtypes and sf_vec_size are valid combinations

        :param ab_dtype: The data type of the A and B operands
        :type ab_dtype: Type[cutlass.Numeric]
        :param sf_dtype: The data type of the scale factor
        :type sf_dtype: Type[cutlass.Numeric]
        :param sf_vec_size: The vector size of the scale factor
        :type sf_vec_size: int
        :param c_dtype: The data type of the output tensor
        :type c_dtype: Type[cutlass.Numeric]

        :return: True if the dtypes and sf_vec_size are valid, False otherwise
        :rtype: bool
        """
        is_valid = True

        # Check valid ab_dtype
        if ab_dtype not in {
            cutlass.Float4E2M1FN,
        }:
            is_valid = False

        # Check valid sf_vec_size
        if sf_vec_size not in {
            16,
        }:
            is_valid = False

        # Check valid sf_dtype
        if sf_dtype not in {cutlass.Float8E4M3FN}:
            is_valid = False

        # Check valid c_dtype
        if c_dtype not in {
            cutlass.Float32,
            cutlass.Float16,
            cutlass.BFloat16,
        }:
            is_valid = False

        return is_valid

    @staticmethod
    def is_valid_layouts(
        ab_dtype: Type[cutlass.Numeric],
        c_dtype: Type[cutlass.Numeric],
        a_major: str,
        b_major: str,
        c_major: str,
    ) -> bool:
        """
        Check if layouts and dtypes are valid combinations

        :param ab_dtype: The data type of the A and B operands
        :type ab_dtype: Type[cutlass.Numeric]
        :param c_dtype: The data type of the output tensor
        :type c_dtype: Type[cutlass.Numeric]
        :param a_major: The major dimension of the A tensor
        :type a_major: str
        :param b_major: The major dimension of the B tensor
        :type b_major: str
        :param c_major: The major dimension of the C tensor
        :type c_major: str

        :return: True if the layouts are valid, False otherwise
        :rtype: bool
        """
        is_valid = True

        # For NVF4 block-scaled GEMM we only support K-major A/B.
        if ab_dtype is cutlass.Float4E2M1FN and not (a_major == "k" and b_major == "k"):
            is_valid = False
        # For the current NVF4 path, only N-major (row-major) output is supported.
        if ab_dtype is cutlass.Float4E2M1FN and c_major == "m":
            is_valid = False

        return is_valid

    @staticmethod
    def is_valid_mma_tiler_and_cluster_shape(
        mma_tiler_mn: Tuple[int, int],
        cluster_shape_mn: Tuple[int, int],
    ) -> bool:
        """
        Check if the mma tiler and cluster shape are valid

        :param mma_tiler_mn: The (M, N) shape of the MMA instruction tiler
        :type mma_tiler_mn: Tuple[int, int]
        :param cluster_shape_mn: The (ClusterM, ClusterN) shape of the CTA cluster
        :type cluster_shape_mn: Tuple[int, int]

        :return: True if the mma tiler and cluster shape are valid, False otherwise
        :rtype: bool
        """
        is_valid = True
        # Skip invalid mma tile shape
        if mma_tiler_mn[0] not in [128, 256]:
            is_valid = False
        if mma_tiler_mn[1] not in [64, 128, 192, 256]:
            is_valid = False
        # Skip illegal cluster shape
        if cluster_shape_mn[0] % (2 if mma_tiler_mn[0] == 256 else 1) != 0:
            is_valid = False
        # Skip invalid cluster shape
        is_power_of_2 = lambda x: x > 0 and (x & (x - 1)) == 0
        if (
            cluster_shape_mn[0] * cluster_shape_mn[1] > 16
            or cluster_shape_mn[0] <= 0
            or cluster_shape_mn[1] <= 0
            # Special cluster shape check for scale factor multicasts.
            # Due to limited size of scale factors, we can't multicast among more than 4 CTAs.
            or cluster_shape_mn[0] > 4
            or cluster_shape_mn[1] > 4
            or not is_power_of_2(cluster_shape_mn[0])
            or not is_power_of_2(cluster_shape_mn[1])
        ):
            is_valid = False
        return is_valid

    @staticmethod
    def is_valid_tensor_alignment(
        m: int,
        n: int,
        k: int,
        l: int,
        ab_dtype: Type[cutlass.Numeric],
        c_dtype: Type[cutlass.Numeric],
        a_major: str,
        b_major: str,
        c_major: str,
    ) -> bool:
        """
        Check if the tensor alignment is valid

        :param m: The number of rows in the A tensor
        :type m: int
        :param n: The number of columns in the B tensor
        :type n: int
        :param k: The number of columns in the A tensor
        :type k: int
        :param l: The number of columns in the C tensor
        :type l: int
        :param ab_dtype: The data type of the A and B operands
        :type ab_dtype: Type[cutlass.Numeric]
        :param c_dtype: The data type of the output tensor
        :type c_dtype: Type[cutlass.Numeric]
        :param a_major: The major axis of the A tensor
        :type a_major: str
        :param b_major: The major axis of the B tensor
        :type b_major: str
        :param c_major: The major axis of the C tensor
        :type c_major: str

        :return: True if the problem shape is valid, False otherwise
        :rtype: bool
        """
        is_valid = True

        def check_contigous_16B_alignment(dtype, is_mode0_major, tensor_shape):
            major_mode_idx = 0 if is_mode0_major else 1
            num_major_elements = tensor_shape[major_mode_idx]
            num_contiguous_elements = 16 * 8 // dtype.width
            return num_major_elements % num_contiguous_elements == 0

        if (
            not check_contigous_16B_alignment(ab_dtype, a_major == "m", (m, k, l))
            or not check_contigous_16B_alignment(ab_dtype, b_major == "n", (n, k, l))
            or not check_contigous_16B_alignment(c_dtype, c_major == "m", (m, n, l))
        ):
            is_valid = False
        return is_valid

    @staticmethod
    def can_implement(
        ab_dtype: Type[cutlass.Numeric],
        sf_dtype: Type[cutlass.Numeric],
        sf_vec_size: int,
        c_dtype: Type[cutlass.Numeric],
        mma_tiler_mn: Tuple[int, int],
        cluster_shape_mn: Tuple[int, int],
        m: int,
        n: int,
        k: int,
        l: int,
        a_major: str,
        b_major: str,
        c_major: str,
    ) -> bool:
        """
        Check if the gemm can be implemented

        :param ab_dtype: The data type of the A and B operands
        :type ab_dtype: Type[cutlass.Numeric]
        :param sf_dtype: The data type of the scale factor tensor
        :type sf_dtype: Type[cutlass.Numeric]
        :param sf_vec_size: The vector size
        :type sf_vec_size: int
        :param c_dtype: The data type of the output tensor
        :type c_dtype: Type[cutlass.Numeric]
        :param mma_tiler_mn: The (M, N) shape of the MMA instruction tiler
        :type mma_tiler_mn: Tuple[int, int]
        :param cluster_shape_mn: The (ClusterM, ClusterN) shape of the CTA cluster
        :type cluster_shape_mn: Tuple[int, int]
        :param m: The number of rows in the A tensor
        :type m: int
        :param n: The number of columns in the B tensor
        :type n: int
        :param k: The number of columns in the A tensor
        :type k: int
        :param l: The number of columns in the C tensor
        :type l: int
        :param a_major: The major axis of the A tensor
        :type a_major: str
        :param b_major: The major axis of the B tensor
        :type b_major: str
        :param c_major: The major axis of the C tensor
        :type c_major: str

        :return: True if the gemm can be implemented, False otherwise
        :rtype: bool
        """
        can_implement = True
        # Skip unsupported types
        if not Sm100BlockScaledPersistentDenseGemmKernel.is_valid_dtypes_and_scale_factor_vec_size(
            ab_dtype, sf_dtype, sf_vec_size, c_dtype
        ):
            can_implement = False
        # Skip unsupported layouts
        if not Sm100BlockScaledPersistentDenseGemmKernel.is_valid_layouts(
            ab_dtype, c_dtype, a_major, b_major, c_major
        ):
            can_implement = False
        # Skip invalid mma tile shape and cluster shape
        if not Sm100BlockScaledPersistentDenseGemmKernel.is_valid_mma_tiler_and_cluster_shape(
            mma_tiler_mn, cluster_shape_mn
        ):
            can_implement = False
        # Skip illegal problem shape for load/store alignment
        if not Sm100BlockScaledPersistentDenseGemmKernel.is_valid_tensor_alignment(
            m, n, k, l, ab_dtype, c_dtype, a_major, b_major, c_major
        ):
            can_implement = False
        return can_implement


# Copyright (c) 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.


# ----------------- Hackathon wrapper -----------------

# Data types
ab_dtype = cutlass.Float4E2M1FN
sf_dtype = cutlass.Float8E4M3FN
c_dtype  = cutlass.Float16

# Default configuration (good baseline for M=128, L=1, N in {4096,7168})
_SF_VEC_SIZE = 16
_MMA_TILER_MN = (128, 64)
_CLUSTER_SHAPE_MN = (1, 2)

# Compile options
_COMPILE_OPT_LEVEL = 2

# Kernel object
_gemm = Sm100BlockScaledPersistentDenseGemmKernel(
    sf_vec_size=_SF_VEC_SIZE,
    mma_tiler_mn=_MMA_TILER_MN,
    cluster_shape_mn=_CLUSTER_SHAPE_MN,
)

# Compile cache keyed by (config, m,n,k,l)
_compiled = {}
_max_active_clusters = {}


def _get_max_active_clusters(cluster_shape_mn):
    cluster_size = cluster_shape_mn[0] * cluster_shape_mn[1]
    if cluster_size in _max_active_clusters:
        return _max_active_clusters[cluster_size]
    hw = utils.HardwareInfo()
    val = hw.get_max_active_clusters(cluster_size)
    _max_active_clusters[cluster_size] = val
    return val



def compile_kernel(problem_size):
    """Compile and cache a specialized kernel for (m,n,k,l) under the active config."""
    ps = tuple(problem_size)
    cfg_key = (_MMA_TILER_MN, _CLUSTER_SHAPE_MN, _SF_VEC_SIZE, _COMPILE_OPT_LEVEL)
    key = (cfg_key, ps)
    if key in _compiled:
        return _compiled[key]

    # Dummy pointers for compilation
    a_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    b_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    sfa_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    sfb_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    c_ptr = make_ptr(c_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)

    layouts = (
        tcgen05.OperandMajorMode.K,
        tcgen05.OperandMajorMode.K,
        utils.LayoutEnum.ROW_MAJOR,
    )

    max_active_clusters = _get_max_active_clusters(_CLUSTER_SHAPE_MN)

    compiled = cute.compile(
        _gemm,
        a_ptr, b_ptr, sfa_ptr, sfb_ptr, c_ptr,
        layouts,
        ps,
        max_active_clusters,
        options=f"--opt-level {_COMPILE_OPT_LEVEL}",
    )

    _compiled[key] = compiled
    return compiled


def custom_kernel(data: input_t) -> output_t:
    a, b, _, _, sfa_permuted, sfb_permuted, c = data

    m, k_packed, l = a.shape
    n, _, _ = b.shape
    # Torch packs 2 FP4 values into one byte (e2m1_x2), so logical K doubles
    k = k_packed * 2

    problem_size = (m, n, k, l)
    compiled = compile_kernel(problem_size)

    a_ptr = make_ptr(ab_dtype, a.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    b_ptr = make_ptr(ab_dtype, b.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    sfa_ptr = make_ptr(sf_dtype, sfa_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    sfb_ptr = make_ptr(sf_dtype, sfb_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    c_ptr = make_ptr(c_dtype, c.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)

    # The compiled callable uses the default CUDA execution queue.
    compiled(a_ptr, b_ptr, sfa_ptr, sfb_ptr, c_ptr, problem_size)
    return c