submission 375647

from typing import Tuple, Type, Union

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

#  -- FFI --

import subprocess
import sys


def install_package():
    """Install apache-tvm-ffi using pip"""
    try:
        print("Installing apache-tvm-ffi...")
        subprocess.check_call([sys.executable, "-m", "pip", "install", "apache-tvm-ffi"])
        print("Successfully installed apache-tvm-ffi!")
        return True
    except subprocess.CalledProcessError as e:
        print(f"Error installing package: {e}")
        return False


success = install_package()


# -- Silu Approxs --


@cute.jit
def silu(x):
    a = x * 0.5
    b = 1.0 + cute.math.tanh(a, fastmath=True)  #  ~= 2 * sigmoid(x)
    return a * b  # = x * sigmoid(x)


# --


class Sm100BlockScaledPersistentDenseGemmKernel:
    def __init__(
        self,
        mma_tiler_mn: Tuple[int, int],
        cluster_shape_mn: Tuple[int, int],
    ):
        self.acc_dtype = cutlass.Float32
        self.use_2cta_instrs = mma_tiler_mn[0] == 256
        self.cluster_shape_mn = cluster_shape_mn
        # K dimension is deferred until _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
        # Warp specialization
        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))

        # ---- PATCH: use same barrier scheme as dense.py
        self.epilog_sync_barrier = pipeline.NamedBarrier(
            barrier_id=1,
            num_threads=32 * len(self.epilog_warp_id),
        )
        self.tmem_alloc_barrier = pipeline.NamedBarrier(
            barrier_id=2,
            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):
        # Compute mma instruction shapes
        self.mma_inst_shape_mn = (
            self.mma_tiler[0],
            self.mma_tiler[1],
        )
        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,
            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,
            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],
        )
        self.cta_tile_shape_mnk_sfb = (
            self.mma_tiler_sfb[0] // cute.size(tiled_mma.thr_id.shape),
            self.mma_tiler_sfb[1],
            self.mma_tiler_sfb[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/SFB
        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,
        )
        self.epi_tile_n = cute.size(self.epi_tile[1])

        # Setup stage counts
        (
            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,
            sf_vec_size,
            self.smem_capacity,
            self.occupancy,
        )

        # Shared memory layouts
        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,
            sf_vec_size,
            self.num_ab_stage,
        )
        self.sfb_smem_layout_staged = blockscaled_utils.make_smem_layout_sfb(
            tiled_mma,
            self.mma_tiler,
            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,
        )

        self.overlapping_accum = False

        sf_atom_mn = 32
        self.num_sfa_tmem_cols = (self.cta_tile_shape_mnk[0] // sf_atom_mn) * mma_inst_tile_k
        self.num_sfb_tmem_cols = (self.cta_tile_shape_mnk_sfb[1] // sf_atom_mn) * mma_inst_tile_k
        # Dual GEMM: two SFB blocks and two accumulator blocks.
        self.num_sf_tmem_cols = self.num_sfa_tmem_cols + 2 * self.num_sfb_tmem_cols
        self.num_accumulator_tmem_cols = self.cta_tile_shape_mnk[1] * self.num_acc_stage
        self.num_accumulator2_tmem_cols = self.cta_tile_shape_mnk[1] * self.num_acc_stage

        self.iter_acc_early_release_in_epilogue = 0

    @cute.jit
    def __call__(
        self,
        a_ptr: cute.Pointer,
        b1_ptr: cute.Pointer,
        b2_ptr: cute.Pointer,
        sfa_ptr: cute.Pointer,
        sfb1_ptr: cute.Pointer,
        sfb2_ptr: cute.Pointer,
        c_ptr: cute.Pointer,
        layouts: cutlass.Constexpr[
            Tuple[tcgen05.OperandMajorMode, tcgen05.OperandMajorMode, utils.LayoutEnum]
        ],
        problem_mnkl: cutlass.Constexpr[Tuple[int, int, int, int]],
        max_active_clusters: cutlass.Constexpr,
        epilogue_op: cutlass.Constexpr = lambda x: x,
    ):
        self.a_dtype = a_ptr.value_type
        self.b_dtype = b1_ptr.value_type
        self.sf_dtype = sfa_ptr.value_type
        self.c_dtype = c_ptr.value_type

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

        self._setup_attributes()

        a_tensor = cute.make_tensor(
            a_ptr,
            cute.make_layout(
                (m, k, l),
                stride=(k, 1, m * k),
            ),
        )

        b1_tensor = cute.make_tensor(
            b1_ptr,
            cute.make_layout(
                (n, k, l),
                stride=(k, 1, n * k),
            ),
        )

        b2_tensor = cute.make_tensor(
            b2_ptr,
            cute.make_layout(
                (n, k, l),
                stride=(k, 1, n * k),
            ),
        )

        c_tensor = cute.make_tensor(
            c_ptr,
            cute.make_layout(
                (m, n, l),
                stride=(n, 1, m * n),
            ),
        )

        # SF layouts identical to dense.py
        sfa_layout = blockscaled_utils.tile_atom_to_shape_SF(a_tensor.shape, sf_vec_size)
        sfa_tensor = cute.make_tensor(sfa_ptr, sfa_layout)

        sfb1_layout = blockscaled_utils.tile_atom_to_shape_SF(b1_tensor.shape, sf_vec_size)
        sfb1_tensor = cute.make_tensor(sfb1_ptr, sfb1_layout)
        sfb2_layout = blockscaled_utils.tile_atom_to_shape_SF(b2_tensor.shape, sf_vec_size)
        sfb2_tensor = cute.make_tensor(sfb2_ptr, sfb2_layout)

        tiled_mma = sm100_utils.make_blockscaled_trivial_tiled_mma(
            self.a_dtype,
            self.a_major_mode,
            self.b_major_mode,
            self.sf_dtype,
            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,
            sf_vec_size,
            cute.nvgpu.tcgen05.CtaGroup.ONE,
            self.mma_inst_shape_mn_sfb,
        )
        atom_thr_size = cute.size(tiled_mma.thr_id.shape)

        # TMA load ops / atoms same as dense.py
        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,
        )

        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_b1, tma_tensor_b1 = cute.nvgpu.make_tiled_tma_atom_B(
            b_op,
            b1_tensor,
            b_smem_layout,
            self.mma_tiler,
            tiled_mma,
            self.cluster_layout_vmnk.shape,
        )
        tma_atom_b2, tma_tensor_b2 = cute.nvgpu.make_tiled_tma_atom_B(
            b_op,
            b2_tensor,
            b_smem_layout,
            self.mma_tiler,
            tiled_mma,
            self.cluster_layout_vmnk.shape,
        )

        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,
        )

        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_sfb1, tma_tensor_sfb1 = cute.nvgpu.make_tiled_tma_atom_B(
            sfb_op,
            sfb1_tensor,
            sfb_smem_layout,
            self.mma_tiler_sfb,
            tiled_mma_sfb,
            self.cluster_layout_sfb_vmnk.shape,
            internal_type=cutlass.Int16,
        )
        tma_atom_sfb2, tma_tensor_sfb2 = cute.nvgpu.make_tiled_tma_atom_B(
            sfb_op,
            sfb2_tensor,
            sfb_smem_layout,
            self.mma_tiler_sfb,
            tiled_mma_sfb,
            self.cluster_layout_sfb_vmnk.shape,
            internal_type=cutlass.Int16,
        )

        # 192‑column special case (copied)
        if cutlass.const_expr(self.cta_tile_shape_mnk[1] == 192):
            x = tma_tensor_sfb1.stride[0][1]
            y = cute.ceil_div(tma_tensor_sfb1.shape[0][1], 4)

            new_shape = (
                (tma_tensor_sfb1.shape[0][0], ((2, 2), y)),
                tma_tensor_sfb1.shape[1],
                tma_tensor_sfb1.shape[2],
            )
            x_times_3 = 3 * x
            new_stride = (
                (tma_tensor_sfb1.stride[0][0], ((x, x), x_times_3)),
                tma_tensor_sfb1.stride[1],
                tma_tensor_sfb1.stride[2],
            )
            tma_tensor_sfb_new_layout = cute.make_layout(new_shape, stride=new_stride)
            tma_tensor_sfb1 = cute.make_tensor(tma_tensor_sfb1.iterator, tma_tensor_sfb_new_layout)
            tma_tensor_sfb2 = cute.make_tensor(tma_tensor_sfb2.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 + 2 * b_copy_size + sfa_copy_size + 2 * sfb_copy_size
        ) * atom_thr_size

        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,
        )

        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

        @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,
            ]
            sB2: 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,
            ]
            sSFB2: cute.struct.Align[
                cute.struct.MemRange[self.sf_dtype, cute.cosize(self.sfb_smem_layout_staged)],
                self.buffer_align_bytes,
            ]

        self.shared_storage = SharedStorage

        self.kernel(
            tiled_mma,
            tiled_mma_sfb,
            tma_atom_a,
            tma_tensor_a,
            tma_atom_b1,
            tma_tensor_b1,
            tma_atom_b2,
            tma_tensor_b2,
            tma_atom_sfa,
            tma_tensor_sfa,
            tma_atom_sfb1,
            tma_tensor_sfb1,
            tma_atom_sfb2,
            tma_tensor_sfb2,
            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

    @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_b1: cute.CopyAtom,
        mB1_nkl: cute.Tensor,
        tma_atom_b2: cute.CopyAtom,
        mB2_nkl: cute.Tensor,
        tma_atom_sfa: cute.CopyAtom,
        mSFA_mkl: cute.Tensor,
        tma_atom_sfb1: cute.CopyAtom,
        mSFB1_nkl: cute.Tensor,
        tma_atom_sfb2: cute.CopyAtom,
        mSFB2_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,
    ):
        warp_idx = cute.arch.warp_idx()
        warp_idx = cute.arch.make_warp_uniform(warp_idx)

        #
        # Prefetch TMA descriptors
        #
        if warp_idx == self.tma_warp_id:
            cpasync.prefetch_descriptor(tma_atom_a)
            cpasync.prefetch_descriptor(tma_atom_b1)
            cpasync.prefetch_descriptor(tma_atom_b2)
            cpasync.prefetch_descriptor(tma_atom_sfa)
            cpasync.prefetch_descriptor(tma_atom_sfb1)
            cpasync.prefetch_descriptor(tma_atom_sfb2)
            cpasync.prefetch_descriptor(tma_atom_c)

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

        #
        # Setup CTA/thread coordinates
        #
        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
        )
        tidx, _, _ = cute.arch.thread_idx()

        #
        # Shared storage and pipelines
        #
        smem = utils.SmemAllocator()
        storage = smem.allocate(self.shared_storage)

        # A/B pipeline
        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,
            defer_sync=True,  # ---- PATCH
        )

        # Accumulator pipeline
        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,
            defer_sync=True,  # ---- PATCH
        )

        # TMEM allocator (identical pattern)
        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,
        )

        # NOTE:
        pipeline.pipeline_init_arrive(cluster_shape_mn=self.cluster_shape_mn, is_relaxed=True)

        #
        # Shared-memory tensors
        #
        sC = storage.sC.get_tensor(c_smem_layout_staged.outer, swizzle=c_smem_layout_staged.inner)
        sA = storage.sA.get_tensor(a_smem_layout_staged.outer, swizzle=a_smem_layout_staged.inner)
        sB = storage.sB.get_tensor(b_smem_layout_staged.outer, swizzle=b_smem_layout_staged.inner)
        sB2 = storage.sB2.get_tensor(b_smem_layout_staged.outer, swizzle=b_smem_layout_staged.inner)
        sSFA = storage.sSFA.get_tensor(sfa_smem_layout_staged)
        sSFB = storage.sSFB.get_tensor(sfb_smem_layout_staged)
        sSFB2 = storage.sSFB2.get_tensor(sfb_smem_layout_staged)

        #
        # Multicast masks for A/B/SFA/SFB
        #
        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 global tensors
        #
        gA_mkl = cute.local_tile(
            mA_mkl,
            cute.slice_(self.mma_tiler, (None, 0, None)),
            (None, None, None),
        )
        gB1_nkl = cute.local_tile(
            mB1_nkl,
            cute.slice_(self.mma_tiler, (0, None, None)),
            (None, None, None),
        )
        gB2_nkl = cute.local_tile(
            mB2_nkl,
            cute.slice_(self.mma_tiler, (0, None, None)),
            (None, None, None),
        )
        gSFA_mkl = cute.local_tile(
            mSFA_mkl,
            cute.slice_(self.mma_tiler, (None, 0, None)),
            (None, None, None),
        )
        gSFB1_nkl = cute.local_tile(
            mSFB1_nkl,
            cute.slice_(self.mma_tiler_sfb, (0, None, None)),
            (None, None, None),
        )
        gSFB2_nkl = cute.local_tile(
            mSFB2_nkl,
            cute.slice_(self.mma_tiler_sfb, (0, None, None)),
            (None, None, None),
        )
        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 tensors for MMA A/B/SFA/SFB/C
        #
        thr_mma = tiled_mma.get_slice(mma_tile_coord_v)
        thr_mma_sfb = tiled_mma_sfb.get_slice(mma_tile_coord_v)
        tCgA = thr_mma.partition_A(gA_mkl)
        tCgB1 = thr_mma.partition_B(gB1_nkl)
        tCgB2 = thr_mma.partition_B(gB2_nkl)
        tCgSFA = thr_mma.partition_A(gSFA_mkl)
        tCgSFB1 = thr_mma_sfb.partition_B(gSFB1_nkl)
        tCgSFB2 = thr_mma_sfb.partition_B(gSFB2_nkl)
        tCgC = thr_mma.partition_C(gC_mnl)

        #
        # Partition TMA A/B/SFA/SFB
        #
        a_cta_layout = cute.make_layout(cute.slice_(cluster_layout_vmnk, (0, 0, None, 0)).shape)
        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),
        )

        b_cta_layout = cute.make_layout(cute.slice_(cluster_layout_vmnk, (0, None, 0, 0)).shape)
        tBsB1, tBgB1 = cpasync.tma_partition(
            tma_atom_b1,
            block_in_cluster_coord_vmnk[1],
            b_cta_layout,
            cute.group_modes(sB, 0, 3),
            cute.group_modes(tCgB1, 0, 3),
        )
        tBsB2, tBgB2 = cpasync.tma_partition(
            tma_atom_b2,
            block_in_cluster_coord_vmnk[1],
            b_cta_layout,
            cute.group_modes(sB2, 0, 3),
            cute.group_modes(tCgB2, 0, 3),
        )

        sfa_cta_layout = a_cta_layout
        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)

        sfb_cta_layout = cute.make_layout(
            cute.slice_(cluster_layout_sfb_vmnk, (0, None, 0, 0)).shape
        )
        tBsSFB1, tBgSFB1 = cute.nvgpu.cpasync.tma_partition(
            tma_atom_sfb1,
            block_in_cluster_coord_sfb_vmnk[1],
            sfb_cta_layout,
            cute.group_modes(sSFB, 0, 3),
            cute.group_modes(tCgSFB1, 0, 3),
        )
        tBsSFB1 = cute.filter_zeros(tBsSFB1)
        tBgSFB1 = cute.filter_zeros(tBgSFB1)

        tBsSFB2, tBgSFB2 = cute.nvgpu.cpasync.tma_partition(
            tma_atom_sfb2,
            block_in_cluster_coord_sfb_vmnk[1],
            sfb_cta_layout,
            cute.group_modes(sSFB2, 0, 3),
            cute.group_modes(tCgSFB2, 0, 3),
        )
        tBsSFB2 = cute.filter_zeros(tBsSFB2)
        tBgSFB2 = cute.filter_zeros(tBgSFB2)

        #
        # Shared/TMEM partition for MMA A/B/Acc
        #
        tCrA = tiled_mma.make_fragment_A(sA)
        tCrB1 = tiled_mma.make_fragment_B(sB)
        tCrB2 = tiled_mma.make_fragment_B(sB2)
        acc_shape = tiled_mma.partition_shape_C(self.mma_tiler[:2])

        tCtAcc_fake = tiled_mma.make_fragment_C(cute.append(acc_shape, self.num_acc_stage))

        #
        # Cluster wait before TMEM alloc
        #
        # NOTE:
        pipeline.pipeline_init_wait(self.cluster_shape_mn)

        #
        # Specialized TMA warp
        #
        if warp_idx == self.tma_warp_id:
            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:
                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],
                )

                tAgA_slice = tAgA[(None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])]
                tBgB1_slice = tBgB1[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]
                tBgB2_slice = tBgB2[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]
                tAgSFA_slice = tAgSFA[(None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])]

                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
                tBgSFB1_slice = tBgSFB1[(None, slice_n, None, mma_tile_coord_mnl[2])]
                tBgSFB2_slice = tBgSFB2[(None, slice_n, None, mma_tile_coord_mnl[2])]

                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)

                for k_tile in cutlass.range(0, k_tile_cnt, 1, unroll=1):
                    ab_pipeline.producer_acquire(ab_producer_state, peek_ab_empty_status)
                    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_b1,
                        tBgB1_slice[(None, ab_producer_state.count)],
                        tBsB1[(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_b2,
                        tBgB2_slice[(None, ab_producer_state.count)],
                        tBsB2[(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_sfb1,
                        tBgSFB1_slice[(None, ab_producer_state.count)],
                        tBsSFB1[(None, ab_producer_state.index)],
                        tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
                        mcast_mask=sfb_full_mcast_mask,
                    )
                    cute.copy(
                        tma_atom_sfb2,
                        tBgSFB2_slice[(None, ab_producer_state.count)],
                        tBsSFB2[(None, ab_producer_state.index)],
                        tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
                        mcast_mask=sfb_full_mcast_mask,
                    )

                    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)

                tile_sched.advance_to_next_work()
                work_tile = tile_sched.get_current_work()

            ab_pipeline.producer_tail(ab_producer_state)

        #
        # Specialized MMA warp
        #
        if warp_idx == self.mma_warp_id:
            tmem.wait_for_alloc()

            acc_tmem_ptr = tmem.retrieve_ptr(self.acc_dtype)
            # Layout: [Acc1][Acc2][SFA][SFB1][SFB2]
            tCtAcc1_base = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)
            acc2_ptr = acc_tmem_ptr + self.num_accumulator_tmem_cols
            tCtAcc2_base = cute.make_tensor(acc2_ptr, tCtAcc_fake.layout)

            # SFA in TMEM
            sfa_tmem_ptr = cute.recast_ptr(
                acc_tmem_ptr + self.num_accumulator_tmem_cols + self.num_accumulator2_tmem_cols,
                dtype=self.sf_dtype,
            )
            tCtSFA_layout = blockscaled_utils.make_tmem_layout_sfa(
                tiled_mma,
                self.mma_tiler,
                sf_vec_size,
                cute.slice_(sfa_smem_layout_staged, (None, None, None, 0)),
            )
            tCtSFA = cute.make_tensor(sfa_tmem_ptr, tCtSFA_layout)

            base_cols = (
                self.num_accumulator_tmem_cols
                + self.num_accumulator2_tmem_cols
                + self.num_sfa_tmem_cols
            )

            # SFB in TMEM
            sfb1_tmem_ptr = cute.recast_ptr(
                acc_tmem_ptr + base_cols,
                dtype=self.sf_dtype,
            )
            tCtSFB_layout = blockscaled_utils.make_tmem_layout_sfb(
                tiled_mma,
                self.mma_tiler,
                sf_vec_size,
                cute.slice_(sfb_smem_layout_staged, (None, None, None, 0)),
            )
            tCtSFB1 = cute.make_tensor(sfb1_tmem_ptr, tCtSFB_layout)

            sfb2_tmem_ptr = cute.recast_ptr(
                acc_tmem_ptr + self.num_sfb_tmem_cols + base_cols,
                dtype=self.sf_dtype,
            )
            tCtSFB2 = cute.make_tensor(sfb2_tmem_ptr, tCtSFB_layout)

            (
                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, tCtSFB1)
            (
                tiled_copy_s2t_sfb2,
                tCsSFB2_compact_s2t,
                tCtSFB2_compact_s2t,
            ) = self.mainloop_s2t_copy_and_partition(sSFB2, tCtSFB2)

            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:
                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],
                )

                acc_stage_index = acc_producer_state.index
                tCtAcc1 = tCtAcc1_base[(None, None, None, acc_stage_index)]
                tCtAcc2 = tCtAcc2_base[(None, None, None, acc_stage_index)]

                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)

                if is_leader_cta:
                    acc_pipeline.producer_acquire(acc_producer_state)

                tCtSFB1_mma = tCtSFB1
                tCtSFB2_mma = tCtSFB2
                if cutlass.const_expr(self.cta_tile_shape_mnk[1] == 192):
                    offset = cutlass.Int32(2 if mma_tile_coord_mnl[1] % 2 == 1 else 0)
                    shifted_ptr = cute.recast_ptr(
                        acc_tmem_ptr
                        + self.num_accumulator_tmem_cols
                        + self.num_accumulator2_tmem_cols
                        + self.num_sfa_tmem_cols
                        + offset,
                        dtype=self.sf_dtype,
                    )
                    tCtSFB1_mma = cute.make_tensor(shifted_ptr, tCtSFB_layout)
                    shifted_ptr2 = cute.recast_ptr(
                        (acc_tmem_ptr + base_cols + offset + self.num_sfb_tmem_cols),
                        dtype=self.sf_dtype,
                    )
                    tCtSFB2_mma = cute.make_tensor(shifted_ptr2, tCtSFB_layout)
                elif cutlass.const_expr(self.cta_tile_shape_mnk[1] == 64):
                    offset = cutlass.Int32((mma_tile_coord_mnl[1] % 2) * 2)
                    shifted_ptr = cute.recast_ptr(
                        acc_tmem_ptr
                        + self.num_accumulator_tmem_cols
                        + self.num_accumulator2_tmem_cols
                        + self.num_sfa_tmem_cols
                        + offset,
                        dtype=self.sf_dtype,
                    )
                    tCtSFB1_mma = cute.make_tensor(shifted_ptr, tCtSFB_layout)
                    shifted_ptr2 = cute.recast_ptr(
                        (acc_tmem_ptr + base_cols + offset + self.num_sfb_tmem_cols),
                        dtype=self.sf_dtype,
                    )
                    tCtSFB2_mma = cute.make_tensor(shifted_ptr2, tCtSFB_layout)

                tiled_mma.set(tcgen05.Field.ACCUMULATE, False)

                for k_tile in range(k_tile_cnt):
                    if is_leader_cta:
                        ab_pipeline.consumer_wait(ab_consumer_state, peek_ab_full_status)

                        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]
                        tCsSFB2_compact_s2t_staged = tCsSFB2_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,
                        )
                        cute.copy(
                            tiled_copy_s2t_sfb2,
                            tCsSFB2_compact_s2t_staged,
                            tCtSFB2_compact_s2t,
                        )
                        cute.arch.fence_view_async_tmem_load()

                        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,
                            )

                            sf_kblock_coord = (None, None, kblock_idx)
                            tiled_mma.set(tcgen05.Field.SFA, tCtSFA[sf_kblock_coord].iterator)
                            # GEMM1: A @ B1
                            tiled_mma.set(
                                tcgen05.Field.SFB,
                                tCtSFB1_mma[sf_kblock_coord].iterator,
                            )
                            cute.gemm(
                                tiled_mma,
                                tCtAcc1,
                                tCrA[kblock_coord],
                                tCrB1[kblock_coord],
                                tCtAcc1,
                            )

                            # GEMM2: A @ B2
                            tiled_mma.set(
                                tcgen05.Field.SFB,
                                tCtSFB2_mma[sf_kblock_coord].iterator,
                            )
                            cute.gemm(
                                tiled_mma,
                                tCtAcc2,
                                tCrA[kblock_coord],
                                tCrB2[kblock_coord],
                                tCtAcc2,
                            )

                            tiled_mma.set(tcgen05.Field.ACCUMULATE, True)

                        ab_pipeline.consumer_release(ab_consumer_state)

                    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)

                if is_leader_cta:
                    acc_pipeline.producer_commit(acc_producer_state)
                acc_producer_state.advance()

                tile_sched.advance_to_next_work()
                work_tile = tile_sched.get_current_work()

            acc_pipeline.producer_tail(acc_producer_state)

        #
        # Specialized epilogue warps
        #
        if warp_idx < self.mma_warp_id:
            tmem.allocate(self.num_tmem_alloc_cols)
            tmem.wait_for_alloc()

            acc_tmem_ptr = tmem.retrieve_ptr(self.acc_dtype)
            tCtAcc1_base = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)
            acc2_ptr = acc_tmem_ptr + self.num_accumulator_tmem_cols
            tCtAcc2_base = cute.make_tensor(acc2_ptr, tCtAcc_fake.layout)

            epi_tidx = tidx
            (
                tiled_copy_t2r,
                tTR_tAcc1_base,
                tTR_rAcc1,
            ) = self.epilog_tmem_copy_and_partition(
                epi_tidx, tCtAcc1_base, tCgC, epi_tile, use_2cta_instrs
            )
            (
                tiled_copy_t2r_2,
                tTR_tAcc2_base,
                tTR_rAcc2,
            ) = self.epilog_tmem_copy_and_partition(
                epi_tidx, tCtAcc2_base, tCgC, epi_tile, use_2cta_instrs
            )

            tTR_rC = cute.make_rmem_tensor(tTR_rAcc1.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)

            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
            )

            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:
                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],
                )

                bSG_gC = bSG_gC_partitioned[
                    (
                        None,
                        None,
                        None,
                        *mma_tile_coord_mnl,
                    )
                ]

                acc_stage_index = acc_consumer_state.index
                reverse_subtile = cutlass.Boolean(False)

                tTR_tAcc1 = tTR_tAcc1_base[(None, None, None, None, None, acc_stage_index)]
                tTR_tAcc2 = tTR_tAcc2_base[(None, None, None, None, None, acc_stage_index)]

                acc_pipeline.consumer_wait(acc_consumer_state)

                tTR_tAcc1 = cute.group_modes(tTR_tAcc1, 3, cute.rank(tTR_tAcc1))
                tTR_tAcc2 = cute.group_modes(tTR_tAcc2, 3, cute.rank(tTR_tAcc2))
                bSG_gC = cute.group_modes(bSG_gC, 1, cute.rank(bSG_gC))

                subtile_cnt = cute.size(tTR_tAcc1.shape, mode=[3])
                num_prev_subtiles = tile_sched.num_tiles_executed * subtile_cnt
                for subtile_idx in cutlass.range(subtile_cnt):
                    real_subtile_idx = subtile_idx

                    tTR_tAcc1_mn = tTR_tAcc1[(None, None, None, real_subtile_idx)]
                    tTR_tAcc2_mn = tTR_tAcc2[(None, None, None, real_subtile_idx)]

                    cute.copy(tiled_copy_t2r, tTR_tAcc1_mn, tTR_rAcc1)
                    cute.copy(tiled_copy_t2r_2, tTR_tAcc2_mn, tTR_rAcc2)

                    x = tiled_copy_r2s.retile(tTR_rAcc1).load()
                    y = tiled_copy_r2s.retile(tTR_rAcc2).load()
                    out = silu(x) * y

                    out = epilogue_op(out.to(self.c_dtype))
                    tRS_rC.store(out)

                    c_buffer = (num_prev_subtiles + real_subtile_idx) % self.num_c_stage
                    cute.copy(
                        tiled_copy_r2s,
                        tRS_rC,
                        tRS_sC[(None, None, None, c_buffer)],
                    )
                    cute.arch.fence_proxy(
                        cute.arch.ProxyKind.async_shared,
                        space=cute.arch.SharedSpace.shared_cta,
                    )
                    self.epilog_sync_barrier.arrive_and_wait()

                    if warp_idx == self.epilog_warp_id[0]:
                        cute.copy(
                            tma_atom_c,
                            bSG_sC[(None, c_buffer)],
                            bSG_gC[(None, real_subtile_idx)],
                        )
                        c_pipeline.producer_commit()
                        c_pipeline.producer_acquire()
                    self.epilog_sync_barrier.arrive_and_wait()

                with cute.arch.elect_one():
                    acc_pipeline.consumer_release(acc_consumer_state)
                acc_consumer_state.advance()

                tile_sched.advance_to_next_work()
                work_tile = tile_sched.get_current_work()

            tmem.relinquish_alloc_permit()
            self.epilog_sync_barrier.arrive_and_wait()
            tmem.free(acc_tmem_ptr)
            c_pipeline.producer_tail()

    def mainloop_s2t_copy_and_partition(
        self,
        sSF: cute.Tensor,
        tSF: cute.Tensor,
    ) -> Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]:
        tCsSF_compact = cute.filter_zeros(sSF)
        tCtSF_compact = cute.filter_zeros(tSF)

        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)

        tCsSF_compact_s2t_ = thr_copy_s2t.partition_S(tCsSF_compact)
        tCsSF_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(tiled_copy_s2t, tCsSF_compact_s2t_)
        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]:
        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,
        )
        tAcc_epi = cute.flat_divide(
            tAcc[((None, None), 0, 0, None)],
            epi_tile,
        )
        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)
        tTR_tAcc = thr_copy_t2r.partition_S(tAcc_epi)

        gC_mnl_epi = cute.flat_divide(gC_mnl[((None, None), 0, 0, None, None, None)], epi_tile)
        tTR_gC = thr_copy_t2r.partition_D(gC_mnl_epi)
        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]:
        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)
        thr_copy_r2s = tiled_copy_r2s.get_slice(tidx)
        tRS_sC = thr_copy_r2s.partition_D(sC)
        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]:
        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)
        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]:
        num_acc_stage = 1 if mma_tiler_mnk[1] == 256 else 2
        num_c_stage = 2

        a_smem_layout_stage_one = sm100_utils.make_smem_layout_a(
            tiled_mma, mma_tiler_mnk, a_dtype, 1
        )
        b_smem_layout_staged_one = sm100_utils.make_smem_layout_b(
            tiled_mma, mma_tiler_mnk, b_dtype, 1
        )
        sfa_smem_layout_staged_one = blockscaled_utils.make_smem_layout_sfa(
            tiled_mma, mma_tiler_mnk, sf_vec_size, 1
        )
        sfb_smem_layout_staged_one = blockscaled_utils.make_smem_layout_sfb(
            tiled_mma, mma_tiler_mnk, sf_vec_size, 1
        )

        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)
            + 2 * cute.size_in_bytes(b_dtype, b_smem_layout_staged_one)
            + cute.size_in_bytes(sf_dtype, sfa_smem_layout_staged_one)
            + 2 * 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

        num_ab_stage = (
            smem_capacity // occupancy - (mbar_helpers_bytes + c_bytes)
        ) // ab_bytes_per_stage

        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]]:
        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


# ---------------------------------------------------------------------------
# Wrapper for GEMM task
# ---------------------------------------------------------------------------

ab_dtype = cutlass.Float4E2M1FN
sf_dtype = cutlass.Float8E4M3FN
c_dtype = cutlass.Float16
sf_vec_size = 16

mma_tiler_mn = (256, 64)
cluster_shape_mn = (2, 1)

gemm = Sm100BlockScaledPersistentDenseGemmKernel(
    mma_tiler_mn,
    cluster_shape_mn,
)

# Pattern A cache: specialize per (m, n, k, l)
_kernel_cache = {}


def compile_kernel_for_shape(m: int, n: int, k: int, l: int):
    global _kernel_cache

    key = (int(m), int(n), int(k), int(l))
    cached = _kernel_cache.get(key)
    if cached is not None:
        return cached

    # Dummy pointers for compilation
    a_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    b1_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    b2_ptr = make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    c_ptr = make_ptr(c_dtype, 0, cute.AddressSpace.gmem, assumed_align=16)
    sfa_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=32)
    sfb1_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=32)
    sfb2_ptr = make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=32)

    # Compile-time constants (layouts)
    a_major_mode = tcgen05.OperandMajorMode.K
    b_major_mode = tcgen05.OperandMajorMode.K
    c_layout = utils.LayoutEnum.ROW_MAJOR

    hardware_info = utils.HardwareInfo()
    max_active_clusters = hardware_info.get_max_active_clusters(
        cluster_shape_mn[0] * cluster_shape_mn[1]
    )

    compiled = cute.compile(
        gemm,
        a_ptr,
        b1_ptr,
        b2_ptr,
        sfa_ptr,
        sfb1_ptr,
        sfb2_ptr,
        c_ptr,
        (a_major_mode, b_major_mode, c_layout),
        key,  # specialize on (m, n, k, l) at compile time
        max_active_clusters,
        options="--opt-level 2 --enable-tvm-ffi",
    )

    _kernel_cache[key] = compiled
    return compiled


def custom_kernel(data: input_t) -> output_t:
    a, b1, b2, _, _, _, sfa_permuted, sfb1_permuted, sfb2_permuted, c = data

    m, k, l = a.shape
    n, _, _ = b1.shape

    k = k * 2

    compiled_gemm = compile_kernel_for_shape(m, n, k, l)

    a_ptr = make_ptr(ab_dtype, a.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    b1_ptr = make_ptr(ab_dtype, b1.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    b2_ptr = make_ptr(ab_dtype, b2.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    c_ptr = make_ptr(c_dtype, c.data_ptr(), cute.AddressSpace.gmem, assumed_align=16)
    sfa_ptr = make_ptr(sf_dtype, sfa_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=32)
    sfb1_ptr = make_ptr(
        sf_dtype, sfb1_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=32
    )
    sfb2_ptr = make_ptr(
        sf_dtype, sfb2_permuted.data_ptr(), cute.AddressSpace.gmem, assumed_align=32
    )

    compiled_gemm(a_ptr, b1_ptr, b2_ptr, sfa_ptr, sfb1_ptr, sfb2_ptr, c_ptr)

    return c