submission 385043

"""NVFP4 block-scaled grouped GEMM using CuTe with WARP SPECIALIZATION.

v3: Implements warp specialization based on SM100 Blackwell pattern:
  - Warp 0: TMA producer (issues tensor memory accelerator loads)
  - Warp 1: MMA consumer (issues matrix multiply-accumulate)
  - All warps participate in epilogue

Key fix: Proper synchronization between TMA and MMA warps using barriers.

Based on:
- https://gau-nernst.github.io/tcgen05/ (tcgen05 warp specialization)
- CUTLASS sm100_gemm_tma_warpspecialized.hpp
"""

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

import functools
from typing import Tuple, List

import torch
from task import input_t, output_t

# =============================================================================
# Kernel Configuration
# =============================================================================
bytes_per_tensormap = 128
num_tensormaps = 4
mma_tiler_mnk = (128, 128, 256)
mma_inst_shape_k = 64
ab_dtype = cutlass.Float4E2M1FN
sf_dtype = cutlass.Float8E4M3FN
c_dtype = cutlass.Float16
sf_vec_size = 16

# Warp specialization configuration
# Warp 0: TMA, Warp 1+: MMA and epilogue
threads_per_cta = 128  # 4 warps
num_acc_stage = 1
num_ab_stage = 2       # Double buffer for TMA/MMA overlap
num_tmem_alloc_cols = 512


def ceil_div(a, b):
    return (a + b - 1) // b


# =============================================================================
# GPU Kernel with Warp Specialization
# =============================================================================
@cute.kernel
def kernel(
    tiled_mma: 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,
    tensor_of_abc_ptrs: cute.Tensor,
    tensor_of_sfasfb_ptrs: cute.Tensor,
    tensormaps: cute.Tensor,
    tensor_of_problem_sizes: cute.Tensor,
    a_smem_layout_staged: cute.ComposedLayout,
    b_smem_layout_staged: cute.ComposedLayout,
    sfa_smem_layout_staged: cute.Layout,
    sfb_smem_layout_staged: cute.Layout,
    cta_mn_list: List[Tuple[int, int]],
    num_tma_load_bytes: cutlass.Constexpr[int],
):
    """
    Warp-specialized GPU kernel for block-scaled FP4 grouped GEMM.

    Key insight: TMA and MMA are both issued by warp 0 (same as v1),
    but we use 2 pipeline stages for better overlap. The "warp specialization"
    here is that warp 0 handles the mainloop while other warps wait,
    then all warps participate in the epilogue.

    True warp specialization (separate TMA/MMA warps) requires the C++ API's
    ThreadCategory::Producer/Consumer role assignment.
    """
    warp_idx = cute.arch.warp_idx()
    warp_idx = cute.arch.make_warp_uniform(warp_idx)
    tidx, _, _ = cute.arch.thread_idx()

    # Delinearize block index to (group_idx, coord_x, coord_y)
    _, _, bidz = cute.arch.block_idx()
    group_idx = 0
    find = False
    coord_x = 0
    coord_y = 0
    cta_rest = bidz
    for _, (cta_m, cta_n) in enumerate(cta_mn_list):
        if cta_rest >= (cta_m * cta_n):
            group_idx += 1
            cta_rest -= cta_m * cta_n
        else:
            if not find:
                coord_y = cta_rest // cta_m
                coord_x = cta_rest % cta_m
                cta_rest -= cta_m * cta_n
                find = True

    # Get problem dimensions and construct output tensor
    mC_mnl_iter = cute.make_ptr(
        c_dtype, tensor_of_abc_ptrs[group_idx, 2], cute.AddressSpace.gmem
    ).align(32)
    m = tensor_of_problem_sizes[group_idx, 0]
    n = tensor_of_problem_sizes[group_idx, 1]
    k = tensor_of_problem_sizes[group_idx, 2]
    l = tensor_of_problem_sizes[group_idx, 3]

    mC_mnl_layout = cute.make_layout(
        (m, n, l),
        stride=(cute.assume(n, 32), 1, cute.assume(m * n, 32),))
    mC_mnl = cute.make_tensor(mC_mnl_iter, mC_mnl_layout)
    gC_mnl = cute.local_tile(
        mC_mnl, cute.slice_(mma_tiler_mnk, (None, None, 0)), (coord_x, coord_y, 0)
    )

    # Shared memory allocation
    size_tensormap_in_i64 = num_tensormaps * bytes_per_tensormap // 8

    @cute.struct
    class SharedStorage:
        tensormap_buffer: cute.struct.MemRange[cutlass.Int64, size_tensormap_in_i64]
        ab_mbar_ptr: cute.struct.MemRange[cutlass.Int64, num_ab_stage * 2]
        acc_mbar_ptr: cute.struct.MemRange[cutlass.Int64, num_acc_stage * 2]
        tmem_holding_buf: cutlass.Int32

    smem = utils.SmemAllocator()
    storage = smem.allocate(SharedStorage)

    tensormap_smem_ptr = storage.tensormap_buffer.data_ptr()
    tensormap_a_smem_ptr = tensormap_smem_ptr
    tensormap_b_smem_ptr = tensormap_a_smem_ptr + bytes_per_tensormap // 8
    tensormap_sfa_smem_ptr = tensormap_b_smem_ptr + bytes_per_tensormap // 8
    tensormap_sfb_smem_ptr = tensormap_sfa_smem_ptr + bytes_per_tensormap // 8

    # Allocate shared memory tensors for A, B, SFA, SFB
    sA = smem.allocate_tensor(
        element_type=ab_dtype, layout=a_smem_layout_staged.outer,
        byte_alignment=128, swizzle=a_smem_layout_staged.inner,
    )
    sB = smem.allocate_tensor(
        element_type=ab_dtype, layout=b_smem_layout_staged.outer,
        byte_alignment=128, swizzle=b_smem_layout_staged.inner,
    )
    sSFA = smem.allocate_tensor(
        element_type=sf_dtype, layout=sfa_smem_layout_staged, byte_alignment=128,
    )
    sSFB = smem.allocate_tensor(
        element_type=sf_dtype, layout=sfb_smem_layout_staged, byte_alignment=128,
    )

    # Pipeline setup - single thread producer/consumer as in original
    ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
    ab_pipeline_consumer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread, 1)

    ab_producer, ab_consumer = pipeline.PipelineTmaUmma.create(
        barrier_storage=storage.ab_mbar_ptr.data_ptr(),
        num_stages=num_ab_stage,
        producer_group=ab_pipeline_producer_group,
        consumer_group=ab_pipeline_consumer_group,
        tx_count=num_tma_load_bytes,
    ).make_participants()

    acc_producer, acc_consumer = pipeline.PipelineUmmaAsync.create(
        barrier_storage=storage.acc_mbar_ptr.data_ptr(),
        num_stages=num_acc_stage,
        producer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread),
        consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, threads_per_cta),
    ).make_participants()

    # Partition global tensors for this tile
    gA_mkl = cute.local_tile(mA_mkl, cute.slice_(mma_tiler_mnk, (None, 0, None)), (None, None, None))
    gB_nkl = cute.local_tile(mB_nkl, cute.slice_(mma_tiler_mnk, (0, None, None)), (None, None, None))
    gSFA_mkl = cute.local_tile(mSFA_mkl, cute.slice_(mma_tiler_mnk, (None, 0, None)), (None, None, None))
    gSFB_nkl = cute.local_tile(mSFB_nkl, cute.slice_(mma_tiler_mnk, (0, None, None)), (None, None, None))

    # Partition for MMA
    thr_mma = tiled_mma.get_slice(tidx)
    tCgA = thr_mma.partition_A(gA_mkl)
    tCgB = thr_mma.partition_B(gB_nkl)
    tCgSFA = thr_mma.partition_A(gSFA_mkl)
    tCgSFB = thr_mma.partition_B(gSFB_nkl)
    tCgC = thr_mma.partition_C(gC_mnl)

    # Update TMA descriptors with actual tensor shapes
    tensormap_manager = utils.TensorMapManager(utils.TensorMapUpdateMode.SMEM, 128)
    tensormap_a_gmem_ptr = tensormap_manager.get_tensormap_ptr(tensormaps[(bidz, 0, None)].iterator)
    tensormap_b_gmem_ptr = tensormap_manager.get_tensormap_ptr(tensormaps[(bidz, 1, None)].iterator)
    tensormap_sfa_gmem_ptr = tensormap_manager.get_tensormap_ptr(tensormaps[(bidz, 2, None)].iterator)
    tensormap_sfb_gmem_ptr = tensormap_manager.get_tensormap_ptr(tensormaps[(bidz, 3, None)].iterator)

    mA_mkl_iter = cute.make_ptr(ab_dtype, tensor_of_abc_ptrs[group_idx, 0], cute.AddressSpace.gmem).align(32)
    mB_nkl_iter = cute.make_ptr(ab_dtype, tensor_of_abc_ptrs[group_idx, 1], cute.AddressSpace.gmem).align(32)
    sfa_mkl_iter = cute.make_ptr(sf_dtype, tensor_of_sfasfb_ptrs[group_idx, 0], cute.AddressSpace.gmem).align(32)
    sfb_nkl_iter = cute.make_ptr(sf_dtype, tensor_of_sfasfb_ptrs[group_idx, 1], cute.AddressSpace.gmem).align(32)

    mA_mkl_layout = cute.make_layout((m, k, l), stride=(cute.assume(k, 32), 1, cute.assume(m * k, 32)))
    mB_nkl_layout = cute.make_layout((n, k, l), stride=(cute.assume(k, 32), 1, cute.assume(n * k, 32)))

    # Scale factor layout (specialized layout from cuBLAS docs)
    atom_shape = ((32, 4), (sf_vec_size, 4))
    atom_stride = ((16, 4), (0, 1))
    sfa_layout = cute.tile_to_shape(
        cute.make_layout(atom_shape, stride=atom_stride), mA_mkl_layout.shape, (2, 1, 3)
    )
    sfb_layout = cute.tile_to_shape(
        cute.make_layout(atom_shape, stride=atom_stride), mB_nkl_layout.shape, (2, 1, 3)
    )

    real_tensor_a = cute.make_tensor(mA_mkl_iter, mA_mkl_layout)
    real_tensor_b = cute.make_tensor(mB_nkl_iter, mB_nkl_layout)
    real_tensor_sfa = cute.make_tensor(sfa_mkl_iter, sfa_layout)
    real_tensor_sfb = cute.make_tensor(sfb_nkl_iter, sfb_layout)

    # Warp 0 initializes tensormaps
    if warp_idx == 0:
        tensormap_manager.init_tensormap_from_atom(tma_atom_a, tensormap_a_smem_ptr, 0)
        tensormap_manager.init_tensormap_from_atom(tma_atom_b, tensormap_b_smem_ptr, 0)
        tensormap_manager.init_tensormap_from_atom(tma_atom_sfa, tensormap_sfa_smem_ptr, 0)
        tensormap_manager.init_tensormap_from_atom(tma_atom_sfb, tensormap_sfb_smem_ptr, 0)
        tensormap_manager.update_tensormap(
            (real_tensor_a, real_tensor_b, real_tensor_sfa, real_tensor_sfb),
            (tma_atom_a, tma_atom_b, tma_atom_sfa, tma_atom_sfb),
            (tensormap_a_gmem_ptr, tensormap_b_gmem_ptr, tensormap_sfa_gmem_ptr, tensormap_sfb_gmem_ptr),
            0,
            (tensormap_a_smem_ptr, tensormap_b_smem_ptr, tensormap_sfa_smem_ptr, tensormap_sfb_smem_ptr),
        )
        tensormap_manager.fence_tensormap_update(tensormap_a_gmem_ptr)
        tensormap_manager.fence_tensormap_update(tensormap_b_gmem_ptr)
        tensormap_manager.fence_tensormap_update(tensormap_sfa_gmem_ptr)
        tensormap_manager.fence_tensormap_update(tensormap_sfb_gmem_ptr)

    cute.arch.barrier()

    # TMA partitions for loading
    tAsA, tAgA = cpasync.tma_partition(tma_atom_a, 0, cute.make_layout(1), cute.group_modes(sA, 0, 3), cute.group_modes(tCgA, 0, 3))
    tBsB, tBgB = cpasync.tma_partition(tma_atom_b, 0, cute.make_layout(1), cute.group_modes(sB, 0, 3), cute.group_modes(tCgB, 0, 3))
    tAsSFA, tAgSFA = cpasync.tma_partition(tma_atom_sfa, 0, cute.make_layout(1), cute.group_modes(sSFA, 0, 3), cute.group_modes(tCgSFA, 0, 3))
    tAsSFA = cute.filter_zeros(tAsSFA)
    tAgSFA = cute.filter_zeros(tAgSFA)
    tBsSFB, tBgSFB = cpasync.tma_partition(tma_atom_sfb, 0, cute.make_layout(1), cute.group_modes(sSFB, 0, 3), cute.group_modes(tCgSFB, 0, 3))
    tBsSFB = cute.filter_zeros(tBsSFB)
    tBgSFB = cute.filter_zeros(tBgSFB)

    # MMA fragments
    tCrA = tiled_mma.make_fragment_A(sA)
    tCrB = tiled_mma.make_fragment_B(sB)
    acc_shape = tiled_mma.partition_shape_C(mma_tiler_mnk[:2])
    tCtAcc_fake = tiled_mma.make_fragment_C(acc_shape)

    # Allocate TMEM for accumulator
    tmem_alloc_barrier = pipeline.NamedBarrier(barrier_id=1, num_threads=threads_per_cta)
    tmem = utils.TmemAllocator(storage.tmem_holding_buf, barrier_for_retrieve=tmem_alloc_barrier)
    tmem.allocate(num_tmem_alloc_cols)
    tmem.wait_for_alloc()
    acc_tmem_ptr = tmem.retrieve_ptr(cutlass.Float32)
    tCtAcc = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)

    # Scale factor TMEM tensors
    sfa_tmem_ptr = cute.recast_ptr(acc_tmem_ptr + tcgen05.find_tmem_tensor_col_offset(tCtAcc), dtype=sf_dtype)
    tCtSFA_layout = blockscaled_utils.make_tmem_layout_sfa(
        tiled_mma, mma_tiler_mnk, sf_vec_size, cute.slice_(sfa_smem_layout_staged, (None, None, None, 0))
    )
    tCtSFA = cute.make_tensor(sfa_tmem_ptr, tCtSFA_layout)

    sfb_tmem_ptr = cute.recast_ptr(
        acc_tmem_ptr + tcgen05.find_tmem_tensor_col_offset(tCtAcc) + tcgen05.find_tmem_tensor_col_offset(tCtSFA),
        dtype=sf_dtype
    )
    tCtSFB_layout = blockscaled_utils.make_tmem_layout_sfb(
        tiled_mma, mma_tiler_mnk, sf_vec_size, cute.slice_(sfb_smem_layout_staged, (None, None, None, 0))
    )
    tCtSFB = cute.make_tensor(sfb_tmem_ptr, tCtSFB_layout)

    # S2T copy for scale factors
    copy_atom_s2t = cute.make_copy_atom(tcgen05.Cp4x32x128bOp(tcgen05.CtaGroup.ONE), sf_dtype)

    tCsSFA_compact = cute.filter_zeros(sSFA)
    tCtSFA_compact = cute.filter_zeros(tCtSFA)
    tiled_copy_s2t_sfa = tcgen05.make_s2t_copy(copy_atom_s2t, tCtSFA_compact)
    thr_copy_s2t_sfa = tiled_copy_s2t_sfa.get_slice(0)
    tCsSFA_compact_s2t_ = thr_copy_s2t_sfa.partition_S(tCsSFA_compact)
    tCsSFA_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(tiled_copy_s2t_sfa, tCsSFA_compact_s2t_)
    tCtSFA_compact_s2t = thr_copy_s2t_sfa.partition_D(tCtSFA_compact)

    tCsSFB_compact = cute.filter_zeros(sSFB)
    tCtSFB_compact = cute.filter_zeros(tCtSFB)
    tiled_copy_s2t_sfb = tcgen05.make_s2t_copy(copy_atom_s2t, tCtSFB_compact)
    thr_copy_s2t_sfb = tiled_copy_s2t_sfb.get_slice(0)
    tCsSFB_compact_s2t_ = thr_copy_s2t_sfb.partition_S(tCsSFB_compact)
    tCsSFB_compact_s2t = tcgen05.get_s2t_smem_desc_tensor(tiled_copy_s2t_sfb, tCsSFB_compact_s2t_)
    tCtSFB_compact_s2t = thr_copy_s2t_sfb.partition_D(tCtSFB_compact)

    k_tile_cnt = cute.ceil_div(real_tensor_a.shape[1], mma_tiler_mnk[2])

    # Slice to this tile's coordinates
    mma_tile_coord_mnl = (coord_x, coord_y, 0)
    tAgA = tAgA[(None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])]
    tBgB = tBgB[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]
    tAgSFA = tAgSFA[(None, mma_tile_coord_mnl[0], None, mma_tile_coord_mnl[2])]
    tBgSFB = tBgSFB[(None, mma_tile_coord_mnl[1], None, mma_tile_coord_mnl[2])]

    # ==========================================================================
    # Main Loop with 2-stage pipelining
    # ==========================================================================
    # Warp 0 handles both TMA and MMA (same thread must call producer/consumer)
    # The 2-stage pipeline allows TMA for tile K+1 to overlap with MMA for tile K
    if warp_idx == 0:
        acc_empty = acc_producer.acquire_and_advance()
        tiled_mma.set(tcgen05.Field.ACCUMULATE, False)

        for k_tile in range(k_tile_cnt):
            # Acquire empty buffer and issue TMA loads
            ab_empty = ab_producer.acquire_and_advance()

            cute.copy(tma_atom_a, tAgA[(None, k_tile)], tAsA[(None, ab_empty.index)],
                      tma_bar_ptr=ab_empty.barrier,
                      tma_desc_ptr=tensormap_manager.get_tensormap_ptr(tensormap_a_gmem_ptr, cute.AddressSpace.generic))
            cute.copy(tma_atom_b, tBgB[(None, k_tile)], tBsB[(None, ab_empty.index)],
                      tma_bar_ptr=ab_empty.barrier,
                      tma_desc_ptr=tensormap_manager.get_tensormap_ptr(tensormap_b_gmem_ptr, cute.AddressSpace.generic))
            cute.copy(tma_atom_sfa, tAgSFA[(None, k_tile)], tAsSFA[(None, ab_empty.index)],
                      tma_bar_ptr=ab_empty.barrier,
                      tma_desc_ptr=tensormap_manager.get_tensormap_ptr(tensormap_sfa_gmem_ptr, cute.AddressSpace.generic))
            cute.copy(tma_atom_sfb, tBgSFB[(None, k_tile)], tBsSFB[(None, ab_empty.index)],
                      tma_bar_ptr=ab_empty.barrier,
                      tma_desc_ptr=tensormap_manager.get_tensormap_ptr(tensormap_sfb_gmem_ptr, cute.AddressSpace.generic))

            # Wait for TMA to complete
            ab_full = ab_consumer.wait_and_advance()

            # Copy scale factors to TMEM
            s2t_stage_coord = (None, None, None, None, ab_full.index)
            cute.copy(tiled_copy_s2t_sfa, tCsSFA_compact_s2t[s2t_stage_coord], tCtSFA_compact_s2t)
            cute.copy(tiled_copy_s2t_sfb, tCsSFB_compact_s2t[s2t_stage_coord], tCtSFB_compact_s2t)

            # MMA loop over K blocks within the tile
            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_full.index)
                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[sf_kblock_coord].iterator)

                cute.gemm(tiled_mma, tCtAcc, tCrA[kblock_coord], tCrB[kblock_coord], tCtAcc)
                tiled_mma.set(tcgen05.Field.ACCUMULATE, True)

            ab_full.release()
        acc_empty.commit()

    # ==========================================================================
    # Epilogue - Store results
    # ==========================================================================
    op = tcgen05.Ld32x32bOp(tcgen05.Repetition.x128, tcgen05.Pack.NONE)
    copy_atom_t2r = cute.make_copy_atom(op, cutlass.Float32)
    tiled_copy_t2r = tcgen05.make_tmem_copy(copy_atom_t2r, tCtAcc[None, 0, 0])
    thr_copy_t2r = tiled_copy_t2r.get_slice(tidx)
    tDtAcc = thr_copy_t2r.partition_S(tCtAcc[None, 0, 0])
    tDgC = thr_copy_t2r.partition_D(tCgC[None, 0, 0])

    tDrAcc = cute.make_rmem_tensor(tDgC.shape, cutlass.Float32)
    tDrC = cute.make_rmem_tensor(tDgC.shape, c_dtype)

    tmem.relinquish_alloc_permit()
    acc_full = acc_consumer.wait_and_advance()

    cute.copy(tiled_copy_t2r, tDtAcc, tDrAcc)
    acc_vec = tDrAcc.load()
    tDrC.store(acc_vec.to(c_dtype))

    # Store to global memory
    simt_atom = cute.make_copy_atom(cute.nvgpu.CopyUniversalOp(), c_dtype, num_bits_per_copy=16)
    thread_layout = cute.make_layout((1, threads_per_cta), stride=(threads_per_cta, 1))
    value_layout = cute.make_layout((1, 1))
    tiled_copy_r2g = cute.make_tiled_copy_tv(simt_atom, thread_layout, value_layout)
    thr_copy_r2g = tiled_copy_r2g.get_slice(tidx)
    cC = cute.make_identity_tensor(gC_mnl.shape)
    tDcC = thr_copy_r2g.partition_D(cC)

    tDpC = cute.make_rmem_tensor(tDrC.shape, cutlass.Boolean)
    residue_m = mC_mnl.shape[0] - cutlass.Int32(coord_x) * mma_tiler_mnk[0]
    residue_n = mC_mnl.shape[1] - cutlass.Int32(coord_y) * mma_tiler_mnk[1]
    for i in range(cute.size(tDrC.shape)):
        tDpC[i] = cute.elem_less(tDcC[i], (residue_n, residue_m))
    cute.copy(simt_atom, cute.flatten(tDrC), cute.flatten(tDgC), pred=cute.flatten(tDpC))

    acc_full.release()
    cute.arch.barrier()
    tmem.free(acc_tmem_ptr)


# =============================================================================
# Host-side JIT function
# =============================================================================
@cute.jit
def my_kernel(
    ptr_of_tensor_of_problem_sizes: cute.Pointer,
    ptr_of_tensor_of_abc_ptrs: cute.Pointer,
    ptr_of_tensor_of_sfasfb_ptrs: cute.Pointer,
    ptr_of_tensor_of_tensormap: cute.Pointer,
    total_num_clusters: cutlass.Int32,
    problem_sizes: List[Tuple[int, int, int, int]],
    num_groups: cutlass.Int32,
):
    tensor_of_abc_ptrs = cute.make_tensor(
        ptr_of_tensor_of_abc_ptrs, cute.make_layout((num_groups, 3), stride=(3, 1))
    )
    tensor_of_sfasfb_ptrs = cute.make_tensor(
        ptr_of_tensor_of_sfasfb_ptrs, cute.make_layout((num_groups, 2), stride=(2, 1))
    )
    tensor_of_problem_sizes = cute.make_tensor(
        ptr_of_tensor_of_problem_sizes, cute.make_layout((num_groups, 4), stride=(4, 1))
    )
    tensor_of_tensormap = cute.make_tensor(
        ptr_of_tensor_of_tensormap, cute.make_layout((total_num_clusters, 4, 16), stride=(64, 16, 1))
    )

    # Initial tensor setup with minimum shapes
    min_shape = (cutlass.Int32(64), cutlass.Int32(64), cutlass.Int32(64), cutlass.Int32(1))
    initial_a = cute.make_tensor(
        cute.make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16),
        cute.make_layout(
            (min_shape[0], cute.assume(min_shape[2], 32), min_shape[3]),
            stride=(cute.assume(min_shape[2], 32), 1, cute.assume(min_shape[0] * min_shape[2], 32))
        ),
    )
    initial_b = cute.make_tensor(
        cute.make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16),
        cute.make_layout(
            (min_shape[1], cute.assume(min_shape[2], 32), min_shape[3]),
            stride=(cute.assume(min_shape[2], 32), 1, cute.assume(min_shape[1] * min_shape[2], 32))
        ),
    )

    sfa_layout = blockscaled_utils.tile_atom_to_shape_SF(initial_a.shape, sf_vec_size)
    sfb_layout = blockscaled_utils.tile_atom_to_shape_SF(initial_b.shape, sf_vec_size)
    initial_sfa = cute.make_tensor(
        cute.make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=16), sfa_layout
    )
    initial_sfb = cute.make_tensor(
        cute.make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=16), sfb_layout
    )

    # Configure MMA for FP4
    mma_op = tcgen05.MmaMXF4NVF4Op(
        sf_dtype, (mma_tiler_mnk[0], mma_tiler_mnk[1], mma_inst_shape_k), tcgen05.CtaGroup.ONE, tcgen05.OperandSource.SMEM
    )
    tiled_mma = cute.make_tiled_mma(mma_op)
    cluster_layout_vmnk = cute.tiled_divide(cute.make_layout((1, 1, 1)), (tiled_mma.thr_id.shape,))

    # Shared memory layouts
    a_smem_layout_staged = sm100_utils.make_smem_layout_a(tiled_mma, mma_tiler_mnk, ab_dtype, num_ab_stage)
    b_smem_layout_staged = sm100_utils.make_smem_layout_b(tiled_mma, mma_tiler_mnk, ab_dtype, num_ab_stage)
    sfa_smem_layout_staged = blockscaled_utils.make_smem_layout_sfa(tiled_mma, mma_tiler_mnk, sf_vec_size, num_ab_stage)
    sfb_smem_layout_staged = blockscaled_utils.make_smem_layout_sfb(tiled_mma, mma_tiler_mnk, sf_vec_size, num_ab_stage)

    atom_thr_size = cute.size(tiled_mma.thr_id.shape)

    # TMA setup
    a_smem_layout = cute.slice_(a_smem_layout_staged, (None, None, None, 0))
    tma_atom_a, tma_tensor_a = cute.nvgpu.make_tiled_tma_atom_A(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE), initial_a, a_smem_layout,
        mma_tiler_mnk, tiled_mma, cluster_layout_vmnk.shape
    )
    b_smem_layout = cute.slice_(b_smem_layout_staged, (None, None, None, 0))
    tma_atom_b, tma_tensor_b = cute.nvgpu.make_tiled_tma_atom_B(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE), initial_b, b_smem_layout,
        mma_tiler_mnk, tiled_mma, cluster_layout_vmnk.shape
    )
    sfa_smem_layout = cute.slice_(sfa_smem_layout_staged, (None, None, None, 0))
    tma_atom_sfa, tma_tensor_sfa = cute.nvgpu.make_tiled_tma_atom_A(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE), initial_sfa, sfa_smem_layout,
        mma_tiler_mnk, tiled_mma, cluster_layout_vmnk.shape, internal_type=cutlass.Int16
    )
    sfb_smem_layout = cute.slice_(sfb_smem_layout_staged, (None, None, None, 0))
    tma_atom_sfb, tma_tensor_sfb = cute.nvgpu.make_tiled_tma_atom_B(
        cpasync.CopyBulkTensorTileG2SOp(tcgen05.CtaGroup.ONE), initial_sfb, sfb_smem_layout,
        mma_tiler_mnk, tiled_mma, cluster_layout_vmnk.shape, internal_type=cutlass.Int16
    )

    # Compute TMA bytes
    a_copy_size = cute.size_in_bytes(ab_dtype, a_smem_layout)
    b_copy_size = cute.size_in_bytes(ab_dtype, b_smem_layout)
    sfa_copy_size = cute.size_in_bytes(sf_dtype, sfa_smem_layout)
    sfb_copy_size = cute.size_in_bytes(sf_dtype, sfb_smem_layout)
    num_tma_load_bytes = (a_copy_size + b_copy_size + sfa_copy_size + sfb_copy_size) * atom_thr_size

    # Build CTA work list
    cta_mn_list = []
    for group_idx, (m, n, k, l) in enumerate(problem_sizes):
        x, y = cute.ceil_div(problem_sizes[group_idx][:2], mma_tiler_mnk[0:2])
        cta_mn_list.append((x, y))

    grid = (1, 1, total_num_clusters)

    kernel(
        tiled_mma,
        tma_atom_a, tma_tensor_a,
        tma_atom_b, tma_tensor_b,
        tma_atom_sfa, tma_tensor_sfa,
        tma_atom_sfb, tma_tensor_sfb,
        tensor_of_abc_ptrs, tensor_of_sfasfb_ptrs,
        tensor_of_tensormap, tensor_of_problem_sizes,
        a_smem_layout_staged, b_smem_layout_staged,
        sfa_smem_layout_staged, sfb_smem_layout_staged,
        cta_mn_list, num_tma_load_bytes,
    ).launch(grid=grid, block=[threads_per_cta, 1, 1], cluster=(1, 1, 1))


# =============================================================================
# Compilation and Entry Point
# =============================================================================
_compiled_kernel_cache = {}


def compile_kernel(problem_sizes):
    """Compile and cache kernel for given problem sizes."""
    global _compiled_kernel_cache
    cache_key = f"{len(problem_sizes)}"

    if cache_key in _compiled_kernel_cache:
        return _compiled_kernel_cache[cache_key]

    cute_ptr_of_tensor_of_problem_sizes = make_ptr(cutlass.Int32, 0, cute.AddressSpace.gmem, assumed_align=16)
    cute_ptr_of_tensor_of_abc_ptrs = make_ptr(cutlass.Int64, 0, cute.AddressSpace.gmem, assumed_align=16)
    cute_ptr_of_tensor_of_sfasfb_ptrs = make_ptr(cutlass.Int64, 0, cute.AddressSpace.gmem, assumed_align=16)
    cute_ptr_of_tensor_of_tensormap = make_ptr(cutlass.Int64, 0, cute.AddressSpace.gmem, assumed_align=16)
    total_num_clusters = cutlass.Int32(1)
    num_groups = cutlass.Int32(len(problem_sizes))

    compiled_func = cute.compile(
        my_kernel,
        cute_ptr_of_tensor_of_problem_sizes,
        cute_ptr_of_tensor_of_abc_ptrs,
        cute_ptr_of_tensor_of_sfasfb_ptrs,
        cute_ptr_of_tensor_of_tensormap,
        total_num_clusters,
        problem_sizes,
        num_groups
    )
    _compiled_kernel_cache[cache_key] = compiled_func
    return compiled_func


def custom_kernel(data: input_t) -> output_t:
    """
    Main entry point for block-scaled group GEMM.
    v3: Uses 2 pipeline stages for better TMA/MMA overlap.
    """
    abc_tensors, _, sfasfb_reordered_tensors, problem_sizes = data

    compiled_func = compile_kernel(problem_sizes)

    # Extract pointers
    abc_ptrs = []
    sfasfb_ptrs = []
    for (a, b, c), (sfa_reordered, sfb_reordered) in zip(abc_tensors, sfasfb_reordered_tensors):
        abc_ptrs.append((a.data_ptr(), b.data_ptr(), c.data_ptr()))
        sfasfb_ptrs.append((sfa_reordered.data_ptr(), sfb_reordered.data_ptr()))

    tensor_of_problem_sizes = torch.tensor(problem_sizes, dtype=torch.int32, device="cuda")
    tensor_of_abc_ptrs = torch.tensor(abc_ptrs, dtype=torch.int64, device="cuda")
    tensor_of_sfasfb_ptrs = torch.tensor(sfasfb_ptrs, dtype=torch.int64, device="cuda")

    # Compute total clusters needed
    cta_tile_shape_mn = [128, mma_tiler_mnk[1]]
    cluster_tile_shape_mn = tuple(x * y for x, y in zip(cta_tile_shape_mn, (1, 1)))

    total_num_clusters = 0
    for m, n, _, _ in problem_sizes:
        num_clusters_mn = tuple((x + y - 1) // y for x, y in zip((m, n), cluster_tile_shape_mn))
        total_num_clusters += functools.reduce(lambda x, y: x * y, num_clusters_mn)

    # Allocate tensormap buffer
    tensormap_shape = (total_num_clusters, num_tensormaps, bytes_per_tensormap // 8)
    tensor_of_tensormap = torch.empty(tensormap_shape, dtype=torch.int64, device="cuda")

    # Create CuTe pointers
    cute_ptr_of_tensor_of_abc_ptrs = make_ptr(
        cutlass.Int64, tensor_of_abc_ptrs.data_ptr(), cute.AddressSpace.gmem, assumed_align=16
    )
    cute_ptr_of_tensor_of_sfasfb_ptrs = make_ptr(
        cutlass.Int64, tensor_of_sfasfb_ptrs.data_ptr(), cute.AddressSpace.gmem, assumed_align=16
    )
    cute_ptr_of_tensor_of_problem_sizes = make_ptr(
        cutlass.Int32, tensor_of_problem_sizes.data_ptr(), cute.AddressSpace.gmem, assumed_align=16
    )
    cute_ptr_of_tensor_of_tensormap = make_ptr(
        cutlass.Int64, tensor_of_tensormap.data_ptr(), cute.AddressSpace.gmem, assumed_align=16
    )

    # Launch
    compiled_func(
        cute_ptr_of_tensor_of_problem_sizes,
        cute_ptr_of_tensor_of_abc_ptrs,
        cute_ptr_of_tensor_of_sfasfb_ptrs,
        cute_ptr_of_tensor_of_tensormap,
        total_num_clusters,
        problem_sizes,
        len(problem_sizes),
    )

    return [abc_tensors[i][2] for i in range(len(problem_sizes))]