submission 498371

- import torch
- from typing import Tuple, List
-  
- def _flatten_reordered(scale: torch.Tensor) -> torch.Tensor:
-     """
-     Convert a scaling tensor that is already in the cuBLAS‑reordered layout
-     ``(32, 4, row_blocks, 4, col_blocks, L)`` into the 1‑D vector expected by
-     ``torch._scaled_mm``.
-  
-     The required order is ``(row_blocks, col_blocks, 32, 4, 4)``; we achieve this
-     with a permutation followed by a contiguous view.
-     """
-     # ``L`` is always 1 in the test suite – drop it if present.
-     if scale.dim() == 6:
-         scale = scale.squeeze(-1)                     # (32,4,Rb,4,Cb)
-  
-     # Permute to bring the row/col block dimensions to the front.
-     # Original: (32, 4, Rb, 4, Cb)  →  (Rb, Cb, 32, 4, 4)
-     return scale.permute(2, 4, 0, 1, 3).contiguous().view(-1)
-  
-  
- def custom_kernel(
-     data: Tuple[
-         List[Tuple[torch.Tensor, torch.Tensor, torch.Tensor]],
-         List[Tuple[torch.Tensor, torch.Tensor]],
-         List[Tuple[torch.Tensor, torch.Tensor]],
-         List[Tuple[int, int, int, int]],
-     ]
- ) -> List[torch.Tensor]:
-     """
-     Grouped NVFP4 block‑scaled GEMM for NVIDIA B200 (NVFP4).
-  
-     For each problem (M, N, K, L) computes
-         C[l] = A[l] @ B[l].T
-     where A and B are packed FP4 tensors (``float4_e2m1fn_x2``) and the
-     per‑block FP8 scaling factors are supplied in the cuBLAS block‑scaled
-     layout (already reordered).  The computation is performed by
-     ``torch._scaled_mm``, which maps to the native B200 FP4 tensor‑core kernel.
-     The result is written back into the provided ``C`` buffer (dtype ``float16``).
-  
-     Parameters
-     ----------
-     data :
-         Tuple containing
-         * ``abc_tensors`` – list of (A, B, C) tensors.
-         * ``sfasfb_tensors`` – unused (original dense scales).
-         * ``sfasfb_reordered_tensors`` – list of (sfa_reordered, sfb_reordered)
-           tensors already in the cuBLAS layout.
-         * ``problem_sizes`` – list of (M, N, K, L) tuples.
-  
-     Returns
-     -------
-     List[torch.Tensor]
-         The output tensors ``C`` (same objects that were passed in).
-     """
-     abc_tensors, _, sfasfb_reordered_tensors, problem_sizes = data
-     results: List[torch.Tensor] = []
-  
-     for (a, b, c), (sfa_reord, sfb_reord), (M, N, K, L) in zip(
-         abc_tensors, sfasfb_reordered_tensors, problem_sizes
-     ):
-         # 1️⃣  Convert the reordered scaling tensors into the flat vectors that
-         #     ``torch._scaled_mm`` expects.  The conversion is tiny compared to
-         #     the GEMM work, so the overhead is negligible.
-         scale_a = _flatten_reordered(sfa_reord).to(a.device)
-         scale_b = _flatten_reordered(sfb_reord).to(b.device)
-  
-         # 2️⃣  Loop over the (trivial) batch dimension L (always 1 in the
-         #     hidden tests, but we keep the loop for completeness).
-         for l_idx in range(L):
-             # A_slice : (M, K/2)
-             # B_slice : (N, K/2)
-             a_slice = a[:, :, l_idx]
-             b_slice = b[:, :, l_idx]
-  
-             # 3️⃣  Core FP4 block‑scaled matrix multiplication.
-             #     ``torch._scaled_mm`` internally dispatches to the B200 FP4
-             #     tensor‑core kernel that applies the per‑block FP8 scaling
-             #     factors.
-             c[:, :, l_idx] = torch._scaled_mm(
-                 a_slice,
-                 b_slice.t(),
-                 scale_a,
-                 scale_b,
-                 bias=None,
-                 out_dtype=torch.float16,
-             )
-  
-         results.append(c)
-  
-     return results
+ 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
+ 
+ 
+ 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  
+ threads_per_cta = 192  
+ epilogue_warp_count = 4
+ mma_warp_id = 0
+ tma_warp_id = 5
+ num_acc_stage = 1
+ num_ab_stage = 2
+ 
+ 
+ def ceil_div(a, b):
+     return (a + b - 1) // b
+ 
+ 
+ def round_tmem_alloc_cols(required_cols: int) -> int:
+     """
+     TMEM allocator accepts power-of-two column counts that are multiples of 32.
+     Valid values are {32, 64, 128, 256, 512}.
+     """
+     valid_cols = (32, 64, 128, 256, 512)
+     need = max(1, int(required_cols))
+     for cols in valid_cols:
+         if need <= cols:
+             return cols
+     return 512
+ 
+ 
+ @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,
+     tensor_of_cta_prefix: cute.Tensor,
+     num_groups: cutlass.Int32,
+     num_tma_load_bytes: cutlass.Constexpr[int],
+ ):
+     """
+     GPU device kernel performing the Group GEMM computation.
+     """
+     warp_idx = cute.arch.warp_idx()
+     warp_idx = cute.arch.make_warp_uniform(warp_idx)
+     tidx, _, _ = cute.arch.thread_idx()
+     is_epilogue_warp = warp_idx < epilogue_warp_count
+     is_mma_warp = warp_idx == mma_warp_id
+     is_tma_warp = warp_idx == tma_warp_id
+ 
+     
+     bidx, bidy, bidz = cute.arch.block_idx()
+     group_idx = cutlass.Int32(0)
+     if num_groups == 2:
+         p1 = tensor_of_cta_prefix[1]
+         if bidz >= p1:
+             group_idx = cutlass.Int32(1)
+     elif num_groups == 8:
+         p1 = tensor_of_cta_prefix[1]
+         p2 = tensor_of_cta_prefix[2]
+         p3 = tensor_of_cta_prefix[3]
+         p4 = tensor_of_cta_prefix[4]
+         p5 = tensor_of_cta_prefix[5]
+         p6 = tensor_of_cta_prefix[6]
+         p7 = tensor_of_cta_prefix[7]
+         if bidz < p4:
+             if bidz < p2:
+                 if bidz < p1:
+                     group_idx = cutlass.Int32(0)
+                 else:
+                     group_idx = cutlass.Int32(1)
+             else:
+                 if bidz < p3:
+                     group_idx = cutlass.Int32(2)
+                 else:
+                     group_idx = cutlass.Int32(3)
+         else:
+             if bidz < p6:
+                 if bidz < p5:
+                     group_idx = cutlass.Int32(4)
+                 else:
+                     group_idx = cutlass.Int32(5)
+             else:
+                 if bidz < p7:
+                     group_idx = cutlass.Int32(6)
+                 else:
+                     group_idx = cutlass.Int32(7)
+     else:
+         left = cutlass.Int32(0)
+         right = num_groups
+         while left < right:
+             mid = (left + right) // 2
+             if tensor_of_cta_prefix[mid + 1] <= bidz:
+                 left = mid + 1
+             else:
+                 right = mid
+         group_idx = left
+     cta_rest = bidz - tensor_of_cta_prefix[group_idx]
+     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]
+     cta_m = ceil_div(m, mma_tiler_mnk[0])
+     coord_y = cta_rest // cta_m
+     coord_x = cta_rest % cta_m
+ 
+     
+     mC_mnl_iter = cute.make_ptr(
+         c_dtype, tensor_of_abc_ptrs[group_idx, 2], cute.AddressSpace.gmem
+     ).align(32)
+ 
+     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)
+     )
+ 
+     
+     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
+     )
+     
+ 
+     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,
+     )
+ 
+     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()
+ 
+     
+     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)
+     )
+     
+     mma_part_slice_idx = tidx
+     if mma_part_slice_idx >= epilogue_warp_count * 32:
+         mma_part_slice_idx = mma_part_slice_idx - epilogue_warp_count * 32
+     thr_mma = tiled_mma.get_slice(mma_part_slice_idx)
+     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)
+ 
+     tensormap_manager = utils.TensorMapManager(
+         utils.TensorMapUpdateMode.GMEM,
+         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),))
+ 
+     sfa_layout = blockscaled_utils.tile_atom_to_shape_SF(
+         mA_mkl_layout.shape, sf_vec_size
+     )
+     sfb_layout = blockscaled_utils.tile_atom_to_shape_SF(
+         mB_nkl_layout.shape, sf_vec_size
+     )
+     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)
+ 
+     if warp_idx == 0:
+         tensormap_manager.init_tensormap_from_atom(
+             tma_atom_a, tensormap_a_gmem_ptr, 0
+         )
+         tensormap_manager.init_tensormap_from_atom(
+             tma_atom_b, tensormap_b_gmem_ptr, 0
+         )
+         tensormap_manager.init_tensormap_from_atom(
+             tma_atom_sfa, tensormap_sfa_gmem_ptr, 0
+         )
+         tensormap_manager.init_tensormap_from_atom(
+             tma_atom_sfb, tensormap_sfb_gmem_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_sfb_gmem_ptr)
+ 
+     cute.arch.barrier()
+ 
+     
+     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)
+ 
+     
+     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)
+ 
+     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)),
+     )
+     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)),
+     )
+ 
+     tCtSFA_fake = cute.make_tensor(
+         cute.make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=16),
+         tCtSFA_layout,
+     )
+     tCtSFB_fake = cute.make_tensor(
+         cute.make_ptr(sf_dtype, 0, cute.AddressSpace.gmem, assumed_align=16),
+         tCtSFB_layout,
+     )
+     acc_cols = tcgen05.find_tmem_tensor_col_offset(tCtAcc_fake)
+     sfa_cols = tcgen05.find_tmem_tensor_col_offset(tCtSFA_fake)
+     sfb_cols = tcgen05.find_tmem_tensor_col_offset(tCtSFB_fake)
+     total_tmem_cols = acc_cols + sfa_cols + sfb_cols
+     alloc_tmem_cols = round_tmem_alloc_cols(total_tmem_cols)
+ 
+     
+     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(alloc_tmem_cols)
+     tmem.wait_for_alloc()
+     acc_tmem_ptr = tmem.retrieve_ptr(cutlass.Float32)
+     tCtAcc = cute.make_tensor(acc_tmem_ptr, tCtAcc_fake.layout)
+ 
+     
+     sfa_tmem_ptr = cute.recast_ptr(
+         acc_tmem_ptr + acc_cols,
+         dtype=sf_dtype,
+     )
+     tCtSFA = cute.make_tensor(sfa_tmem_ptr, tCtSFA_layout)
+     sfb_tmem_ptr = cute.recast_ptr(
+         acc_tmem_ptr
+         + acc_cols
+         + sfa_cols,
+         dtype=sf_dtype,
+     )
+     tCtSFB = cute.make_tensor(sfb_tmem_ptr, tCtSFB_layout)
+ 
+     
+     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])
+ 
+     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])]
+ 
+     if is_tma_warp:
+         for k_tile in range(k_tile_cnt):
+             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,
+                 ),
+             )
+ 
+     if is_mma_warp:
+         acc_empty = acc_producer.acquire_and_advance()
+         tiled_mma.set(tcgen05.Field.ACCUMULATE, False)
+         accumulate_enabled = False
+         num_kblocks = cute.size(tCrA, mode=[2])
+         for k_tile in range(k_tile_cnt):
+             ab_full = ab_consumer.wait_and_advance()
+ 
+             s2t_stage_coord = (None, None, None, None, ab_full.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,
+             )
+ 
+             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,
+                 )
+                 if not accumulate_enabled:
+                     tiled_mma.set(tcgen05.Field.ACCUMULATE, True)
+                     accumulate_enabled = True
+ 
+             ab_full.release()
+         acc_empty.commit()
+ 
+     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])
+     epilogue_slice_idx = tidx
+     if not is_epilogue_warp:
+         epilogue_slice_idx = 0
+     thr_copy_t2r = tiled_copy_t2r.get_slice(epilogue_slice_idx)
+     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()
+ 
+     simt_atom_128 = cute.make_copy_atom(
+         cute.nvgpu.CopyUniversalOp(), c_dtype, num_bits_per_copy=128
+     )
+     simt_atom_fast = cute.make_copy_atom(
+         cute.nvgpu.CopyUniversalOp(), c_dtype, num_bits_per_copy=64
+     )
+     simt_atom = cute.make_copy_atom(
+         cute.nvgpu.CopyUniversalOp(), c_dtype, num_bits_per_copy=16
+     )
+     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]
+     full_m_tile = residue_m >= mma_tiler_mnk[0]
+     full_n_tile = residue_n >= mma_tiler_mnk[1]
+     thread_row = tidx
+     row_valid = thread_row < residue_m
+ 
+     if is_epilogue_warp:
+         cute.copy(tiled_copy_t2r, tDtAcc, tDrAcc)
+         tDrC.store(tDrAcc.load().to(c_dtype))
+ 
+         if full_m_tile and full_n_tile:
+             cute.copy(simt_atom_128, cute.flatten(tDrC), cute.flatten(tDgC))
+         elif full_n_tile:
+             if row_valid:
+                 cute.copy(simt_atom_128, cute.flatten(tDrC), cute.flatten(tDgC))
+         else:
+             if row_valid:
+                 tDpC = cute.make_rmem_tensor(tDrC.shape, cutlass.Boolean)
+                 for i in cutlass.range(cute.size(tDrC.shape), unroll_full=True):
+                     tDpC[i] = i < residue_n
+                 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)
+     pass
+ 
+ 
+ @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_cta_prefix: 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_cta_prefix = cute.make_tensor(
+         ptr_of_tensor_of_cta_prefix, cute.make_layout((num_groups + 1), stride=(1))
+     )
+     tensor_of_tensormap = cute.make_tensor(
+         ptr_of_tensor_of_tensormap, cute.make_layout((total_num_clusters, 4, 16), stride=(64, 16, 1))
+     )
+ 
+     min_a_shape = (cutlass.Int32(64), cutlass.Int32(64), cutlass.Int32(64), cutlass.Int32(1))
+     min_b_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_a_shape[0], cute.assume(min_a_shape[2], 32), min_a_shape[3]),
+             stride=(
+                 cute.assume(min_a_shape[2], 32),
+                 1,
+                 cute.assume(min_a_shape[0] * min_a_shape[2], 32),
+             ),
+         ),
+     )
+     initial_b = cute.make_tensor(
+         cute.make_ptr(ab_dtype, 0, cute.AddressSpace.gmem, assumed_align=16,),
+         cute.make_layout(
+             (min_b_shape[1], cute.assume(min_b_shape[2], 32), min_b_shape[3]),
+             stride=(
+                 cute.assume(min_b_shape[2], 32),
+                 1,
+                 cute.assume(min_b_shape[1] * min_b_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)
+ 
+     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,),
+     )
+ 
+     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)
+ 
+     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,
+     )
+ 
+     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
+ 
+     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,
+         
+         tensor_of_cta_prefix,
+         num_groups,
+ 
+         num_tma_load_bytes,
+     ).launch(
+         grid=grid,
+         block=[threads_per_cta, 1, 1],
+         cluster=(1, 1, 1),
+     )
+     return
+ 
+ 
+ _compiled_kernel_cache = {}
+ _runtime_meta_cache = {}
+ def compile_kernel(problem_sizes):
+     """
+     Compile the kernel once and cache it using problem_sizes as the key.
+     This should be called before any timing measurements.
+ 
+     Returns:
+         The compiled kernel function
+     """
+     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_cta_prefix = make_ptr(
+         cutlass.Int32, 0, cute.AddressSpace.gmem, assumed_align=16,
+     )
+     total_num_clusters = cutlass.Int32(1)
+     num_groups = cutlass.Int32(len(problem_sizes))
+     cute_ptr_of_tensor_of_tensormap = make_ptr(
+         cutlass.Int64, 0, cute.AddressSpace.gmem, assumed_align=16,
+     )
+     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_cta_prefix,
+         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:
+     """
+     Execute the block-scaled group GEMM kernel.
+     
+     This is the main entry point called by the evaluation framework.
+     It converts PyTorch tensors to CuTe tensors, launches the kernel,
+     and returns the result.
+     
+     Args:
+         data: Tuple of (abc_tensors, sfasfb_tensors, problem_sizes) where:
+             abc_tensors: list of tuples (a, b, c) where 
+                 a is torch.Tensor[float4e2m1fn_x2] of shape [m, k // 2, l]
+                 b is torch.Tensor[float4e2m1fn_x2] of shape [n, k // 2, l]
+                 c is torch.Tensor[float16] of shape [m, n, l]
+             sfasfb_tensors: list of tuples (sfa, sfb) where 
+                 sfa is torch.Tensor[float8_e4m3fnuz] of shape [m, k // 16, l]
+                 sfb is torch.Tensor[float8_e4m3fnuz] of shape [n, k // 16, l]
+             problem_sizes: list of tuples (m, n, k, l)
+             each group has its own a, b, c, sfa, sfb with different m, n, k, l problem sizes
+             l should always be 1 for each group.
+             list size is the number of groups.
+     
+     Returns:
+         list of c tensors where c is torch.Tensor[float16] of shape [m, n, l] for each group
+     """
+     abc_tensors, _, sfasfb_reordered_tensors, problem_sizes = data
+ 
+     global _runtime_meta_cache
+     compiled_func = compile_kernel(problem_sizes)
+ 
+     runtime_key = tuple(tuple(int(x) for x in mnkl) for mnkl in problem_sizes)
+     runtime_meta = _runtime_meta_cache.get(runtime_key)
+     if runtime_meta is None:
+         tensor_of_problem_sizes = torch.tensor(
+             problem_sizes, dtype=torch.int32, device="cuda"
+         )
+ 
+         cta_tile_shape_mn = [mma_tiler_mnk[0], 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
+         cta_prefix = [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)
+             )
+             group_clusters = functools.reduce(lambda x, y: x * y, num_clusters_mn)
+             total_num_clusters += group_clusters
+             cta_prefix.append(total_num_clusters)
+         tensor_of_cta_prefix = torch.tensor(cta_prefix, dtype=torch.int32, device="cuda")
+ 
+         tensormap_shape = (
+             total_num_clusters,
+             num_tensormaps,
+             bytes_per_tensormap // 8,
+         )
+         tensor_of_tensormap = torch.empty(tensormap_shape, dtype=torch.int64, device="cuda")
+         num_groups_local = len(problem_sizes)
+         tensor_of_abc_ptrs = torch.empty((num_groups_local, 3), dtype=torch.int64, device="cuda")
+         tensor_of_sfasfb_ptrs = torch.empty((num_groups_local, 2), dtype=torch.int64, device="cuda")
+         host_abc_ptrs = torch.empty((num_groups_local, 3), dtype=torch.int64, pin_memory=True)
+         host_sfasfb_ptrs = torch.empty((num_groups_local, 2), dtype=torch.int64, pin_memory=True)
+         runtime_meta = {
+             "tensor_of_problem_sizes": tensor_of_problem_sizes,
+             "tensor_of_cta_prefix": tensor_of_cta_prefix,
+             "tensor_of_tensormap": tensor_of_tensormap,
+             "tensor_of_abc_ptrs": tensor_of_abc_ptrs,
+             "tensor_of_sfasfb_ptrs": tensor_of_sfasfb_ptrs,
+             "host_abc_ptrs": host_abc_ptrs,
+             "host_sfasfb_ptrs": host_sfasfb_ptrs,
+             "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,
+             ),
+             "total_num_clusters": total_num_clusters,
+             "num_groups": len(problem_sizes),
+             "last_abc_ptrs": [[0, 0, 0] for _ in range(num_groups_local)],
+             "last_sfasfb_ptrs": [[0, 0] for _ in range(num_groups_local)],
+         }
+         _runtime_meta_cache[runtime_key] = runtime_meta
+     else:
+         tensor_of_problem_sizes = runtime_meta["tensor_of_problem_sizes"]
+         tensor_of_cta_prefix = runtime_meta["tensor_of_cta_prefix"]
+         tensor_of_tensormap = runtime_meta["tensor_of_tensormap"]
+         tensor_of_abc_ptrs = runtime_meta["tensor_of_abc_ptrs"]
+         tensor_of_sfasfb_ptrs = runtime_meta["tensor_of_sfasfb_ptrs"]
+         host_abc_ptrs = runtime_meta["host_abc_ptrs"]
+         host_sfasfb_ptrs = runtime_meta["host_sfasfb_ptrs"]
+ 
+     total_num_clusters = runtime_meta["total_num_clusters"]
+     num_groups = runtime_meta["num_groups"]
+ 
+     last_abc_ptrs = runtime_meta["last_abc_ptrs"]
+     last_sfasfb_ptrs = runtime_meta["last_sfasfb_ptrs"]
+     ptrs_changed = False
+     for i, ((a, b, c), (sfa_reordered, sfb_reordered), _) in enumerate(
+         zip(abc_tensors, sfasfb_reordered_tensors, problem_sizes)
+     ):
+         a_ptr = a.data_ptr()
+         b_ptr = b.data_ptr()
+         c_ptr = c.data_ptr()
+         sfa_ptr = sfa_reordered.data_ptr()
+         sfb_ptr = sfb_reordered.data_ptr()
+         if (
+             last_abc_ptrs[i][0] != a_ptr
+             or last_abc_ptrs[i][1] != b_ptr
+             or last_abc_ptrs[i][2] != c_ptr
+             or last_sfasfb_ptrs[i][0] != sfa_ptr
+             or last_sfasfb_ptrs[i][1] != sfb_ptr
+         ):
+             ptrs_changed = True
+             last_abc_ptrs[i][0] = a_ptr
+             last_abc_ptrs[i][1] = b_ptr
+             last_abc_ptrs[i][2] = c_ptr
+             last_sfasfb_ptrs[i][0] = sfa_ptr
+             last_sfasfb_ptrs[i][1] = sfb_ptr
+ 
+     if ptrs_changed:
+         for i in range(num_groups):
+             host_abc_ptrs[i, 0] = last_abc_ptrs[i][0]
+             host_abc_ptrs[i, 1] = last_abc_ptrs[i][1]
+             host_abc_ptrs[i, 2] = last_abc_ptrs[i][2]
+             host_sfasfb_ptrs[i, 0] = last_sfasfb_ptrs[i][0]
+             host_sfasfb_ptrs[i, 1] = last_sfasfb_ptrs[i][1]
+         tensor_of_abc_ptrs.copy_(host_abc_ptrs, non_blocking=True)
+         tensor_of_sfasfb_ptrs.copy_(host_sfasfb_ptrs, non_blocking=True)
+ 
+     cute_ptr_of_tensor_of_abc_ptrs = runtime_meta["cute_ptr_of_tensor_of_abc_ptrs"]
+     cute_ptr_of_tensor_of_sfasfb_ptrs = runtime_meta["cute_ptr_of_tensor_of_sfasfb_ptrs"]
+     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_cta_prefix = make_ptr(
+         cutlass.Int32,
+         tensor_of_cta_prefix.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,
+     )
+ 
+     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_cta_prefix,
+         cute_ptr_of_tensor_of_tensormap,
+         total_num_clusters,
+         problem_sizes,
+         num_groups,
+     )
+ 
+     res = []
+     for i in range(num_groups):
+         res.append(abc_tensors[i][2])
+     return res