quack-kernels 0.1.10__py3-none-any.whl → 0.2.0__py3-none-any.whl

quack/dense_gemm_sm90.py

@@ -1,47 +1,43 @@
-# Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: BSD-3-Clause
-
-# Redistribution and use in source and binary forms, with or without
-# modification, are permitted provided that the following conditions are met:
-
-# 1. Redistributions of source code must retain the above copyright notice, this
-# list of conditions and the following disclaimer.
-
-# 2. Redistributions in binary form must reproduce the above copyright notice,
-# this list of conditions and the following disclaimer in the documentation
-# and/or other materials provided with the distribution.
-
-# 3. Neither the name of the copyright holder nor the names of its
-# contributors may be used to endorse or promote products derived from
-# this software without specific prior written permission.
-
-# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
-# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
-# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
-# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
-# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
-# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
-# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
-# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
-# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
-# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-import argparse
-from typing import Tuple, Type
+# Based on the cute-dsl example:
+# https://github.com/NVIDIA/cutlass/blob/main/examples/python/CuTeDSL/hopper/dense_gemm.py
+
+import enum
+from typing import Tuple, Type, Callable, Optional, Union
+from dataclasses import dataclass
+from functools import partial
 import math
-import cuda.bindings.driver as cuda
 
-import torch
+from torch import Tensor
+
+import cuda.bindings.driver as cuda
 
 import cutlass
 import cutlass.cute as cute
-import cutlass.cute.testing as testing
-import cutlass.utils as utils
 import cutlass.pipeline as pipeline
-import cutlass.torch as cutlass_torch
-from cutlass.cute.runtime import from_dlpack
 from cutlass.cute.nvgpu import cpasync, warp, warpgroup
 import cutlass.utils.hopper_helpers as sm90_utils
+from cutlass import Int32, Float32, Boolean, const_expr
+from cutlass.utils import LayoutEnum
+import cutlass.torch as cutlass_torch
+from cutlass.cute.runtime import make_ptr
+
+
+from quack.cute_dsl_utils import ParamsBase, ArgumentsBase
+from quack.tile_scheduler import (
+    TileSchedulerOptions,
+    TileSchedulerArguments,
+    TileScheduler,
+    VarlenMTileSchedulerArguments,
+    VarlenMTileScheduler,
+)
+from quack.varlen_utils import VarlenArguments
+from quack.tensormap_manager import TensorMapManagerSm90
+
+# return PipelineStateWAdvance instead of PipelineState
+from quack.pipeline import make_pipeline_state, PipelineTmaCpAsync
+import quack.utils as utils
+from quack.cute_dsl_utils import get_max_active_clusters
+from quack.gemm_wrapper_utils import GemmWrapperBase
 
 """
 A high-performance batched dense GEMM (C = A * B) example for the NVIDIA Hopper architecture
@@ -66,31 +62,6 @@ Hopper WGMMA instructions operate as follows:
 - Read matrix B from SMEM
 - Perform MMA operation and store the result in Accumulator(register)
 
-To run this example:
-
-.. code-block:: bash
-
-    python examples/hopper/dense_gemm.py                                   \
-      --mnkl 8192,8192,8192,1 --tile_shape_mnk 128,256,64                  \
-      --cluster_shape_mn 1,1 --a_dtype Float16 --b_dtype Float16           \
-      --d_dtype Float16 --acc_dtype Float32                                \
-      --a_major k --b_major k --d_major n
-
-The above example command compute batched gemm with M=8192, N=8192, K=8192,
-batch_count=1. The Hopper WGMMA tile shape is 128x256x64 and the cluster shape
-is (1,1). The input, mma accumulator and output data type are set as fp16, fp32
-and fp16, respectively.
-
-To collect performance with NCU profiler:
-
-.. code-block:: bash
-
-    ncu python examples/hopper/dense_gemm.py                               \
-      --mnkl 8192,8192,8192,1 --tile_shape_mnk 128,256,64                  \
-      --cluster_shape_mn 1,1 --a_dtype Float16 --b_dtype Float16           \
-      --d_dtype Float16 --acc_dtype Float32                                \
-      --a_major k --b_major k --d_major n
-
 Constraints:
 * Supported input data types: fp16, fp8 (e4m3fn, e5m2)
 * For fp16 types, A and B must have the same data type
@@ -103,107 +74,29 @@ Constraints:
 * Cluster shape M/N must be positive and power of 2, total cluster size <= 4
 * The contiguous dimension of A/B/C tensors must be at least 16 bytes aligned,
   i.e, number of elements is a multiple of 8, 16 for Float16, and Float8, respectively.
-* OOB tiles are not allowed when TMA store is disabled
 """
 
 
-# /////////////////////////////////////////////////////////////////////////////
-#  Helpers to parse args
-# /////////////////////////////////////////////////////////////////////////////
-def parse_comma_separated_ints(s: str):
-    try:
-        return tuple([int(x.strip()) for x in s.split(",")])
-    except ValueError:
-        raise argparse.ArgumentTypeError("Invalid format. Expected comma-separated integers.")
-
-
-def parse_arguments() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Example of MxNxKxL GEMM on Hopper.")
-
-    parser.add_argument(
-        "--mnkl",
-        type=parse_comma_separated_ints,
-        default=(4096, 4096, 4096, 1),
-        help="mnkl dimensions (comma-separated)",
-    )
-    parser.add_argument(
-        "--tile_shape_mnk",
-        type=parse_comma_separated_ints,
-        default=(128, 256, 64),
-        help="Cta tile shape (comma-separated)",
-    )
-    parser.add_argument(
-        "--cluster_shape_mn",
-        type=parse_comma_separated_ints,
-        choices=[(1, 1), (2, 1), (1, 2), (2, 2)],
-        default=(1, 1),
-        help="Cluster shape (comma-separated)",
-    )
-    parser.add_argument(
-        "--a_dtype",
-        type=cutlass.dtype,
-        default=cutlass.BFloat16,
-    )
-    parser.add_argument(
-        "--b_dtype",
-        type=cutlass.dtype,
-        default=cutlass.BFloat16,
-    )
-    parser.add_argument(
-        "--d_dtype",
-        type=cutlass.dtype,
-        default=cutlass.BFloat16,
-    )
-    parser.add_argument(
-        "--acc_dtype",
-        type=cutlass.dtype,
-        default=cutlass.Float32,
-    )
-    parser.add_argument("--a_major", choices=["k", "m"], type=str, default="k")
-    parser.add_argument("--b_major", choices=["k", "n"], type=str, default="k")
-    parser.add_argument("--d_major", choices=["n", "m"], type=str, default="n")
-    parser.add_argument("--tolerance", type=float, default=1e-01, help="Tolerance for validation")
-    parser.add_argument("--warmup_iterations", type=int, default=0, help="Warmup iterations")
-    parser.add_argument(
-        "--iterations",
-        type=int,
-        default=1,
-        help="Number of iterations to run the kernel",
-    )
-    parser.add_argument("--skip_ref_check", action="store_true", help="Skip reference checking")
-    parser.add_argument(
-        "--use_cold_l2",
-        action="store_true",
-        default=False,
-        help="Use circular buffer tensor sets to ensure L2 cold cache",
-    )
-
-    args = parser.parse_args()
-
-    if len(args.mnkl) != 4:
-        parser.error("--mnkl must contain exactly 4 values")
-    if len(args.tile_shape_mnk) != 3:
-        parser.error("--tile_shape_mnk must contain exactly 3 values")
-    if len(args.cluster_shape_mn) != 2:
-        parser.error("--cluster_shape_mn must contain exactly 2 values")
-
-    return args
-
-
-# /////////////////////////////////////////////////////////////////////////////
-#  Host setup and device kernel launch
-# /////////////////////////////////////////////////////////////////////////////
+class NamedBarrierGemm(enum.IntEnum):
+    Epilogue = enum.auto()  # starts from 1 as barrier 0 is reserved for sync_threads()
+    # For mainloop load warps to signal that the epilogue load warp can start.
+    # This is to avoid loading C too early, interfering with loading A and B.
+    EpilogueLoad = enum.auto()
+    MmaWG0 = enum.auto()
+    MmaWG1 = enum.auto()
+    EpiWG0 = enum.auto()
+    EpiWG1 = enum.auto()
 
 
-class HopperWgmmaGemmKernel:
+class GemmSm90:
     """
     This class implements batched matrix multiplication (C = A x B) with support for various data types
     and architectural features specific to Hopper GPUs.
 
     :param acc_dtype: Data type for accumulation during computation
     :type acc_dtype: type[cutlass.Numeric]
-    :param tile_shape_mnk: Shape of the CTA tile (M,N,K)
-    :type tile_shape_mnk: Tuple[int, int, int]
+    :param tile_shape_mn: Shape of the CTA tile (M,N)
+    :type tile_shape_mn: Tuple[int, int, int]
     :param cluster_shape_mnk: Cluster dimensions (M,N,K) for parallel processing
     :type cluster_shape_mnk: Tuple[int, int, int]
 
@@ -221,25 +114,39 @@ class HopperWgmmaGemmKernel:
         - Float32 (for all floating point inputs)
 
     :note: Constraints:
-        - CTA tile M must be 64/128
-        - CTA tile N must be 64/128/256
-        - CTA tile K must be 64
         - Cluster shape M/N must be positive and power of 2, total cluster size <= 4
 
     Example:
-        >>> gemm = HopperWgmmaGemmKernel(
+        >>> gemm = GemmSm90(
         ...     acc_dtype=cutlass.Float32,
-        ...     tile_shape_mnk=(128, 256, 64),
+        ...     tile_shape_mn=(128, 256),
         ...     cluster_shape_mnk=(1, 1, 1)
         ... )
         >>> gemm(a_tensor, b_tensor, c_tensor, stream)
     """
 
+    bytes_per_tensormap = 128
+
+    @dataclass
+    class EpilogueArguments(ArgumentsBase):
+        alpha: Optional[Float32 | cute.Tensor] = None
+        beta: Optional[Float32 | cute.Tensor] = None
+
+    @dataclass
+    class EpilogueParams(ParamsBase):
+        alpha: Optional[Float32 | cute.Tensor] = None
+        beta: Optional[Float32 | cute.Tensor] = None
+
     def __init__(
         self,
         acc_dtype: Type[cutlass.Numeric],
-        tile_shape_mnk: Tuple[int, int, int],
+        a_dtype: Type[cutlass.Numeric],
+        tile_shape_mn: Tuple[int, int],
         cluster_shape_mnk: Tuple[int, int, int],
+        pingpong: bool = False,
+        is_persistent: bool = True,
+        fp8_fast_accum: bool = False,
+        gather_A: bool = False,
     ):
         """
         Initializes the configuration for a Hopper dense GEMM kernel.
@@ -249,59 +156,106 @@ class HopperWgmmaGemmKernel:
 
         :param acc_dtype: Data type for accumulation during computation
         :type acc_dtype: type[cutlass.Numeric]
-        :param tile_shape_mnk: Shape of the CTA tile (M,N,K)
-        :type tile_shape_mnk: Tuple[int, int, int]
+        :param tile_shape_mn: Shape of the CTA tile (M,N)
+        :type tile_shape_mn: Tuple[int, int]
         :param cluster_shape_mnk: Cluster dimensions (M,N,K) for parallel processing
         :type cluster_shape_mnk: Tuple[int, int, int]
         """
 
         self.acc_dtype = acc_dtype
+        self.pingpong = pingpong
+        self.is_persistent = is_persistent
+        if self.pingpong:
+            assert self.is_persistent, "Pingpong gemm requires persistent scheduler"
+        self.fp8_slow_accum = not fp8_fast_accum and a_dtype.width == 8
+        self.gather_A = gather_A
+        if gather_A:
+            assert cluster_shape_mnk[1] == 1, "Cluster shape N must be 1 for gather A "
 
         self.cluster_shape_mnk = cluster_shape_mnk
-        self.tile_shape_mnk = tuple(tile_shape_mnk)
-        tile_M, tile_N = tile_shape_mnk[0], tile_shape_mnk[1]
+        # K dimension is deferred in _setup_attributes
+        self.tile_shape_mnk = (*tile_shape_mn, 1)
+        tile_M, tile_N = self.tile_shape_mnk[0], self.tile_shape_mnk[1]
         # check the cta tile shape
-        # if tile_M not in [64, 128, 192, 256]:
-        # TODO: M=192 currently doesn't work
-        if tile_M not in [64, 128, 256]:
-            raise ValueError("CTA tile shape M must be 64/128/192/256")
-        if tile_M == 192:  # special case
-            if not (tile_N % 32 == 0 and tile_N <= 288):
-                raise ValueError(
-                    "If tile_m == 192, CTA tile shape N must be divisible by 32 and <= 288"
-                )
+        if not self.pingpong:
+            if tile_M not in [64, 128, 192, 256, 320]:
+                raise ValueError("CTA tile shape M must be 64/128/192/256/320")
+            if tile_M in [192, 320]:  # special case
+                tile_N_max = 256 if tile_M == 192 else 160
+                if not (tile_N % 32 == 0 and tile_N <= tile_N_max):
+                    raise ValueError(
+                        f"If tile_m == {tile_M}, CTA tile shape N must be divisible by 32 and <= {tile_N_max}"
+                    )
+            else:
+                if not (
+                    (tile_N % 16 == 0 and tile_N <= 256) or (tile_N % 32 == 0 and tile_N <= 512)
+                ):
+                    raise ValueError(
+                        "CTA tile shape N must be divisible by 16 and <= 256, or divisible by 32 and <= 512"
+                    )
         else:
-            if not ((tile_N % 16 == 0 and tile_N <= 256) or (tile_N % 32 == 0 and tile_N <= 512)):
-                raise ValueError(
-                    "CTA tile shape N must be divisible by 16 and <= 256, or divisible by 32 and <= 512"
-                )
-        if not self.tile_shape_mnk[2] % 16 == 0:
-            raise ValueError("CTA tile shape K must be divisible by 16")
-
-        if tile_M == 192:  # Special case
-            atom_layout_m, atom_layout_n = 1, 2
+            if tile_M not in [64, 128, 192]:
+                raise ValueError("CTA tile shape M must be 64/128/192 if pingpong")
+            tile_N_max = 256 if tile_M == 64 else (208 if tile_M == 128 else 128)
+            if not (tile_N % 16 == 0 and tile_N <= tile_N_max):
+                raise ValueError(f"CTA tile shape N must be divisible by 16 and <= {tile_N_max}")
+
+        if not self.pingpong:
+            if tile_M == 320:  # tile_M / 64 is not even so we have to split along N
+                atom_layout_m, atom_layout_n = 1, 2
+            elif tile_M == 192:
+                if tile_N <= 128:
+                    atom_layout_m, atom_layout_n = 3, 1
+                else:
+                    atom_layout_m, atom_layout_n = 1, 2
+            else:
+                atom_layout_m = self.tile_shape_mnk[0] // 64 if self.tile_shape_mnk[0] < 256 else 2
+                atom_layout_n = 1
+            assert atom_layout_m in [1, 2, 3] and atom_layout_n in [1, 2]
         else:
-            atom_layout_m = tile_shape_mnk[0] // 64 if tile_shape_mnk[0] < 256 else 2
-            atom_layout_n = 1
-        assert atom_layout_m in [1, 2] and atom_layout_n in [1, 2]
+            atom_layout_m, atom_layout_n = 1, 1
         self.atom_layout_mnk = (atom_layout_m, atom_layout_n, 1)
 
-        self.num_mcast_ctas_a = self.cluster_shape_mnk[1]
+        self.num_mcast_ctas_a = self.cluster_shape_mnk[1] if not self.gather_A else 1
         self.num_mcast_ctas_b = self.cluster_shape_mnk[0]
         self.is_a_mcast = self.num_mcast_ctas_a > 1
         self.is_b_mcast = self.num_mcast_ctas_b > 1
 
         self.occupancy = 1
-        self.mma_warp_groups = math.prod(self.atom_layout_mnk)
+        self.mma_warp_groups = math.prod(self.atom_layout_mnk) * (1 if not self.pingpong else 2)
+        if self.pingpong:
+            assert self.mma_warp_groups == 2
+        assert self.mma_warp_groups in [1, 2, 3]
         self.num_threads_per_warp_group = 128
         self.threads_per_cta = (self.mma_warp_groups + 1) * self.num_threads_per_warp_group
-        self.smem_capacity = utils.get_smem_capacity_in_bytes("sm_90")
-        self.num_mma_threads = self.mma_warp_groups * self.num_threads_per_warp_group
-
-        regs_per_thread = math.prod(self.tile_shape_mnk) // self.num_mma_threads
-        heavy_register_pressure = regs_per_thread >= 208
-        self.num_regs_load = 40 if not heavy_register_pressure else 24
-        self.num_regs_mma = 232 if not heavy_register_pressure else 240
+        self.smem_capacity = cutlass.utils.get_smem_capacity_in_bytes("sm_90")
+        self.num_epi_threads = (
+            self.mma_warp_groups if not self.pingpong else 1
+        ) * self.num_threads_per_warp_group
+        self.num_ab_load_warps = 1 if not self.gather_A else 4
+        self.num_ab_load_threads = cute.arch.WARP_SIZE * self.num_ab_load_warps
+        self.num_epi_load_threads = cute.arch.WARP_SIZE * 1
+        self.ab_load_warp_id = self.mma_warp_groups * 4
+        self.epi_load_warp_id = self.ab_load_warp_id + self.num_ab_load_warps
+
+        regs_per_thread = math.prod(self.tile_shape_mnk[:2]) // (
+            math.prod(self.atom_layout_mnk) * self.num_threads_per_warp_group
+        )
+        if self.fp8_slow_accum:
+            regs_per_thread *= 2
+        if not self.gather_A:
+            if self.mma_warp_groups == 3:
+                self.num_regs_load, self.num_regs_mma = 32, 160
+            else:
+                heavy_register_pressure = regs_per_thread >= 208
+                self.num_regs_load, self.num_regs_mma = (
+                    (40, 232) if not heavy_register_pressure else (24, 240)
+                )
+        else:
+            if self.mma_warp_groups == 3:
+                self.num_regs_load, self.num_regs_mma = 56, 152
+            else:
+                self.num_regs_load, self.num_regs_mma = (56, 224)
 
         self.ab_stage = None
         self.epi_stage = None
@@ -314,7 +268,7 @@ class HopperWgmmaGemmKernel:
         self.shared_storage = None
         self.buffer_align_bytes = 1024
 
-    def _setup_attributes(self):
+    def _setup_attributes(self, epilogue_args: Optional[EpilogueArguments]):
         """Set up configurations that are dependent on GEMM inputs
 
         This method configures various attributes based on the input tensor properties
@@ -328,26 +282,67 @@ class HopperWgmmaGemmKernel:
         - Computing A/B/C shared memory layout
         """
 
-        self.cta_layout_mnk = cute.make_layout(self.cluster_shape_mnk)
+        self.tiled_mma = sm90_utils.make_trivial_tiled_mma(
+            self.a_dtype,
+            self.b_dtype,
+            self.a_layout.sm90_mma_major_mode(),
+            self.b_layout.sm90_mma_major_mode(),
+            self.acc_dtype,
+            self.atom_layout_mnk,
+            tiler_mn=(64, self.tile_shape_mnk[1] // self.atom_layout_mnk[1]),
+        )
+        if const_expr(self.atom_layout_mnk[1] > 1):
+            # If N dimension is split among 2 WGs, we need to permute the N dimension so
+            # that in the epilogue, WG0 and WG1 can write to epi smem of size e.g. (64, 32)
+            # containing accumulators that are next to each other in the N dimension.
+            # Without permutation WG0 would write to epi smem of size (64, 16) and
+            # WG1 would write to a separate epi smem of size (64, 16) that's far away.
+            atom_n = self.atom_layout_mnk[1]
+            permutation_n = cute.make_ordered_layout(
+                (8, self.tile_shape_mnk[1] // atom_n // 8, atom_n), order=(0, 2, 1)
+            )
+            self.tiled_mma = cute.make_tiled_mma(
+                cute.make_mma_atom(self.tiled_mma.op),
+                self.atom_layout_mnk,
+                permutation_mnk=(None, permutation_n, None),
+            )
+        mma_inst_shape_k = cute.size(self.tiled_mma.shape_mnk, mode=[2])
+        mma_inst_tile_k = 4
+        self.tile_shape_mnk = (
+            self.tile_shape_mnk[0],
+            self.tile_shape_mnk[1],
+            mma_inst_shape_k * mma_inst_tile_k,
+        )
+
+        self.cluster_layout_mnk = cute.make_layout(self.cluster_shape_mnk)
 
-        is_cooperative = math.prod(self.atom_layout_mnk) > 1
         self.epi_tile = self._sm90_compute_tile_shape_or_override(
-            self.tile_shape_mnk, self.d_dtype, is_cooperative=is_cooperative
+            self.tile_shape_mnk,
+            self.atom_layout_mnk,
+            self.d_dtype,
         )
 
         # Compute stage before compute smem layout
-        self.ab_stage, self.epi_stage = self._compute_stages(
+        self.ab_stage, self.epi_stage, self.epi_c_stage = self._compute_stages(
             self.tile_shape_mnk,
+            self.epi_tile,
             self.a_dtype,
             self.b_dtype,
+            self.d_dtype,
+            self.c_dtype,
+            epilogue_args,
             self.smem_capacity,
             self.occupancy,
+            # epi_smem will reuse smem ab if not persistent.
+            overlap_sD_sA=not self.is_persistent,
         )
+        self.sched_stage = 2 if self.pingpong else 1
 
         (
             self.a_smem_layout_staged,
             self.b_smem_layout_staged,
             self.epi_smem_layout_staged,
+            self.epi_c_smem_layout_staged,
         ) = self._make_smem_layouts(
             self.tile_shape_mnk,
             self.epi_tile,
@@ -359,6 +354,9 @@ class HopperWgmmaGemmKernel:
             self.d_dtype,
             self.d_layout,
             self.epi_stage,
+            self.c_dtype,
+            self.c_layout,
+            self.epi_c_stage,
         )
 
     @cute.jit
@@ -366,7 +364,12 @@ class HopperWgmmaGemmKernel:
         self,
         mA: cute.Tensor,
         mB: cute.Tensor,
-        mD: cute.Tensor,
+        mD: Optional[cute.Tensor],
+        mC: Optional[cute.Tensor],
+        epilogue_args: Optional[ArgumentsBase],
+        scheduler_args: TileSchedulerOptions,
+        varlen_args: Optional[VarlenArguments],
+        mAIdx: Optional[cute.Tensor],
         stream: cuda.CUstream,
     ):
         """Execute the GEMM operation in steps:
@@ -389,36 +392,44 @@ class HopperWgmmaGemmKernel:
         # setup static attributes before smem/grid/tma computation
         self.a_dtype = mA.element_type
         self.b_dtype = mB.element_type
-        self.d_dtype = mD.element_type
-        self.a_layout = utils.LayoutEnum.from_tensor(mA)
-        self.b_layout = utils.LayoutEnum.from_tensor(mB)
-        self.d_layout = utils.LayoutEnum.from_tensor(mD)
-
-        if cutlass.const_expr(self.a_dtype.width == 16 and self.a_dtype != self.b_dtype):
+        self.d_dtype = mD.element_type if mD is not None else None
+        self.c_dtype = mC.element_type if mC is not None else None
+        self.a_layout = LayoutEnum.from_tensor(mA)
+        self.b_layout = LayoutEnum.from_tensor(mB)
+        self.d_layout = LayoutEnum.from_tensor(mD) if mD is not None else None
+        self.c_layout = LayoutEnum.from_tensor(mC) if mC is not None else None
+
+        if const_expr(self.a_dtype.width == 16 and self.a_dtype != self.b_dtype):
             raise TypeError(f"Type mismatch: {self.a_dtype} != {self.b_dtype}")
-        if cutlass.const_expr(self.a_dtype.width != self.b_dtype.width):
+        if const_expr(self.a_dtype.width != self.b_dtype.width):
             raise TypeError(f"Type width mismatch: {self.a_dtype.width} != {self.b_dtype.width}")
-        if cutlass.const_expr(self.a_dtype.width != 16 and self.a_dtype.width != 8):
+        if const_expr(self.a_dtype.width != 16 and self.a_dtype.width != 8):
             raise TypeError("a_dtype should be float16 or float8")
+        assert (mAIdx is not None) == self.gather_A
 
-        self._setup_attributes()
-
-        tiled_mma = sm90_utils.make_trivial_tiled_mma(
-            self.a_dtype,
-            self.b_dtype,
-            self.a_layout.sm90_mma_major_mode(),
-            self.b_layout.sm90_mma_major_mode(),
-            self.acc_dtype,
-            self.atom_layout_mnk,
-            tiler_mn=(64, self.tile_shape_mnk[1] // self.atom_layout_mnk[1]),
-        )
-
-        tma_atom_a, tma_tensor_a = self._make_tma_atoms_and_tensors(
-            mA,
-            self.a_smem_layout_staged,
-            (self.tile_shape_mnk[0], self.tile_shape_mnk[2]),
-            self.cluster_shape_mnk[1],
+        # Assume all strides are divisible by 128 bits except the last stride
+        new_stride = lambda t: tuple(
+            cute.assume(s, divby=128 // t.element_type.width) if not cute.is_static(s) else s
+            for s in t.stride
         )
+        mA, mD = [
+            cute.make_tensor(t.iterator, cute.make_layout(t.shape, stride=new_stride(t)))
+            if t is not None
+            else None
+            for t in (mA, mD)
+        ]
+
+        self._setup_attributes(epilogue_args)
+
+        if const_expr(not self.gather_A):
+            tma_atom_a, tma_tensor_a = self._make_tma_atoms_and_tensors(
+                mA,
+                self.a_smem_layout_staged,
+                (self.tile_shape_mnk[0], self.tile_shape_mnk[2]),
+                self.cluster_shape_mnk[1],
+            )
+        else:
+            tma_atom_a, tma_tensor_a = None, None
 
         tma_atom_b, tma_tensor_b = self._make_tma_atoms_and_tensors(
             mB,
@@ -427,17 +438,89 @@ class HopperWgmmaGemmKernel:
             self.cluster_shape_mnk[0],
         )
 
-        tma_atom_d, tma_tensor_d = self._make_tma_store_atoms_and_tensors(
-            mD,
-            self.epi_smem_layout_staged,
-            self.epi_tile,
+        if const_expr(mD is not None):
+            tma_atom_d, tma_tensor_d = self._make_tma_epi_atoms_and_tensors(
+                mD, self.epi_smem_layout_staged, self.epi_tile, store_or_load="store"
+            )
+        else:
+            tma_atom_d, tma_tensor_d = None, None
+
+        if const_expr(mC is not None):
+            tma_atom_c, tma_tensor_c = self._make_tma_epi_atoms_and_tensors(
+                mC, self.epi_c_smem_layout_staged, self.epi_tile, store_or_load="load"
+            )
+        else:
+            tma_atom_c, tma_tensor_c = None, None
+
+        epilogue_params = self.epi_to_underlying_arguments(epilogue_args)
+
+        if const_expr(varlen_args is None):
+            varlen_args = VarlenArguments()
+        if const_expr(varlen_args.mCuSeqlensM is None):
+            num_problems = (
+                mD.shape[2]
+                if mD is not None
+                else (
+                    mB.shape[2]
+                    if varlen_args.mCuSeqlensK is None
+                    else varlen_args.mCuSeqlensK.shape[0] - 1
+                )
+            )
+            problem_shape_ntile_mnl = (
+                cute.ceil_div(mA.shape[0], self.tile_shape_mnk[0]),
+                cute.ceil_div(mB.shape[0], self.tile_shape_mnk[1]),
+                num_problems,
+            )
+            TileSchedulerCls = self.get_scheduler_class()
+            tile_sched_args = self.get_scheduler_arguments(problem_shape_ntile_mnl, scheduler_args)
+        else:
+            assert mD is not None or not self.gather_A
+            problem_shape_ntile_mnl = (
+                None,
+                cute.ceil_div(mB.shape[0], self.tile_shape_mnk[1]),
+                varlen_args.mCuSeqlensM.shape[0] - 1,
+            )
+            TileSchedulerCls = VarlenMTileScheduler
+            tile_sched_args = VarlenMTileSchedulerArguments(
+                problem_shape_ntile_mnl=problem_shape_ntile_mnl,
+                total_m=mD.shape[0] if mD is not None else mAIdx.shape[0],
+                cu_seqlens_m=varlen_args.mCuSeqlensM,
+                raster_order=scheduler_args.raster_order,
+                group_size=scheduler_args.max_swizzle_size,
+                tile_shape_mn=self.tile_shape_mnk[:2],
+                cluster_shape_mnk=self.cluster_shape_mnk,
+                tile_count_semaphore=scheduler_args.tile_count_semaphore,
+                is_persistent=self.is_persistent,
+            )
+        tile_sched_params = TileSchedulerCls.to_underlying_arguments(tile_sched_args)
+        grid = TileSchedulerCls.get_grid_shape(
+            tile_sched_params, scheduler_args.max_active_clusters
         )
 
-        grid = self._compute_grid(mD, self.tile_shape_mnk, self.cluster_shape_mnk)
+        epi_smem_size = (
+            cute.cosize(self.epi_smem_layout_staged) if self.is_persistent and mD is not None else 0
+        )
+        epi_c_smem_size = cute.cosize(self.epi_c_smem_layout_staged) if mC is not None else 0
 
         @cute.struct
         class SharedStorage:
-            mainloop_pipeline_array_ptr: cute.struct.MemRange[cutlass.Int64, self.ab_stage * 2]
+            ab_pipeline_array_ptr: cute.struct.MemRange[cutlass.Int64, self.ab_stage * 2]
+            epi_pipeline_array_ptr: cute.struct.MemRange[cutlass.Int64, self.epi_c_stage * 2]
+            sched_pipeline_array_ptr: cute.struct.MemRange[cutlass.Int64, self.sched_stage * 2]
+            tile_count: cute.struct.MemRange[cutlass.Int32, self.sched_stage]
+            sD: cute.struct.Align[
+                cute.struct.MemRange[
+                    self.d_dtype if self.d_dtype is not None else Int32, epi_smem_size
+                ],
+                self.buffer_align_bytes,
+            ]
+            sC: cute.struct.Align[
+                cute.struct.MemRange[
+                    self.c_dtype if self.c_dtype is not None else Int32, epi_c_smem_size
+                ],
+                self.buffer_align_bytes,
+            ]
+            epi: self.epi_get_smem_struct(epilogue_params)
             sA: cute.struct.Align[
                 cute.struct.MemRange[self.a_dtype, cute.cosize(self.a_smem_layout_staged)],
                 self.buffer_align_bytes,
@@ -452,16 +535,26 @@ class HopperWgmmaGemmKernel:
         # Launch the kernel synchronously
         self.kernel(
             tma_atom_a,
-            tma_tensor_a,
+            tma_tensor_a if const_expr(not self.gather_A) else mA,
             tma_atom_b,
             tma_tensor_b,
             tma_atom_d,
             tma_tensor_d,
-            tiled_mma,
-            self.cta_layout_mnk,
+            tma_atom_c,
+            tma_tensor_c,
+            epilogue_params,
+            mAIdx,
+            varlen_args.mCuSeqlensM,
+            varlen_args.mCuSeqlensK,
+            varlen_args.mTensormaps,
+            self.tiled_mma,
+            self.cluster_layout_mnk,
             self.a_smem_layout_staged,
             self.b_smem_layout_staged,
             self.epi_smem_layout_staged,
+            self.epi_c_smem_layout_staged,
+            tile_sched_params,
+            TileSchedulerCls,
         ).launch(
             grid=grid,
             block=[self.threads_per_cta, 1, 1],
@@ -476,17 +569,27 @@ class HopperWgmmaGemmKernel:
     @cute.kernel
     def kernel(
         self,
-        tma_atom_a: cute.CopyAtom,
+        tma_atom_a: Optional[cute.CopyAtom],
         mA_mkl: cute.Tensor,
         tma_atom_b: cute.CopyAtom,
         mB_nkl: cute.Tensor,
-        tma_atom_d: cute.CopyAtom,
-        mD_mnl: cute.Tensor,
+        tma_atom_d: Optional[cute.CopyAtom],
+        mD_mnl: Optional[cute.Tensor],
+        tma_atom_c: Optional[cute.CopyAtom],
+        mC_mnl: Optional[cute.Tensor],
+        epilogue_params: ParamsBase,
+        mAIdx: Optional[cute.Tensor],
+        cu_seqlens_m: Optional[cute.Tensor],
+        cu_seqlens_k: Optional[cute.Tensor],
+        tensormaps: Optional[cute.Tensor],
         tiled_mma: cute.TiledMma,
-        cta_layout_mnk: cute.Layout,
-        a_smem_layout_staged: cute.ComposedLayout,
-        b_smem_layout_staged: cute.ComposedLayout,
-        epi_smem_layout_staged: cute.ComposedLayout,
+        cluster_layout_mnk: cute.Layout,
+        a_smem_layout: cute.ComposedLayout,
+        b_smem_layout: cute.ComposedLayout,
+        epi_smem_layout: cute.ComposedLayout,
+        epi_c_smem_layout: cute.ComposedLayout,
+        tile_sched_params: ParamsBase,
+        TileSchedulerCls: cutlass.Constexpr[Callable],
     ):
         """
         GPU device kernel performing the batched GEMM computation.
@@ -505,32 +608,31 @@ class HopperWgmmaGemmKernel:
         :type mD_mnl: cute.Tensor
         :param tiled_mma: Tiled MMA object
         :type tiled_mma: cute.TiledMma
-        :param cta_layout_mnk: CTA layout
-        :type cta_layout_mnk: cute.Layout
-        :param a_smem_layout_staged: Shared memory layout for A
-        :type a_smem_layout_staged: cute.ComposedLayout
-        :param b_smem_layout_staged: Shared memory layout for B
-        :type b_smem_layout_staged: cute.ComposedLayout
-        :param epi_smem_layout_staged: Shared memory layout for epilogue
-        :type epi_smem_layout_staged: cute.ComposedLayout
+        :param cluster_layout_mnk: CTA layout
+        :type cluster_layout_mnk: cute.Layout
+        :param a_smem_layout: Shared memory layout for A
+        :type a_smem_layout: cute.ComposedLayout
+        :param b_smem_layout: Shared memory layout for B
+        :type b_smem_layout: cute.ComposedLayout
+        :param epi_smem_layout: Shared memory layout for epilogue
+        :type epi_smem_layout: cute.ComposedLayout
         """
 
+        varlen_m = const_expr(cu_seqlens_m is not None)
+        varlen_k = const_expr(cu_seqlens_k is not None)
+        assert not (varlen_m and varlen_k)
+        has_D = const_expr(mD_mnl is not None)
+        has_C = const_expr(mC_mnl is not None)
+
         warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
 
         # /////////////////////////////////////////////////////////////////////////////
         #  Prefetch Tma desc
         # /////////////////////////////////////////////////////////////////////////////
-        # if warp_idx == 0:
-        if warp_idx == self.mma_warp_groups * 4:
-            cpasync.prefetch_descriptor(tma_atom_a)
-            cpasync.prefetch_descriptor(tma_atom_b)
-            cpasync.prefetch_descriptor(tma_atom_d)
-
-        a_smem_layout = cute.slice_(a_smem_layout_staged, (None, None, 0))
-        b_smem_layout = cute.slice_(b_smem_layout_staged, (None, None, 0))
-        tma_copy_bytes = cute.size_in_bytes(self.a_dtype, a_smem_layout) + cute.size_in_bytes(
-            self.b_dtype, b_smem_layout
-        )
+        if warp_idx == self.ab_load_warp_id:
+            for tma_atom in (tma_atom_a, tma_atom_b, tma_atom_d, tma_atom_c):
+                if const_expr(tma_atom is not None):
+                    cpasync.prefetch_descriptor(tma_atom)
 
         # /////////////////////////////////////////////////////////////////////////////
         #  Alloc and init AB full/empty + ACC full mbar (pipeline)
@@ -538,164 +640,321 @@ class HopperWgmmaGemmKernel:
         smem = cutlass.utils.SmemAllocator()
         storage = smem.allocate(self.shared_storage)
 
-        # Threads/warps participating in this pipeline
-        mainloop_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
-        # Each warp will constribute to the arrive count with the number of mcast size
-        mcast_size = self.num_mcast_ctas_a + self.num_mcast_ctas_b - 1
-        consumer_arrive_cnt = mcast_size * (self.num_mma_threads // cute.arch.WARP_SIZE)
-        mainloop_pipeline_consumer_group = pipeline.CooperativeGroup(
-            pipeline.Agent.Thread, consumer_arrive_cnt
-        )
-
-        cta_layout_vmnk = cute.make_layout((1, *cta_layout_mnk.shape))
-        mainloop_pipeline = pipeline.PipelineTmaAsync.create(
-            barrier_storage=storage.mainloop_pipeline_array_ptr.data_ptr(),
-            num_stages=self.ab_stage,
-            producer_group=mainloop_pipeline_producer_group,
-            consumer_group=mainloop_pipeline_consumer_group,
-            tx_count=tma_copy_bytes,
-            cta_layout_vmnk=cta_layout_vmnk,
+        ab_pipeline = self.make_ab_pipeline(
+            a_smem_layout=cute.slice_(a_smem_layout, (None, None, 0)),
+            b_smem_layout=cute.slice_(b_smem_layout, (None, None, 0)),
+            tiled_mma=tiled_mma,
+            cluster_layout_vmnk=cute.make_layout((1, *cluster_layout_mnk.shape)),
+            ab_pipeline_mbar_ptr=storage.ab_pipeline_array_ptr.data_ptr(),
         )
+        epi_pipeline = None
+        if const_expr(has_C):
+            epi_pipeline = self.make_epi_pipeline(
+                c_smem_layout=cute.slice_(epi_c_smem_layout, (None, None, 0)),
+                epi_pipeline_mbar_ptr=storage.epi_pipeline_array_ptr.data_ptr(),
+            )
+        sched_pipeline = None
+        tile_count = None
+        if const_expr(tile_sched_params.tile_count_semaphore is not None):
+            # Dynamic persistent scheduler
+            sched_pipeline = self.make_sched_pipeline(
+                cluster_layout_mnk,
+                sched_pipeline_mbar_ptr=storage.sched_pipeline_array_ptr.data_ptr(),
+                varlen_k=varlen_k,
+            )
+            tile_count = storage.tile_count.get_tensor((self.sched_stage,))
 
         # ///////////////////////////////////////////////////////////////////////////////
         #  Generate smem tensor A/B
         # ///////////////////////////////////////////////////////////////////////////////
-        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)
-        sD_ptr = cute.recast_ptr(sA.iterator, epi_smem_layout_staged.inner, dtype=self.d_dtype)
-        sD = cute.make_tensor(sD_ptr, epi_smem_layout_staged.outer)
-
-        # ///////////////////////////////////////////////////////////////////////////////
-        #  Get cta/warp/thread idx
-        # ///////////////////////////////////////////////////////////////////////////////
-
-        cidx, cidy, _ = cute.arch.cluster_idx()
-        cdimx, cdimy, _ = cute.arch.cluster_dim()
-        cluster_id = cidx + cdimx * cidy
+        sA = storage.sA.get_tensor(a_smem_layout.outer, swizzle=a_smem_layout.inner)
+        sB = storage.sB.get_tensor(b_smem_layout.outer, swizzle=b_smem_layout.inner)
+        sD = None
+        if const_expr(has_D):
+            if const_expr(not self.is_persistent):
+                sD_ptr = cute.recast_ptr(sA.iterator, epi_smem_layout.inner, dtype=self.d_dtype)
+                sD = cute.make_tensor(sD_ptr, epi_smem_layout.outer)
+            else:
+                sD = storage.sD.get_tensor(epi_smem_layout.outer, swizzle=epi_smem_layout.inner)
+        sC = None
+        if const_expr(has_C):
+            sC = storage.sC.get_tensor(epi_c_smem_layout.outer, swizzle=epi_c_smem_layout.inner)
+        epi_smem_tensors = self.epi_get_smem_tensors(epilogue_params, storage)
+
+        # Get tensormap buffer address
+        tensormap_manager = None
+        tensormap_a_ptr, tensormap_b_ptr, tensormap_d_ptr = None, None, None
+        if const_expr(varlen_m or varlen_k):
+            tensormap_manager = TensorMapManagerSm90(
+                cutlass.utils.TensorMapUpdateMode.GMEM, GemmSm90.bytes_per_tensormap
+            )
+            # equivalent to bidx + bidy * gridDim.x + bidxz * gridDim.x * gridDim.y
+            tensormap_workspace_idx = cute.make_layout(cute.arch.grid_dim())(cute.arch.block_idx())
+            if const_expr(varlen_m):
+                tensormap_d_idx = warp_idx // 4 if const_expr(self.pingpong) else 0
+                tensormap_d_ptr = tensormap_manager.get_tensormap_ptr(
+                    tensormaps[tensormap_workspace_idx, tensormap_d_idx, None].iterator
+                )
+            else:
+                assert varlen_k
+                tensormap_a_ptr = tensormap_manager.get_tensormap_ptr(
+                    tensormaps[tensormap_workspace_idx, 0, None].iterator
+                )
+                tensormap_b_ptr = tensormap_manager.get_tensormap_ptr(
+                    tensormaps[tensormap_workspace_idx, 1, None].iterator
+                )
 
-        # CTA Swizzle to promote L2 data reuse
-        group_size_m = 8
-        s_shape = (
-            (group_size_m, cdimx // group_size_m),
-            cdimy,
+        TileSchedulerCls = partial(
+            TileSchedulerCls.create, tile_sched_params, tile_count, sched_pipeline
         )
-        s_stride = ((1, cdimy * group_size_m), group_size_m)
-        s_layout = cute.make_layout(s_shape, stride=s_stride)
-        num_reg_cids = cute.size(s_shape)
-        cid_m, cid_n = s_layout.get_flat_coord(cluster_id % num_reg_cids)
-
-        # Deal with the tail part
-        if cluster_id >= num_reg_cids:
-            tail_size_m = cdimx % group_size_m
-            tail_layout = cute.make_layout((tail_size_m, cdimy), stride=(1, tail_size_m))
-            tail_cid = cluster_id - num_reg_cids
-            tail_cid_m, tail_cid_n = tail_layout.get_flat_coord(tail_cid)
-            cid_m = cute.size(s_shape, mode=[0]) + tail_cid_m
-            cid_n = tail_cid_n
-
-        # Get the pid from cluster id
-        bidx_in_cluster = cute.arch.block_in_cluster_idx()
-        pid_m = cid_m * self.cluster_shape_mnk[0] + bidx_in_cluster[0]
-        pid_n = cid_n * self.cluster_shape_mnk[1] + bidx_in_cluster[1]
-
-        _, _, bidz = cute.arch.block_idx()
-        tile_coord_mnkl = (pid_m, pid_n, None, bidz)
-        cta_rank_in_cluster = cute.arch.make_warp_uniform(cute.arch.block_idx_in_cluster())
-        cluster_coord_mnk = cta_layout_mnk.get_flat_coord(cta_rank_in_cluster)
-
-        k_tile_cnt = cute.ceil_div(cute.size(mA_mkl.shape[1]), self.tile_shape_mnk[2])
-
-        if warp_idx >= self.mma_warp_groups * 4:
+
+        if warp_idx >= self.ab_load_warp_id:
             cute.arch.warpgroup_reg_dealloc(self.num_regs_load)
-            if warp_idx == self.mma_warp_groups * 4:
+            if (
+                warp_idx >= self.ab_load_warp_id
+                and warp_idx < self.ab_load_warp_id + self.num_ab_load_warps
+            ):
+                is_tma_warp = self.num_ab_load_warps == 1 or warp_idx == self.ab_load_warp_id
+                if const_expr(varlen_k):
+                    # initialize tensormap for A & B
+                    tensormap_manager.init_tensormap_from_atom(
+                        tma_atom_a,
+                        tensormap_a_ptr,
+                        is_tma_warp,
+                    )
+                    tensormap_manager.init_tensormap_from_atom(
+                        tma_atom_b,
+                        tensormap_b_ptr,
+                        is_tma_warp,
+                    )
                 # ///////////////////////////////////////////////////////////////////////////////
                 # Get mcast mask
                 # ///////////////////////////////////////////////////////////////////////////////
+                cta_rank_in_cluster = cute.arch.make_warp_uniform(cute.arch.block_idx_in_cluster())
+                cluster_coord_mnk = cluster_layout_mnk.get_flat_coord(cta_rank_in_cluster)
                 a_mcast_mask = cute.make_layout_image_mask(
-                    cta_layout_mnk, cluster_coord_mnk, mode=1
+                    cluster_layout_mnk, cluster_coord_mnk, mode=1
                 )
                 b_mcast_mask = cute.make_layout_image_mask(
-                    cta_layout_mnk, cluster_coord_mnk, mode=0
+                    cluster_layout_mnk, cluster_coord_mnk, mode=0
                 )
                 a_mcast_mask = a_mcast_mask if self.is_a_mcast else 0
                 b_mcast_mask = b_mcast_mask if self.is_b_mcast else 0
-                mainloop_producer_state = pipeline.make_pipeline_state(
+
+                # Persistent tile scheduling loop
+                is_scheduler_warp = self.num_ab_load_warps == 1 or warp_idx == self.ab_load_warp_id
+                if const_expr(cute.size(cluster_layout_mnk) > 1):
+                    is_scheduler_warp = is_scheduler_warp and cute.arch.block_idx_in_cluster() == 0
+                tile_scheduler = TileSchedulerCls(is_scheduler_warp=is_scheduler_warp)
+                work_tile = tile_scheduler.initial_work_tile_info()
+                ab_producer_state = make_pipeline_state(
                     pipeline.PipelineUserType.Producer, self.ab_stage
                 )
-                # ///////////////////////////////////////////////////////////////////////////////
-                #  Local_tile partition global tensors
-                # ///////////////////////////////////////////////////////////////////////////////
-                # (bM, bK, RestK)
-                gA_mkl = cute.local_tile(
-                    mA_mkl, self.tile_shape_mnk, tile_coord_mnkl, proj=(1, None, 1)
-                )
-                # (bN, bK, RestK)
-                gB_nkl = cute.local_tile(
-                    mB_nkl, self.tile_shape_mnk, tile_coord_mnkl, proj=(None, 1, 1)
-                )
-                # //////////////////////////////////////////////////////////////////////////////
-                #  Partition shared tensor for TMA load A/B
-                # //////////////////////////////////////////////////////////////////////////////
-                #  TMA load A partition_S/D
-                a_cta_layout = cute.make_layout(cute.slice_(cta_layout_mnk, (0, None, 0)).shape)
-                a_cta_crd = cluster_coord_mnk[1]
-                tAsA, tAgA_mkl = cpasync.tma_partition(
-                    tma_atom_a,
-                    a_cta_crd,
-                    a_cta_layout,
-                    cute.group_modes(sA, 0, 2),
-                    cute.group_modes(gA_mkl, 0, 2),
-                )
-                # TMA load B partition_S/D
-                b_cta_layout = cute.make_layout(cute.slice_(cta_layout_mnk, (None, 0, 0)).shape)
-                b_cta_crd = cluster_coord_mnk[0]
-                tBsB, tBgB_nkl = cpasync.tma_partition(
-                    tma_atom_b,
-                    b_cta_crd,
-                    b_cta_layout,
-                    cute.group_modes(sB, 0, 2),
-                    cute.group_modes(gB_nkl, 0, 2),
-                )
-                # /////////////////////////////////////////////////////////////////////////////
-                # TMA load
-                # /////////////////////////////////////////////////////////////////////////////
-                for k_tile in cutlass.range(k_tile_cnt, unroll=1):
-                    # Wait for A/B buffers to be empty before loading into them
-                    # Also sets the transaction barrier for the A/B buffers
-                    mainloop_pipeline.producer_acquire(mainloop_producer_state)
-                    # /////////////////////////////////////////////////////////////////////////////
-                    #  TMA load A/B
-                    # /////////////////////////////////////////////////////////////////////////////
-                    cute.copy(
-                        tma_atom_a,
-                        tAgA_mkl[None, k_tile],
-                        tAsA[None, mainloop_producer_state.index],
-                        tma_bar_ptr=mainloop_pipeline.producer_get_barrier(mainloop_producer_state),
-                        mcast_mask=a_mcast_mask,
+                if const_expr(varlen_k):
+                    # wait tensormap initialization complete before update
+                    tensormap_manager.fence_tensormap_initialization()
+                # batch index of last tile
+                last_batch_idx = cutlass.Int32(-1)
+                while work_tile.is_valid_tile:
+                    tile_coord_mnkl = work_tile.tile_idx
+                    batch_idx = tile_coord_mnkl[3]
+                    if const_expr(varlen_k):
+                        is_group_changed = batch_idx != last_batch_idx
+                        last_batch_idx = batch_idx
+                        if is_group_changed:
+                            # construct tensor A/B based on real address, shape and stride information
+                            tensormap_manager.update_tensormap_shape(
+                                (tensormap_a_ptr, tensormap_b_ptr),
+                                is_manager_warp=is_tma_warp,
+                                shapes=(cu_seqlens_k[batch_idx + 1], cu_seqlens_k[batch_idx + 1]),
+                                orders=(
+                                    0 if const_expr(self.a_layout == LayoutEnum.ROW_MAJOR) else 1,
+                                    0 if const_expr(self.b_layout == LayoutEnum.ROW_MAJOR) else 1,
+                                ),
+                                tensormap_smem_ptr=None,
+                            )
+                    # ///////////////////////////////////////////////////////////////////////////
+                    #  Local_tile partition global tensors
+                    # ///////////////////////////////////////////////////////////////////////////
+                    if const_expr(not self.gather_A):
+                        if const_expr(varlen_m):
+                            mA_mk = cute.domain_offset((cu_seqlens_m[batch_idx], 0), mA_mkl)
+                        elif const_expr(varlen_k):
+                            mA_mk = cute.domain_offset((0, cu_seqlens_k[batch_idx]), mA_mkl)
+                        else:
+                            mA_mk = mA_mkl[None, None, batch_idx]
+                        # (bM, bK, RestK)
+                        gA_k = cute.local_tile(
+                            mA_mk,
+                            cute.select(self.tile_shape_mnk, [0, 2]),
+                            (tile_coord_mnkl[0], None),
+                        )
+                    else:
+                        mA_mk = mA_mkl
+                        if const_expr(varlen_m):
+                            mAIdx_mk = cute.domain_offset((cu_seqlens_m[batch_idx],), mAIdx)
+                        elif const_expr(varlen_k):
+                            mAIdx_mk = cute.domain_offset((cu_seqlens_k[batch_idx],), mAIdx)
+                        else:
+                            mAIdx_mk = mAIdx[None, batch_idx]
+                        gAIdx = cute.local_tile(
+                            mAIdx_mk, (self.tile_shape_mnk[0],), (tile_coord_mnkl[0],)
+                        )
+                    if const_expr(varlen_k):
+                        mB_nk = cute.domain_offset((0, cu_seqlens_k[batch_idx]), mB_nkl)
+                    else:
+                        mB_nk = mB_nkl[None, None, batch_idx]
+                    # (bN, bK, RestK)
+                    gB_k = cute.local_tile(
+                        mB_nk, cute.select(self.tile_shape_mnk, [1, 2]), (tile_coord_mnkl[1], None)
                     )
-                    cute.copy(
+                    # //////////////////////////////////////////////////////////////////////////
+                    #  Partition shared tensor for TMA load A/B
+                    # //////////////////////////////////////////////////////////////////////////
+                    if const_expr(varlen_k):
+                        # ensure the update to tensormap has completed before using it
+                        if is_group_changed and is_tma_warp:
+                            tensormap_manager.fence_tensormap_update(tensormap_a_ptr)
+                            tensormap_manager.fence_tensormap_update(tensormap_b_ptr)
+                        tma_desc_a_ptr = tensormap_manager.get_tensormap_ptr(
+                            tensormap_a_ptr, cute.AddressSpace.generic
+                        )
+                        tma_desc_b_ptr = tensormap_manager.get_tensormap_ptr(
+                            tensormap_b_ptr, cute.AddressSpace.generic
+                        )
+                    else:
+                        tma_desc_a_ptr, tma_desc_b_ptr = None, None
+                    #  TMA load A partition_S/D
+                    a_cta_layout = cute.make_layout(
+                        cute.slice_(cluster_layout_mnk, (0, None, 0)).shape
+                    )
+                    a_cta_crd = cluster_coord_mnk[1]
+                    if const_expr(not self.gather_A):
+                        # ((atom_v, rest_v), STAGE), ((atom_v, rest_v), RestK)
+                        tAsA, tAgA_k = cpasync.tma_partition(
+                            tma_atom_a,
+                            a_cta_crd,
+                            a_cta_layout,
+                            cute.group_modes(sA, 0, 2),
+                            cute.group_modes(gA_k, 0, 2),
+                        )
+                        copy_A = partial(
+                            cute.copy,
+                            tma_atom_a,
+                            mcast_mask=a_mcast_mask,
+                            tma_desc_ptr=tma_desc_a_ptr,
+                        )
+                    else:
+                        tiled_copy_A = self._make_gmem_tiled_copy_A(
+                            mA_mkl.element_type, self.a_layout, self.num_ab_load_threads
+                        )
+                        tidx = (
+                            cute.arch.thread_idx()[0]
+                            - self.mma_warp_groups * self.num_threads_per_warp_group
+                        )
+                        thr_copy_A = tiled_copy_A.get_slice(tidx)
+                        # (atom_v, CPY_M, 1, STAGE)
+                        tAsA = thr_copy_A.partition_D(sA)
+                        assert tAsA.shape[2] == 1
+                        tAsA = cute.group_modes(cute.slice_(tAsA, (None, None, 0, None)), 0, 2)
+                        copy_A = partial(cute.copy, tiled_copy_A)
+                    # TMA load B partition_S/D
+                    b_cta_layout = cute.make_layout(
+                        cute.slice_(cluster_layout_mnk, (None, 0, 0)).shape
+                    )
+                    b_cta_crd = cluster_coord_mnk[0]
+                    # ((atom_v, rest_v), STAGE), ((atom_v, rest_v), RestK)
+                    tBsB, tBgB_k = cpasync.tma_partition(
                         tma_atom_b,
-                        tBgB_nkl[None, k_tile],
-                        tBsB[None, mainloop_producer_state.index],
-                        tma_bar_ptr=mainloop_pipeline.producer_get_barrier(mainloop_producer_state),
-                        mcast_mask=b_mcast_mask,
+                        b_cta_crd,
+                        b_cta_layout,
+                        cute.group_modes(sB, 0, 2),
+                        cute.group_modes(gB_k, 0, 2),
                     )
-                    # Mainloop pipeline's producer commit is a NOP
-                    mainloop_pipeline.producer_commit(mainloop_producer_state)
-                    mainloop_producer_state.advance()
-                mainloop_pipeline.producer_tail(mainloop_producer_state)
-
-        if warp_idx < self.mma_warp_groups * 4:
+                    copy_B = partial(
+                        cute.copy, tma_atom_b, mcast_mask=b_mcast_mask, tma_desc_ptr=tma_desc_b_ptr
+                    )
+                    k_len = (
+                        cu_seqlens_k[batch_idx + 1] - cu_seqlens_k[batch_idx]
+                        if const_expr(varlen_k)
+                        else mA_mkl.shape[1]
+                    )
+                    k_tile_cnt = cute.ceil_div(k_len, self.tile_shape_mnk[2])
+                    if const_expr(not self.gather_A):
+                        ab_producer_state = self.load_AB(
+                            ab_pipeline,
+                            ab_producer_state,
+                            copy_A,
+                            tAgA_k,
+                            tAsA,
+                            copy_B,
+                            tBgB_k,
+                            tBsB,
+                            k_tile_cnt,
+                        )
+                    else:
+                        limit_m = (
+                            mAIdx.shape[0]
+                            if const_expr(cu_seqlens_m is None)
+                            else cu_seqlens_m[batch_idx + 1] - cu_seqlens_m[batch_idx]
+                        )
+                        ab_producer_state = self.load_AB_gather_A(
+                            ab_pipeline,
+                            ab_producer_state,
+                            thr_copy_A,
+                            mA_mk,
+                            tAsA,
+                            gAIdx,
+                            copy_B,
+                            tBgB_k,
+                            tBsB,
+                            k_tile_cnt,
+                            limit_A=(
+                                limit_m - tile_coord_mnkl[0] * self.tile_shape_mnk[0],
+                                mA_mk.shape[1],
+                            ),
+                        )
+                    tile_scheduler.fetch_next_work(is_scheduler_warp=is_scheduler_warp)
+                    tile_scheduler.broadcast_next_work(is_scheduler_warp=is_scheduler_warp)
+                    tile_scheduler.advance_to_next_work(is_scheduler_warp=is_scheduler_warp)
+                    work_tile = tile_scheduler.get_current_work()
+                    # End of persistent scheduler loop
+                if const_expr(self.pingpong and not varlen_k):
+                    # Need to write the tile_idx to smem for the next WG in the pingpong mode
+                    # tile_scheduler.advance_to_next_work(is_scheduler_warp=is_scheduler_warp)
+                    tile_scheduler.broadcast_next_work(is_scheduler_warp=is_scheduler_warp)
+                    tile_scheduler.advance_to_next_work(is_scheduler_warp=is_scheduler_warp)
+                ab_pipeline.producer_tail(ab_producer_state)
+                if is_scheduler_warp:
+                    tile_scheduler.producer_tail()
+
+        if warp_idx < self.ab_load_warp_id:
             cute.arch.warpgroup_reg_alloc(self.num_regs_mma)
+            is_tma_warp = Boolean(
+                (not self.pingpong and warp_idx == 0)
+                or (self.pingpong and (warp_idx == 0 or warp_idx == 4))
+            )
+            if const_expr(varlen_m):
+                # initialize tensormap for D
+                tensormap_manager.init_tensormap_from_atom(
+                    tma_atom_d,
+                    tensormap_d_ptr,
+                    is_manager_warp=is_tma_warp,
+                )
             # //////////////////////////////////////////////////////////////////////////////
             #  Partition global tensor for TiledMMA_A/B/C
             # //////////////////////////////////////////////////////////////////////////////
             tidx, _, _ = cute.arch.thread_idx()
             warp_group_idx = cute.arch.make_warp_uniform(tidx // self.num_threads_per_warp_group)
+            if const_expr(self.pingpong):
+                tidx = tidx % self.num_threads_per_warp_group
             warp_group_thread_layout = cute.make_layout(
-                self.mma_warp_groups, stride=self.num_threads_per_warp_group
+                self.mma_warp_groups if not self.pingpong else 1,
+                stride=self.num_threads_per_warp_group,
+            )
+            thr_mma = tiled_mma.get_slice(
+                warp_group_thread_layout(warp_group_idx if not self.pingpong else 0)
             )
-            thr_mma = tiled_mma.get_slice(warp_group_thread_layout(warp_group_idx))
 
             # //////////////////////////////////////////////////////////////////////////////
             #  Make fragments
@@ -705,148 +964,818 @@ class HopperWgmmaGemmKernel:
 
             acc_shape = tiled_mma.partition_shape_C(cute.select(self.tile_shape_mnk, mode=[0, 1]))
             acc = cute.make_fragment(acc_shape, self.acc_dtype)
-
-            mainloop_consumer_read_state = pipeline.make_pipeline_state(
-                pipeline.PipelineUserType.Consumer, self.ab_stage
+            acc_slow = None
+            if const_expr(self.fp8_slow_accum):
+                acc_slow = cute.make_fragment(acc_shape, self.acc_dtype)
+
+            if const_expr(self.pingpong):
+                if warp_group_idx == 0:
+                    # WG0 needs a start signal at the very beginning
+                    self.pingpong_barrier_arrive(warp_group_idx=0, stage="mma")
+                    self.pingpong_barrier_arrive(warp_group_idx=0, stage="epi")
+
+            k_tile_cnt_static = cute.ceil_div(mA_mkl.shape[1], self.tile_shape_mnk[2])
+            c_tile_cnt = cute.size(cute.ceil_div(self.tile_shape_mnk[:2], self.epi_tile))
+
+            ab_read_state = make_pipeline_state(pipeline.PipelineUserType.Consumer, self.ab_stage)
+            epi_read_state = make_pipeline_state(
+                pipeline.PipelineUserType.Consumer, self.epi_c_stage
             )
-            mainloop_consumer_release_state = pipeline.make_pipeline_state(
-                pipeline.PipelineUserType.Consumer, self.ab_stage
+            epi_producer_state = make_pipeline_state(
+                pipeline.PipelineUserType.Producer, self.epi_c_stage
             )
+            tile_scheduler = TileSchedulerCls()
+            work_tile = None
+            if const_expr(self.pingpong):
+                if const_expr(varlen_k):
+                    work_tile = tile_scheduler.initial_work_tile_info()
+                if warp_idx >= 4:
+                    # Advance 2nd Math WG pipeline states to the end of 1st Math WG
+                    epi_read_state.advance_iters(c_tile_cnt)
+                    epi_producer_state.advance_iters(c_tile_cnt)
+                    if const_expr(not varlen_k):
+                        ab_read_state.advance_iters(k_tile_cnt_static)
+                    else:
+                        batch_idx = work_tile.tile_idx[3]
+                        k_len = cu_seqlens_k[batch_idx + 1] - cu_seqlens_k[batch_idx]
+                        k_tile_cnt = cute.ceil_div(k_len, self.tile_shape_mnk[2])
+                        ab_read_state.advance_iters(k_tile_cnt)
+                    tile_scheduler.advance_to_next_work()
+                    if const_expr(varlen_k):
+                        work_tile = tile_scheduler.get_current_work()
+                if const_expr(not varlen_k):
+                    work_tile = tile_scheduler.initial_work_tile_info()
+            else:
+                work_tile = tile_scheduler.initial_work_tile_info()
+            if const_expr(varlen_m):
+                # wait tensormap initialization complete before update
+                tensormap_manager.fence_tensormap_initialization()
+            # batch index of last tile
+            last_batch_idx = cutlass.Int32(-1)
+            while work_tile.is_valid_tile:
+                tile_coord_mnkl = work_tile.tile_idx
+                batch_idx = tile_coord_mnkl[3]
+                if const_expr(varlen_m):
+                    is_group_changed = batch_idx != last_batch_idx
+                    last_batch_idx = batch_idx
+                    if is_group_changed:
+                        # construct tensor D based on real address, shape and stride information
+                        tensormap_manager.update_tensormap_shape(
+                            (tensormap_d_ptr,),
+                            is_manager_warp=is_tma_warp,
+                            shapes=(cu_seqlens_m[batch_idx + 1],),
+                            orders=(0 if const_expr(self.d_layout.is_m_major_c()) else 1,),
+                            tensormap_smem_ptr=None,
+                        )
+
+                k_len = (
+                    cu_seqlens_k[batch_idx + 1] - cu_seqlens_k[batch_idx]
+                    if const_expr(varlen_k)
+                    else mA_mkl.shape[1]
+                )
+                k_tile_cnt = cute.ceil_div(k_len, self.tile_shape_mnk[2])
+                ab_read_state, tiled_mma = self.mma(
+                    ab_pipeline,
+                    ab_read_state,
+                    tiled_mma,
+                    tCrA,
+                    tCrB,
+                    acc,
+                    acc_slow,
+                    k_tile_cnt,
+                    warp_group_idx,
+                )
+                if const_expr(varlen_k):
+                    if k_tile_cnt == 0:
+                        acc.fill(0.0)
+
+                # /////////////////////////////////////////////////////////////////////////////
+                #  EPILOGUE
+                # /////////////////////////////////////////////////////////////////////////////
+                if const_expr(self.pingpong):
+                    self.pingpong_barrier_sync(warp_group_idx, "epi")
 
-            # /////////////////////////////////////////////////////////////////////////////
-            #  Prologue MMAs
-            # /////////////////////////////////////////////////////////////////////////////
-            k_pipe_mmas = 1
-            peek_ab_full_status = cutlass.Boolean(1)
-            if mainloop_consumer_read_state.count < k_tile_cnt:
-                peek_ab_full_status = mainloop_pipeline.consumer_try_wait(
-                    mainloop_consumer_read_state
+                epilogue_barrier = pipeline.NamedBarrier(
+                    barrier_id=int(NamedBarrierGemm.Epilogue), num_threads=self.num_epi_threads
                 )
-            tiled_mma.set(warpgroup.Field.ACCUMULATE, False)
-            num_k_blocks = cute.size(tCrA, mode=[2])
-            for k_tile in cutlass.range_constexpr(k_pipe_mmas):
-                # Wait for A/B buffer to be ready
-                mainloop_pipeline.consumer_wait(mainloop_consumer_read_state, peek_ab_full_status)
-                warpgroup.fence()
-                for k_block_idx in cutlass.range(num_k_blocks, unroll_full=True):
-                    k_block_coord = (None, None, k_block_idx, mainloop_consumer_read_state.index)
-                    cute.gemm(tiled_mma, acc, tCrA[k_block_coord], tCrB[k_block_coord], acc)
-                    tiled_mma.set(warpgroup.Field.ACCUMULATE, True)
-                warpgroup.commit_group()
-                mainloop_consumer_read_state.advance()
-                peek_ab_full_status = cutlass.Boolean(1)
-                if mainloop_consumer_read_state.count < k_tile_cnt:
-                    peek_ab_full_status = mainloop_pipeline.consumer_try_wait(
-                        mainloop_consumer_read_state
-                    )
 
-            # /////////////////////////////////////////////////////////////////////////////
-            #  MAINLOOP
-            # /////////////////////////////////////////////////////////////////////////////
-            for k_tile in cutlass.range(k_pipe_mmas, k_tile_cnt, unroll=1):
-                # Wait for TMA copies to complete
-                mainloop_pipeline.consumer_wait(mainloop_consumer_read_state, peek_ab_full_status)
-                # WGMMA
-                warpgroup.fence()
-                for k_block_idx in cutlass.range(num_k_blocks, unroll_full=True):
-                    k_block_coord = (None, None, k_block_idx, mainloop_consumer_read_state.index)
-                    cute.gemm(tiled_mma, acc, tCrA[k_block_coord], tCrB[k_block_coord], acc)
-                warpgroup.commit_group()
-                # Wait on the wgmma barrier for previous k_pipe_mmas wgmmas to complete
-                warpgroup.wait_group(k_pipe_mmas)
-                mainloop_pipeline.consumer_release(mainloop_consumer_release_state)
-                mainloop_consumer_read_state.advance()
-                mainloop_consumer_release_state.advance()
-                peek_ab_full_status = cutlass.Boolean(1)
-                if mainloop_consumer_read_state.count < k_tile_cnt:
-                    peek_ab_full_status = mainloop_pipeline.consumer_try_wait(
-                        mainloop_consumer_read_state
+                if const_expr(varlen_m):
+                    # ensure the update to tensormap has completed before using it
+                    if is_group_changed and is_tma_warp:
+                        tensormap_manager.fence_tensormap_update(tensormap_d_ptr)
+                    tma_desc_d_ptr = tensormap_manager.get_tensormap_ptr(
+                        tensormap_d_ptr, cute.AddressSpace.generic
                     )
-            warpgroup.wait_group(0)
-            for k_tile in cutlass.range(k_pipe_mmas, unroll=1):
-                mainloop_pipeline.consumer_release(mainloop_consumer_release_state)
-                mainloop_consumer_release_state.advance()
-
-            # /////////////////////////////////////////////////////////////////////////////
-            #  EPILOGUE
-            # /////////////////////////////////////////////////////////////////////////////
-
-            # Wait for all warp groups in the thread block to finish, because smem for tensor A in
-            # the mainloop is reused in the epilogue.
-            cute.arch.barrier(barrier_id=1, number_of_threads=self.num_mma_threads)
-
-            copy_atom_r2s = sm90_utils.sm90_get_smem_store_op(
-                self.d_layout,
-                elem_ty_d=self.d_dtype,
-                elem_ty_acc=self.acc_dtype,
-            )
-            copy_atom_D = cute.make_copy_atom(
-                warp.StMatrix8x8x16bOp(self.d_layout.is_m_major_c(), 4),
-                self.d_dtype,
-            )
-            tiled_copy_D_Atom = cute.make_tiled_copy_C_atom(copy_atom_D, tiled_mma)
-            tiled_copy_r2s = cute.make_tiled_copy_S(copy_atom_r2s, tiled_copy_D_Atom)
-            # (R2S, R2S_M, R2S_N, PIPE_D)
-            tRS_sD = tiled_copy_r2s.get_slice(tidx).partition_D(sD)
-            # (R2S, R2S_M, R2S_N)
-            tRS_rAcc = tiled_copy_r2s.retile(acc)
-
-            # (bM, bN)
-            gD_mnl = cute.local_tile(
-                mD_mnl, self.tile_shape_mnk, tile_coord_mnkl, proj=(1, 1, None)
+                else:
+                    tma_desc_d_ptr = None
+
+                if const_expr(has_D):
+                    bSG_sD, bSG_gD = self.epilog_gmem_copy_and_partition(
+                        tma_atom_d,
+                        mD_mnl,
+                        self.tile_shape_mnk[:2],
+                        self.epi_tile,
+                        sD,
+                        tile_coord_mnkl,
+                        cu_seqlens_m,
+                    )
+                    copy_D = partial(cute.copy, tma_atom_d, tma_desc_ptr=tma_desc_d_ptr)
+                else:
+                    bSG_sD, bSG_gD, copy_D = None, None, None
+                if const_expr(has_C):
+                    bGS_sC, bGS_gC = self.epilog_gmem_copy_and_partition(
+                        tma_atom_c,
+                        mC_mnl,
+                        self.tile_shape_mnk[:2],
+                        self.epi_tile,
+                        sC,
+                        tile_coord_mnkl,
+                        cu_seqlens_m,
+                    )
+                    copy_C = partial(cute.copy, tma_atom_c)
+                    epi_load_g2s = partial(self.epi_load_g2s, epi_pipeline, copy_C, bGS_gC, bGS_sC)
+                else:
+                    epi_load_g2s = None
+
+                d_dtype_for_layout = self.d_dtype if self.d_dtype is not None else cutlass.BFloat16
+                tiled_copy_r2s, tRS_rAcc, tRS_rD, tRS_sD = self.epilog_smem_store_and_partition(
+                    tiled_mma, self.d_layout, d_dtype_for_layout, acc, sD, tidx
+                )
+                if const_expr(has_C):
+                    tiled_copy_s2r, tRS_rC, tSR_rC, tSR_sC = self.epilog_smem_load_and_partition(
+                        tiled_mma, self.c_layout, self.c_dtype, sC, tRS_rD.layout, tidx
+                    )
+                else:
+                    tiled_copy_s2r, tSR_sC, tRS_rC, tSR_rC = None, None, None, None
+
+                # Wait for all warp groups in the thread block to finish, because smem for tensor
+                # A in the mainloop is reused in the epilogue if not persistent.
+                if const_expr(not self.is_persistent):
+                    epilogue_barrier.arrive_and_wait()
+
+                self.epi_visit_acc(epilogue_params, acc, tiled_mma, tile_coord_mnkl, tidx)
+
+                epi_read_state, epi_producer_state = self.epilogue(
+                    epilogue_params,
+                    epi_smem_tensors,
+                    epi_pipeline,
+                    epi_read_state,
+                    epi_producer_state,
+                    tiled_mma,
+                    tRS_rAcc,
+                    tRS_rD,
+                    tRS_rC,
+                    tiled_copy_r2s,
+                    tRS_sD,
+                    tiled_copy_s2r,
+                    tSR_rC,
+                    tSR_sC,
+                    copy_D,
+                    bSG_sD,
+                    bSG_gD,
+                    epi_load_g2s,
+                    tile_coord_mnkl,
+                    cu_seqlens_m,
+                    epilogue_barrier,
+                    tile_scheduler,
+                    tidx,
+                    is_tma_warp,
+                )
+
+                if const_expr(self.pingpong):
+                    # With pingpong, 2 WGs write two different output tiles to the same smem,
+                    # so we have to make sure the smem content is done reading before signaling
+                    # the next WG's epilogue.
+                    if is_tma_warp:
+                        cute.arch.cp_async_bulk_wait_group(0, read=True)
+                    self.pingpong_barrier_arrive(1 - warp_group_idx, stage="epi")
+
+                if const_expr(not self.pingpong):
+                    tile_scheduler.advance_to_next_work()
+                    work_tile = tile_scheduler.get_current_work()
+                else:  # Skip a tile for pingpong
+                    # Update starting load/store pipeline states for the next tile
+                    epi_read_state.advance_iters(c_tile_cnt)
+                    epi_producer_state.advance_iters(c_tile_cnt)
+                    # Update starting mainloop pipeline state for the next tile
+                    if const_expr(not varlen_k):
+                        ab_read_state.advance_iters(k_tile_cnt_static)
+                        tile_scheduler.advance_to_next_work(advance_count=self.mma_warp_groups)
+                        work_tile = tile_scheduler.get_current_work()
+                    else:
+                        tile_scheduler.advance_to_next_work()
+                        work_tile = tile_scheduler.get_current_work()
+                        if work_tile.is_valid_tile:
+                            batch_idx = work_tile.tile_idx[3]
+                            k_len = cu_seqlens_k[batch_idx + 1] - cu_seqlens_k[batch_idx]
+                            k_tile_cnt = cute.ceil_div(k_len, self.tile_shape_mnk[2])
+                            ab_read_state.advance_iters(k_tile_cnt)
+                            tile_scheduler.advance_to_next_work()
+                            work_tile = tile_scheduler.get_current_work()
+                # End of persistent scheduler loop
+
+            if const_expr(not self.pingpong):
+                if is_tma_warp:
+                    cute.arch.cp_async_bulk_wait_group(0, read=True)
+
+    @cute.jit
+    def load_AB(
+        self,
+        ab_pipeline: cutlass.pipeline.PipelineAsync,
+        ab_producer_state: cutlass.pipeline.PipelineState,
+        copy_A: Callable,
+        tAgA: cute.Tensor,
+        tAsA: cute.Tensor,
+        copy_B: Callable,
+        tBgB: cute.Tensor,
+        tBsB: cute.Tensor,
+        k_tile_cnt: Int32,
+    ) -> cutlass.pipeline.PipelineState:
+        # Peek (try_wait) AB buffer empty for k_block = prefetch_k_tile_cnt
+        peek_ab_empty_status = Boolean(True)
+        if 0 < k_tile_cnt:
+            peek_ab_empty_status = ab_pipeline.producer_try_acquire(ab_producer_state)
+        # /////////////////////////////////////////////////////////////////////////
+        # TMA load
+        # /////////////////////////////////////////////////////////////////////////
+        for k_tile in cutlass.range(k_tile_cnt, unroll=1):
+            # Wait for A/B buffers to be empty before loading into them
+            # Also sets the transaction barrier for the A/B buffers
+            ab_pipeline.producer_acquire(ab_producer_state, peek_ab_empty_status)
+            tma_bar_ptr = ab_pipeline.producer_get_barrier(ab_producer_state)
+            copy_A(tAgA[None, k_tile], tAsA[None, ab_producer_state.index], tma_bar_ptr=tma_bar_ptr)
+            copy_B(tBgB[None, k_tile], tBsB[None, ab_producer_state.index], tma_bar_ptr=tma_bar_ptr)
+            # Mainloop pipeline's producer commit is a NOP
+            ab_pipeline.producer_commit(ab_producer_state)
+            ab_producer_state.advance()
+            peek_ab_empty_status = Boolean(True)
+            if k_tile + 1 < k_tile_cnt:
+                peek_ab_empty_status = ab_pipeline.producer_try_acquire(ab_producer_state)
+        return ab_producer_state
+
+    @cute.jit
+    def load_AB_gather_A(
+        self,
+        ab_pipeline: cutlass.pipeline.PipelineAsync,
+        ab_producer_state: cutlass.pipeline.PipelineState,
+        thr_copy_A: cute.core.ThrCopy,
+        mA: cute.Tensor,
+        tAsA: cute.Tensor,
+        gAIdx: cute.Tensor,
+        copy_B: Callable,
+        tBgB: cute.Tensor,
+        tBsB: cute.Tensor,
+        k_tile_cnt: Int32,
+        limit_A: Tuple[Int32, Int32],
+    ) -> cutlass.pipeline.PipelineState:
+        # (atom_v, CPY_M, 1, RestK)
+        limit_m, limit_k = limit_A
+        limit_m = min(limit_m, self.tile_shape_mnk[0])  # To avoid writing beyond smem limit
+        cA = cute.make_identity_tensor(cute.select(self.tile_shape_mnk, [0, 2]))
+        tAcA = thr_copy_A.partition_S(cA)
+        t0AcA = thr_copy_A.get_slice(0).partition_S(cA)
+        # Instead of comparing tAcA to limit_m, we instead compare t0AcA to limit_m - tAcA[0][0]
+        # since we know that tAcA[m][0] = t0AcA[m][0] + tAcA[0][0].
+        # This is so that when we do the comparison, t0AcA is known at compile time.
+        limit_m = limit_m - tAcA[0][0]
+        # Read indices for A
+        rows_per_thread = const_expr(cute.size(tAcA.shape, mode=[1]))
+        m_idx = cute.make_fragment(rows_per_thread, Int32)
+        for m in cutlass.range(rows_per_thread):
+            row_idx = tAcA[0, m, 0][0]
+            if t0AcA[0, m, 0][0] < limit_m:
+                m_idx[m] = gAIdx[row_idx]
+            else:
+                m_idx[m] = -1
+        elems_per_load = cute.size(tAsA.shape[0][0])
+        # (m, (bK, RestK))
+        mA_k = cute.logical_divide(mA, (None, self.tile_shape_mnk[2]))
+        warp_idx = cute.arch.make_warp_uniform(cute.arch.warp_idx())
+        # Peek (try_wait) AB buffer empty for k_block = prefetch_k_tile_cnt
+        peek_ab_empty_status = Boolean(True)
+        if 0 < k_tile_cnt:
+            peek_ab_empty_status = ab_pipeline.producer_try_acquire(ab_producer_state)
+        # /////////////////////////////////////////////////////////////////////////
+        # TMA load on B and cp.async on A
+        # /////////////////////////////////////////////////////////////////////////
+        copy_A = partial(cute.copy, thr_copy_A)
+        for k_tile in cutlass.range(k_tile_cnt - 1, unroll=1):
+            # Wait for A/B buffers to be empty before loading into them
+            # Also sets the transaction barrier for the A/B buffers
+            ab_pipeline.producer_acquire(
+                ab_producer_state,
+                peek_ab_empty_status,
+                # A tiny bit faster to rotate the warp that does TMA
+                is_tma_warp=warp_idx == self.ab_load_warp_id + (k_tile % self.num_ab_load_warps),
             )
-            tcgc_for_tma_partition = cute.zipped_divide(gD_mnl, self.epi_tile)
-            bSG_sD, bSG_gD = cpasync.tma_partition(
-                tma_atom_d,
-                0,
-                cute.make_layout(1),
-                cute.group_modes(sD, 0, 2),
-                tcgc_for_tma_partition,
+            # A bit faster to load B first while we calculate the predicate for A
+            if warp_idx == self.ab_load_warp_id + (k_tile % self.num_ab_load_warps):
+                copy_B(
+                    tBgB[None, k_tile],
+                    tBsB[None, ab_producer_state.index],
+                    tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
+                )
+            # (m, bK)
+            mA_cur = mA_k[None, (None, k_tile)]
+            for m in cutlass.range_constexpr(tAcA.shape[1]):
+                # (elems_per_load, thread_per_row)
+                mA_row = cute.tiled_divide(mA_cur[m_idx[m], None], (elems_per_load,))
+                if t0AcA[0, m, 0][0] < limit_m:
+                    # There's only 1 load per row
+                    assert cute.size(tAcA.shape, mode=[2]) == 1
+                    ki = tAcA[0, 0, 0][1] // elems_per_load
+                    copy_A(mA_row[None, ki], tAsA[(None, m), ab_producer_state.index])
+            # This tells mbarrier to track the completion of cp.async
+            ab_pipeline.producer_commit(ab_producer_state)
+            ab_producer_state.advance()
+            peek_ab_empty_status = Boolean(True)
+            if k_tile + 1 < k_tile_cnt:
+                peek_ab_empty_status = ab_pipeline.producer_try_acquire(ab_producer_state)
+        # bound checking in the K dimension on the last k_tile
+        if 0 < k_tile_cnt:
+            k_tile = k_tile_cnt - 1
+            ab_pipeline.producer_acquire(
+                ab_producer_state,
+                peek_ab_empty_status,
+                is_tma_warp=warp_idx == self.ab_load_warp_id + (k_tile % self.num_ab_load_warps),
             )
+            if warp_idx == self.ab_load_warp_id + (k_tile % self.num_ab_load_warps):
+                copy_B(
+                    tBgB[None, k_tile],
+                    tBsB[None, ab_producer_state.index],
+                    tma_bar_ptr=ab_pipeline.producer_get_barrier(ab_producer_state),
+                )
+            assert tAcA.shape[2] == 1  # there's only 1 load along the K dimension
+            tApA = cute.make_fragment(1, Boolean)
+            tApA[0] = tAcA[0, 0, 0][1] < limit_k
+            # (m, bK)
+            mA_cur = mA_k[None, (None, k_tile)]
+            for m in cutlass.range_constexpr(tAcA.shape[1]):
+                # (elems_per_load, thread_per_row)
+                mA_row = cute.tiled_divide(mA_cur[m_idx[m], None], (elems_per_load,))
+                if t0AcA[0, m, 0][0] < limit_m:
+                    # There's only 1 load per row
+                    assert cute.size(tAcA.shape, mode=[2]) == 1
+                    ki = tAcA[0, 0, 0][1] // elems_per_load
+                    # copy_A(mA_row[None, ki], tAsA[(None, m), ab_producer_state.index], pred=tApA)
+                    # TODO
+                    copy_A(mA_row[None, ki], tAsA[(None, m), ab_producer_state.index])
+            ab_pipeline.producer_commit(ab_producer_state)
+            ab_producer_state.advance()
+        return ab_producer_state
 
-            epi_tile_num = cutlass.const_expr(cute.size(tcgc_for_tma_partition, mode=[1]))
-            epi_tile_shape = tcgc_for_tma_partition.shape[1]
+    @cute.jit
+    def mma(
+        self,
+        ab_pipeline: cutlass.pipeline.PipelineAsync,
+        ab_read_state: cutlass.pipeline.PipelineState,
+        tiled_mma: cute.TiledMma,
+        tCrA: cute.Tensor,
+        tCrB: cute.Tensor,
+        acc: cute.Tensor,
+        acc_slow: Optional[cute.Tensor],
+        k_tile_cnt: Int32,
+        warp_group_idx: Int32,
+    ) -> Tuple[cutlass.pipeline.PipelineState, cute.TiledMma]:
+        # /////////////////////////////////////////////////////////////////////////////
+        #  Prologue MMAs
+        # /////////////////////////////////////////////////////////////////////////////
+        k_pipe_mmas = 1
+        ab_release_state = ab_read_state.clone()
+        num_prologue_mma = min(k_pipe_mmas, k_tile_cnt)
+        if const_expr(self.pingpong):
+            self.pingpong_barrier_sync(warp_group_idx, stage="mma")
+        peek_ab_full_status = Boolean(True)
+        if 0 < k_tile_cnt:
+            peek_ab_full_status = ab_pipeline.consumer_try_wait(ab_read_state)
+        tiled_mma.set(warpgroup.Field.ACCUMULATE, False)
+        num_k_blocks = cute.size(tCrA, mode=[2])
+        for k_tile in cutlass.range(num_prologue_mma):
+            # Wait for A/B buffer to be ready
+            ab_pipeline.consumer_wait(ab_read_state, peek_ab_full_status)
+            warpgroup.fence()
+            for k_blk_idx in cutlass.range(num_k_blocks, unroll_full=True):
+                k_blk_coord = (None, None, k_blk_idx, ab_read_state.index)
+                cute.gemm(tiled_mma, acc, tCrA[k_blk_coord], tCrB[k_blk_coord], acc)
+                tiled_mma.set(warpgroup.Field.ACCUMULATE, True)
+            warpgroup.commit_group()
+            ab_read_state.advance()
+            peek_ab_full_status = Boolean(True)
+            if k_tile + 1 < k_tile_cnt:
+                peek_ab_full_status = ab_pipeline.consumer_try_wait(ab_read_state)
+        # If k_tile_cnt == 0, this is not correct. But we will set acc to 0 in the mainloop
+        # in that case.
+        if const_expr(self.fp8_slow_accum):
+            warpgroup.wait_group(0)
+            acc_slow.store(acc.load())
 
-            for epi_idx in cutlass.range_constexpr(epi_tile_num):
-                # Copy from acc to D registers
-                tRS_rD = cute.make_fragment_like(tRS_sD[None, None, None, 0], self.acc_dtype)
-                for epi_v in cutlass.range_constexpr(cute.size(tRS_rD)):
-                    tRS_rD[epi_v] = tRS_rAcc[epi_idx * cute.size(tRS_rD) + epi_v]
-                # Type conversion
-                tRS_rD_out = cute.make_fragment_like(tRS_rD, self.d_dtype)
-                tRS_rD_out.store(tRS_rD.load().to(self.d_dtype))
-                # Copy from D registers to shared memory
-                epi_buffer = epi_idx % cute.size(tRS_sD, mode=[3])
-                # cute.arch.barrier(barrier_id=1, number_of_threads=self.num_mma_threads)
-                cute.copy(tiled_copy_r2s, tRS_rD_out, tRS_sD[(None, None, None, epi_buffer)])
+        # /////////////////////////////////////////////////////////////////////////////
+        #  MAINLOOP
+        # /////////////////////////////////////////////////////////////////////////////
+        for k_tile in cutlass.range(num_prologue_mma, k_tile_cnt, unroll=1):
+            # Wait for TMA copies to complete
+            ab_pipeline.consumer_wait(ab_read_state, peek_ab_full_status)
+            # WGMMA
+            warpgroup.fence()
+            if const_expr(self.fp8_slow_accum):
+                tiled_mma.set(warpgroup.Field.ACCUMULATE, False)
+            for k_blk_idx in cutlass.range(num_k_blocks, unroll_full=True):
+                k_blk_coord = (None, None, k_blk_idx, ab_read_state.index)
+                cute.gemm(tiled_mma, acc, tCrA[k_blk_coord], tCrB[k_blk_coord], acc)
+                tiled_mma.set(warpgroup.Field.ACCUMULATE, True)
+            warpgroup.commit_group()
+            # Wait on the wgmma barrier for previous k_pipe_mmas wgmmas to complete
+            if const_expr(not self.fp8_slow_accum):
+                warpgroup.wait_group(k_pipe_mmas)
+            else:
+                warpgroup.wait_group(0)
+                acc_slow.store(acc_slow.load() + acc.load())
+            ab_pipeline.consumer_release(ab_release_state)
+            ab_read_state.advance()
+            ab_release_state.advance()
+            peek_ab_full_status = Boolean(True)
+            if k_tile + 1 < k_tile_cnt:
+                peek_ab_full_status = ab_pipeline.consumer_try_wait(ab_read_state)
+        if const_expr(self.pingpong):
+            # Cue for next WG's MMA to start
+            self.pingpong_barrier_arrive(1 - warp_group_idx, stage="mma")
+        if const_expr(not self.fp8_slow_accum):
+            # fp8_slow_accum would already called wait_group(0) inside the loop
+            warpgroup.wait_group(0)
+        for k_tile in cutlass.range(num_prologue_mma, unroll=1):
+            ab_pipeline.consumer_release(ab_release_state)
+            ab_release_state.advance()
+        if const_expr(self.fp8_slow_accum):
+            acc.store(acc_slow.load())
+        # If we don't return the tiled_mma, we get compiler error
+        # "operand #0 does not dominate this use"
+        return ab_read_state, tiled_mma
+
+    @cute.jit
+    def epilogue(
+        self,
+        params: EpilogueParams,
+        epi_smem_tensors: Tuple[cute.Tensor, ...],
+        epi_pipeline: cutlass.pipeline.PipelineAsync,
+        epi_read_state: cutlass.pipeline.PipelineState,
+        epi_producer_state: cutlass.pipeline.PipelineState,
+        tiled_mma: cute.TiledMma,
+        tRS_rAcc: cute.Tensor,
+        tRS_rD: cute.Tensor,
+        tRS_rC: Optional[cute.Tensor],
+        tiled_copy_r2s: cute.core.ThrCopy,
+        tRS_sD: cute.Tensor,
+        tiled_copy_s2r: Optional[cute.core.ThrCopy],
+        tSR_rC: Optional[cute.Tensor],
+        tSR_sC: Optional[cute.Tensor],
+        copy_D: Optional[Callable],
+        bSG_sD: cute.Tensor,
+        bSG_gD: cute.Tensor,
+        epi_load_g2s: Optional[Callable],
+        tile_coord_mnkl: cute.Coord,
+        cu_seqlens_m: Optional[cute.Tensor],
+        epilogue_barrier: cutlass.pipeline.NamedBarrier,
+        tile_scheduler,
+        tidx: Int32,
+        is_tma_warp: Boolean,
+    ) -> Tuple[cutlass.pipeline.PipelineState, cutlass.pipeline.PipelineState]:
+        has_C = const_expr(tRS_rC is not None)
+        has_D = const_expr(copy_D is not None)
+        # We iterate over epi tiles in the N dimension first before the M dimension
+        epi_tile_shape = cute.zipped_divide(
+            cute.make_layout(self.tile_shape_mnk[:2]), self.epi_tile
+        ).shape[1]
+        epi_tile_layout = cute.make_layout(epi_tile_shape, stride=(epi_tile_shape[1], 1))
+        epi_tile_num = cute.size(epi_tile_shape)
+        num_prev_subtiles = tile_scheduler.num_tiles_executed * epi_tile_num
+
+        if const_expr(epi_load_g2s is not None):
+            for epi_idx in cutlass.range(min(epi_tile_num, self.epi_c_stage), unroll=1):
+                epi_producer_state = epi_load_g2s(epi_producer_state, epi_idx, is_tma_warp)
+
+        for epi_idx in cutlass.range_constexpr(epi_tile_num):
+            # Copy from acc to D registers
+            for epi_v in cutlass.range_constexpr(cute.size(tRS_rD)):
+                tRS_rD[epi_v] = tRS_rAcc[epi_idx * cute.size(tRS_rD) + epi_v]
+            if const_expr(has_C):
+                epi_pipeline.consumer_wait(epi_read_state)
+                cute.copy(tiled_copy_s2r, tSR_sC[None, None, None, epi_read_state.index], tSR_rC)
+                # Fence to make sure shared memory read is visible to TMA load
                 cute.arch.fence_proxy(
                     cute.arch.ProxyKind.async_shared, space=cute.arch.SharedSpace.shared_cta
                 )
-                # barrier for sync
-                cute.arch.barrier(barrier_id=1, number_of_threads=self.num_mma_threads)
-                # Get the global memory coordinate for the current epi tile.
-                epi_tile_layout = cute.make_layout(epi_tile_shape, stride=(epi_tile_shape[1], 1))
-                gmem_coord = epi_tile_layout.get_hier_coord(epi_idx)
-                # Copy from shared memory to global memory
-                if warp_idx == 0:
-                    cute.copy(tma_atom_d, bSG_sD[(None, epi_buffer)], bSG_gD[(None, gmem_coord)])
-                    cute.arch.cp_async_bulk_commit_group()
-                    # TODO: when moving to persistent maybe we always need this wait_group
-                    if epi_idx >= self.epi_stage - 1:
-                        cute.arch.cp_async_bulk_wait_group(self.epi_stage - 1, read=True)
-                if epi_idx >= self.epi_stage - 1:
-                    cute.arch.barrier(barrier_id=1, number_of_threads=self.num_mma_threads)
-
-            if warp_idx == 0:
-                cute.arch.cp_async_bulk_wait_group(0, read=True)
+                cute.arch.sync_warp()
+                with cute.arch.elect_one():
+                    epi_pipeline.consumer_release(epi_read_state)
+                epi_read_state.advance()
+            if const_expr(epi_load_g2s is not None and epi_idx + self.epi_c_stage < epi_tile_num):
+                epi_producer_state = epi_load_g2s(
+                    epi_producer_state, epi_idx + self.epi_c_stage, is_tma_warp
+                )
+            tRS_rEpi = self.epi_visit_acc_subtile(params, tRS_rD, tRS_rC)
+            epi_buffer = (num_prev_subtiles + epi_idx) % self.epi_stage
+            # Copy from D registers to shared memory
+            if const_expr(has_D):
+                # Type conversion
+                tRS_rD_out = cute.make_fragment_like(tRS_rD, self.d_dtype)
+                tRS_rD_out.store(tRS_rD.load().to(self.d_dtype))
+                cute.copy(tiled_copy_r2s, tRS_rD_out, tRS_sD[None, None, None, epi_buffer])
+            # Fence and barrier to make sure shared memory store is visible to TMA store
+            cute.arch.fence_proxy(
+                cute.arch.ProxyKind.async_shared, space=cute.arch.SharedSpace.shared_cta
+            )
+            epilogue_barrier.arrive_and_wait()
+            # Get the global memory coordinate for the current epi tile
+            gmem_coord = epi_tile_layout.get_hier_coord(epi_idx)
+            # Copy from shared memory to global memory
+            if is_tma_warp:
+                if const_expr(has_D):
+                    copy_D(bSG_sD[None, epi_buffer], bSG_gD[None, gmem_coord])
+                cute.arch.cp_async_bulk_commit_group()
+                cute.arch.cp_async_bulk_wait_group(self.epi_stage - 1, read=True)
+            epilogue_barrier.arrive_and_wait()
+
+        return epi_read_state, epi_producer_state
+
+    @cute.jit
+    def epi_load_g2s(
+        self,
+        epi_pipeline: cutlass.pipeline.PipelineAsync,
+        copy_C: Callable,
+        bGS_gC: cute.Tensor,
+        bGS_sC: cute.Tensor,
+        epi_producer_state: cutlass.pipeline.PipelineState,
+        epi_idx: Int32,
+        should_load: Boolean,
+    ) -> cutlass.pipeline.PipelineState:
+        # We iterate over epi tiles in the N dimension first before the M dimension
+        epi_tile_layout = cute.make_layout(bGS_gC.shape[1], stride=(bGS_gC.shape[1][1], 1))
+        if should_load:
+            epi_pipeline.producer_acquire(epi_producer_state)
+            # Get the global memory coordinate for the current epi tile
+            gmem_coord = epi_tile_layout.get_hier_coord(epi_idx)
+            copy_C(
+                bGS_gC[None, gmem_coord],
+                bGS_sC[None, epi_producer_state.index],
+                tma_bar_ptr=epi_pipeline.producer_get_barrier(epi_producer_state),
+            )
+            # Epi pipeline's producer commit is a NOP
+            epi_pipeline.producer_commit(epi_producer_state)
+        epi_producer_state.advance()
+        return epi_producer_state
+
+    def epi_visit_acc_subtile(
+        self,
+        params: EpilogueParams,
+        tRS_rD: cute.Tensor,
+        tRS_rC: Optional[cute.Tensor] = None,
+    ) -> Optional[cute.Tensor]:
+        # Apply alpha scaling to accumulator if alpha is provided (not None)
+        if const_expr(hasattr(params, "alpha") and params.alpha is not None):
+            alpha = utils.load_scalar_or_pointer(params.alpha)
+            tRS_rD.store(tRS_rD.load() * alpha)
+        # Apply C with beta scaling
+        if const_expr(tRS_rC is not None):
+            if const_expr(not hasattr(params, "beta") or params.beta is None):
+                # beta is None, default behavior: add C (beta=1.0)
+                tRS_rD.store(tRS_rD.load() + tRS_rC.load().to(tRS_rD.element_type))
+            else:
+                beta = utils.load_scalar_or_pointer(params.beta)
+                tRS_rD.store(tRS_rD.load() + beta * tRS_rC.load().to(tRS_rD.element_type))
+        return None
+
+    def get_scheduler_class(self):
+        """Return the scheduler class to use. Override in subclasses for custom schedulers."""
+        return TileScheduler
+
+    def get_scheduler_arguments(self, problem_shape_ntile_mnl, scheduler_args):
+        """Create scheduler arguments. Override in subclasses for custom schedulers."""
+        return TileSchedulerArguments(
+            problem_shape_ntile_mnl=problem_shape_ntile_mnl,
+            raster_order=scheduler_args.raster_order,
+            group_size=scheduler_args.max_swizzle_size,
+            cluster_shape_mnk=self.cluster_shape_mnk,
+            tile_count_semaphore=scheduler_args.tile_count_semaphore,
+            batch_idx_permute=scheduler_args.batch_idx_permute,
+            is_persistent=self.is_persistent,
+        )
+
+    def epi_visit_acc(
+        self,
+        params: EpilogueParams,
+        acc: cute.Tensor,
+        tiled_mma: cute.TiledMma,
+        tile_coord_mnkl: cute.Coord,
+        tidx: Int32,
+    ) -> None:
+        pass
+
+    def epi_to_underlying_arguments(
+        self, args: EpilogueArguments, *, loc=None, ip=None
+    ) -> EpilogueParams:
+        return GemmSm90.EpilogueParams(alpha=args.alpha, beta=args.beta)
 
     @staticmethod
+    def epi_smem_bytes_per_stage(
+        args: Optional[EpilogueArguments],
+        tile_shape_mnk: Tuple[int, int, int],
+        epi_tile: Tuple[int, int],
+    ) -> int:
+        return 0
+
+    def epi_get_smem_struct(self, params: EpilogueParams):
+        return cute.struct.MemRange[cutlass.Int32, 0]  # Dummy struct
+
+    def epi_get_smem_tensors(self, params: EpilogueParams, storage) -> Tuple[cute.Tensor, ...]:
+        return tuple()
+
+    def pingpong_barrier_sync(self, warp_group_idx: Int32, stage: str):
+        assert stage in ["mma", "epi"]
+        barrier = NamedBarrierGemm.MmaWG0 if stage == "mma" else NamedBarrierGemm.EpiWG0
+        cute.arch.barrier(
+            barrier_id=int(barrier) + warp_group_idx,
+            number_of_threads=2 * self.num_threads_per_warp_group,
+        )
+
+    def pingpong_barrier_arrive(self, warp_group_idx: Int32, stage: str):
+        assert stage in ["mma", "epi"]
+        barrier = NamedBarrierGemm.MmaWG0 if stage == "mma" else NamedBarrierGemm.EpiWG0
+        cute.arch.barrier_arrive(
+            barrier_id=int(barrier) + warp_group_idx,
+            number_of_threads=2 * self.num_threads_per_warp_group,
+        )
+
+    def epilog_smem_copy_atom(self, tiled_mma: cute.TiledMma) -> cute.TiledCopy:
+        copy_atom_C = cute.make_copy_atom(
+            warp.StMatrix8x8x16bOp(
+                self.d_layout.is_m_major_c() if self.d_layout is not None else False,
+                num_matrices=4 if self.epi_tile[1] % 16 == 0 else 2,
+            ),
+            cutlass.Float16,  # this is just to get the right source layout
+        )
+        tiled_copy_C_atom = cute.make_tiled_copy_C_atom(copy_atom_C, tiled_mma)
+        return tiled_copy_C_atom
+
+    def epilog_smem_store_and_partition(
+        self,
+        tiled_mma: cute.TiledMma,
+        d_layout: Optional[LayoutEnum],
+        dtype: Type[cutlass.Numeric],
+        acc: cute.Tensor,
+        sD: cute.Tensor,
+        tidx: Int32,
+    ) -> Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]:
+        if d_layout is None:
+            d_layout = LayoutEnum.ROW_MAJOR
+        tiled_copy_C_atom = self.epilog_smem_copy_atom(tiled_mma)
+        # Doesn't work with tile_N % 8 == 0 but tile_n % 16 != since this always
+        # get st.matrix with num_matrices=4
+        copy_atom_r2s = sm90_utils.sm90_get_smem_store_op(
+            d_layout, elem_ty_d=dtype, elem_ty_acc=self.acc_dtype
+        )
+        tiled_copy_r2s = cute.make_tiled_copy_S(copy_atom_r2s, tiled_copy_C_atom)
+        # (R2S, R2S_M, R2S_N, PIPE_D)
+        thr_copy_r2s = tiled_copy_r2s.get_slice(tidx)
+        tRS_sD = thr_copy_r2s.partition_D(sD) if sD is not None else None
+        # (R2S, R2S_M, R2S_N)
+        tRS_rAcc = tiled_copy_r2s.retile(acc)
+        sD_shape = sD.shape[:2] if sD is not None else self.epi_tile
+        tRS_rD_shape = thr_copy_r2s.partition_S(cute.make_identity_tensor(sD_shape)).shape
+        tRS_rD = cute.make_fragment(tRS_rD_shape, self.acc_dtype)
+        return tiled_copy_r2s, tRS_rAcc, tRS_rD, tRS_sD
+
+    def epilog_smem_load_and_partition(
+        self,
+        tiled_mma: cute.TiledMma,
+        c_layout: LayoutEnum,
+        dtype: Type[cutlass.Numeric],
+        sC: cute.Tensor,
+        tRS_rD_layout: cutlass.Layout,
+        tidx: Int32,
+    ) -> Tuple[cute.TiledCopy, cute.Tensor, cute.Tensor]:
+        tiled_copy_C_atom = self.epilog_smem_copy_atom(tiled_mma)
+        copy_atom_s2r = utils.sm90_get_smem_load_op(c_layout, dtype)
+        tiled_copy_s2r = cute.make_tiled_copy_S(copy_atom_s2r, tiled_copy_C_atom)
+        thr_copy_s2r = tiled_copy_s2r.get_slice(tidx)
+        tSR_sC = thr_copy_s2r.partition_S(sC)
+        tRS_rC = cute.make_fragment(tRS_rD_layout, dtype)
+        tSR_rC = thr_copy_s2r.retile(tRS_rC)
+        return tiled_copy_s2r, tRS_rC, tSR_rC, tSR_sC
+
+    def epilog_gmem_copy_and_partition(
+        self,
+        atom: Union[cute.CopyAtom, cute.TiledCopy],
+        mD_mnl: cute.Tensor,
+        tile_shape_mn: cute.Tile,
+        epi_tile: cute.Tile,
+        sD: cute.Tensor,
+        tile_coord_mnkl: cute.Coord,
+        cu_seqlens_m: Optional[cute.Tensor] = None,
+    ) -> Tuple[cute.Tensor, cute.Tensor]:
+        batch_idx = tile_coord_mnkl[3]
+        if const_expr(cu_seqlens_m is not None):
+            mD_mn = cute.domain_offset((cu_seqlens_m[batch_idx], 0), mD_mnl)
+        else:
+            mD_mn = mD_mnl[None, None, batch_idx]
+        # (bM, bN)
+        gD = cute.local_tile(mD_mn, tile_shape_mn, tile_coord_mnkl[:2])
+        tDgD_for_tma_partition = cute.zipped_divide(gD, epi_tile)
+        bSG_sD, bSG_gD = cpasync.tma_partition(
+            atom,
+            0,
+            cute.make_layout(1),
+            cute.group_modes(sD, 0, 2),
+            tDgD_for_tma_partition,
+        )
+        return bSG_sD, bSG_gD
+
+    def make_ab_pipeline(
+        self,
+        a_smem_layout: cute.Layout | cute.ComposedLayout,
+        b_smem_layout: cute.Layout | cute.ComposedLayout,
+        tiled_mma: cute.TiledMma,
+        cluster_layout_vmnk: cute.Layout,
+        ab_pipeline_mbar_ptr: cute.Pointer,
+    ):
+        # Threads/warps participating in this pipeline
+        producer_cnt = 1 if const_expr(not self.gather_A) else 1 + self.num_ab_load_threads
+        ab_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread, producer_cnt)
+        # Each warp will contribute to the arrive count with the number of mcast size
+        mcast_size = self.num_mcast_ctas_a + self.num_mcast_ctas_b - 1
+        consumer_arrive_cnt = mcast_size * (tiled_mma.size // cute.arch.WARP_SIZE)
+        ab_pipeline_consumer_group = pipeline.CooperativeGroup(
+            pipeline.Agent.Thread, consumer_arrive_cnt
+        )
+        pipeline_cls = pipeline.PipelineTmaAsync if not self.gather_A else PipelineTmaCpAsync
+        tma_copy_bytes = cute.size_in_bytes(self.b_dtype, b_smem_layout)
+        if const_expr(not self.gather_A):
+            tma_copy_bytes += cute.size_in_bytes(self.a_dtype, a_smem_layout)
+        return pipeline_cls.create(
+            barrier_storage=ab_pipeline_mbar_ptr,
+            num_stages=self.ab_stage,
+            producer_group=ab_pipeline_producer_group,
+            consumer_group=ab_pipeline_consumer_group,
+            tx_count=tma_copy_bytes,
+            cta_layout_vmnk=cluster_layout_vmnk,
+        )
+
+    def make_epi_pipeline(
+        self, c_smem_layout: cute.Layout | cute.ComposedLayout, epi_pipeline_mbar_ptr: cute.Pointer
+    ):
+        # Threads/warps participating in this pipeline
+        epi_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
+        # Each warp will contribute 1 to the arrive count
+        consumer_arrive_cnt = self.num_epi_threads // cute.arch.WARP_SIZE
+        epi_pipeline_consumer_group = pipeline.CooperativeGroup(
+            pipeline.Agent.Thread, consumer_arrive_cnt
+        )
+        tma_copy_c_bytes = cute.size_in_bytes(self.c_dtype, c_smem_layout)
+        return pipeline.PipelineTmaAsync.create(
+            barrier_storage=epi_pipeline_mbar_ptr,
+            num_stages=self.epi_c_stage,
+            producer_group=epi_pipeline_producer_group,
+            consumer_group=epi_pipeline_consumer_group,
+            tx_count=tma_copy_c_bytes,
+        )
+
+    def make_sched_pipeline(
+        self, cluster_layout_mnk: cute.Layout, sched_pipeline_mbar_ptr: cute.Pointer, varlen_k: bool
+    ):
+        # Threads/warps participating in this pipeline
+        sched_pipeline_producer_group = pipeline.CooperativeGroup(pipeline.Agent.Thread)
+        cluster_size = cute.size(cluster_layout_mnk)
+        # Each warp that are not the scheduler warp will contribute 1 to the arrive count
+        # If pingpong and varlen_k, then all 8 mma warps will participate in the scheduler barrier
+        # at each round. If pingpong and not varlen_k, then only 4 mma warp will participate.
+        consumer_arrive_cnt = (
+            (self.mma_warp_groups if not (self.pingpong and not varlen_k) else 1) * 4
+            + self.num_ab_load_warps
+        ) * cluster_size - 1
+        sched_pipeline_consumer_group = pipeline.CooperativeGroup(
+            pipeline.Agent.Thread, consumer_arrive_cnt
+        )
+        return pipeline.PipelineAsync.create(
+            barrier_storage=sched_pipeline_mbar_ptr,
+            num_stages=self.sched_stage,
+            producer_group=sched_pipeline_producer_group,
+            consumer_group=sched_pipeline_consumer_group,
+            # If there's cluster, the consumers must arrive at the mbar of CTA 0 in the cluster.
+            consumer_mask=None if const_expr(cluster_size == 1) else 0,
+        )
+
+    @classmethod
     def _compute_stages(
+        cls,
         tile_shape_mnk: Tuple[int, int, int],
+        epi_tile: Tuple[int, int],
         a_dtype: Type[cutlass.Numeric],
         b_dtype: Type[cutlass.Numeric],
+        d_dtype: Optional[Type[cutlass.Numeric]],
+        c_dtype: Optional[Type[cutlass.Numeric]],
+        epilogue_args: Optional[EpilogueArguments],
         smem_capacity: int,
         occupancy: int,
+        overlap_sD_sA: bool,
     ) -> Tuple[int, int]:
         """Computes the number of stages for A/B/C operands based on heuristics.
 
@@ -866,10 +1795,20 @@ class HopperWgmmaGemmKernel:
         :rtype: Tuple[int, int]
         """
 
-        # epi_stage = 4 if tile_shape_mnk[1] % 32 == 0 else 8
-        epi_stage = 4
-        # epi_smem will reuse smem ab.
-        epi_bytes = 0
+        epi_stage = 4 if epi_tile[1] <= 16 else 2
+        if overlap_sD_sA:
+            epi_bytes = 0
+        else:
+            d_bytes_per_stage = (
+                cute.size(epi_tile) * d_dtype.width // 8 if d_dtype is not None else 0
+            )
+            epi_bytes_per_stage = d_bytes_per_stage + cls.epi_smem_bytes_per_stage(
+                epilogue_args, tile_shape_mnk, epi_tile
+            )
+            epi_bytes = epi_bytes_per_stage * epi_stage
+        epi_c_stage = 0 if c_dtype is None else (4 if epi_tile[1] <= 16 else 2)
+        if c_dtype is not None:
+            epi_bytes += cute.size(epi_tile) * c_dtype.width // 8 * epi_c_stage
 
         a_shape = cute.slice_(tile_shape_mnk, (None, 0, None))
         b_shape = cute.slice_(tile_shape_mnk, (0, None, None))
@@ -878,16 +1817,21 @@ class HopperWgmmaGemmKernel:
         )
         mbar_helpers_bytes = 1024
 
-        ab_stage = (
-            (smem_capacity - occupancy * 1024) // occupancy - mbar_helpers_bytes - epi_bytes
-        ) // ab_bytes_per_stage
-        return ab_stage, epi_stage
+        remaining_bytes = smem_capacity // occupancy - mbar_helpers_bytes - epi_bytes
+        ab_stage = remaining_bytes // ab_bytes_per_stage
+
+        # Refine epilogue stages:
+        # Calculate remaining smem after allocating for A/B stages and reserved bytes
+        # Add remaining unused smem to epilogue
+        if not overlap_sD_sA and epi_bytes_per_stage > 0:
+            epi_stage += (remaining_bytes - ab_bytes_per_stage * ab_stage) // epi_bytes_per_stage
+        return ab_stage, epi_stage, epi_c_stage
 
     @staticmethod
     def _sm90_compute_tile_shape_or_override(
         tile_shape_mnk: Tuple[int, int, int],
-        element_type: Type[cutlass.Numeric],
-        is_cooperative: bool = False,
+        atom_layout_mnk: Tuple[int, int, int],
+        element_type: Optional[Type[cutlass.Numeric]] = None,
         epi_tile_override: Tuple[int, int] | None = None,
     ) -> Tuple[int, int]:
         """Compute the epilogue tile shape or use override if provided.
@@ -906,33 +1850,42 @@ class HopperWgmmaGemmKernel:
         """
         if epi_tile_override is not None:
             return epi_tile_override
-        if is_cooperative:
-            if cute.size(tile_shape_mnk, mode=[0]) == 192:
-                tile_m = 192
-                tile_n = math.gcd(32, cute.size(tile_shape_mnk, mode=[1]) // 2)
-            else:
-                tile_m = math.gcd(128, cute.size(tile_shape_mnk, mode=[0]))
-                tile_n = math.gcd(32, cute.size(tile_shape_mnk, mode=[1]))
-            return (tile_m, tile_n)
+        if tile_shape_mnk[0] % 128 == 0 and atom_layout_mnk[0] > 1:
+            tile_m = math.gcd(128, cute.size(tile_shape_mnk, mode=[0]))
+            tile_n = math.gcd(32, cute.size(tile_shape_mnk, mode=[1]))
+        elif tile_shape_mnk[0] % 192 == 0 and atom_layout_mnk[0] > 1:
+            tile_m = math.gcd(192, cute.size(tile_shape_mnk, mode=[0]))
+            tile_n = math.gcd(32, cute.size(tile_shape_mnk, mode=[1]))
         else:
-            n_perf = 64 if element_type.width == 8 else 32
+            # In the case of tile shape 128 x N but atom_layout 1 x 2, we need to set
+            # epi_tile_m = 64. If epi_tile_m = 128, the epilogue would iterate along the
+            # M dimension first, then move to the N dimension. But the accumulator in registers
+            # iterate along the N dimension first, then move to the M dimension.
+            # We could change the epilogue to accommodate this,
+            # but it's easier to just set epi_tile_m = 64.
+            n_perf = 64 if element_type is not None and element_type.width == 8 else 32
             tile_m = math.gcd(64, cute.size(tile_shape_mnk, mode=[0]))
             tile_n = math.gcd(n_perf, cute.size(tile_shape_mnk, mode=[1]))
-            return (tile_m, tile_n)
+        return (tile_m, tile_n)
 
     @staticmethod
     def _make_smem_layouts(
         tile_shape_mnk: Tuple[int, int, int],
         epi_tile: Tuple[int, int],
         a_dtype: Type[cutlass.Numeric],
-        a_layout: utils.LayoutEnum,
+        a_layout: LayoutEnum,
         b_dtype: Type[cutlass.Numeric],
-        b_layout: utils.LayoutEnum,
+        b_layout: LayoutEnum,
         ab_stage: int,
-        d_dtype: Type[cutlass.Numeric],
-        d_layout: utils.LayoutEnum,
+        d_dtype: Optional[Type[cutlass.Numeric]],
+        d_layout: LayoutEnum,
         epi_stage: int,
-    ) -> Tuple[cute.ComposedLayout, cute.ComposedLayout, cute.ComposedLayout]:
+        c_dtype: Optional[Type[cutlass.Numeric]],
+        c_layout: Optional[LayoutEnum],
+        epi_c_stage: int,
+    ) -> Tuple[
+        cute.ComposedLayout, cute.ComposedLayout, cute.ComposedLayout, Optional[cute.ComposedLayout]
+    ]:
         """Create shared memory layouts for A, B, and C tensors.
 
         :param tile_shape_mnk: CTA tile shape (M,N,K)
@@ -942,17 +1895,17 @@ class HopperWgmmaGemmKernel:
         :param a_dtype: Data type for matrix A
         :type a_dtype: type[cutlass.Numeric]
         :param a_layout: Layout enum for matrix A
-        :type a_layout: utils.LayoutEnum
+        :type a_layout: LayoutEnum
         :param b_dtype: Data type for matrix B
         :type b_dtype: type[cutlass.Numeric]
         :param b_layout: Layout enum for matrix B
-        :type b_layout: utils.LayoutEnum
+        :type b_layout: LayoutEnum
         :param ab_stage: Number of stages for A/B tensors
         :type ab_stage: int
-        :param d_dtype: Data type for output matrix C
+        :param d_dtype: Data type for output matrix D
         :type d_dtype: type[cutlass.Numeric]
         :param d_layout: Layout enum for the output matrix C
-        :type d_layout: utils.LayoutEnum
+        :type d_layout: LayoutEnum
         :param epi_stage: Number of epilogue stages
         :type epi_stage: int
 
@@ -965,11 +1918,7 @@ class HopperWgmmaGemmKernel:
         b_is_k_major = b_layout.sm90_mma_major_mode() == warpgroup.OperandMajorMode.K
         a_major_mode_size = tile_shape_mnk[2 if a_is_k_major else 0]
         a_smem_layout_atom = warpgroup.make_smem_layout_atom(
-            sm90_utils.get_smem_layout_atom(
-                a_layout,
-                a_dtype,
-                a_major_mode_size,
-            ),
+            sm90_utils.get_smem_layout_atom(a_layout, a_dtype, a_major_mode_size),
             a_dtype,
         )
         a_smem_layout_staged = cute.tile_to_shape(
@@ -982,11 +1931,7 @@ class HopperWgmmaGemmKernel:
 
         b_major_mode_size = tile_shape_mnk[2 if b_is_k_major else 1]
         b_smem_layout_atom = warpgroup.make_smem_layout_atom(
-            sm90_utils.get_smem_layout_atom(
-                b_layout,
-                b_dtype,
-                b_major_mode_size,
-            ),
+            sm90_utils.get_smem_layout_atom(b_layout, b_dtype, b_major_mode_size),
             b_dtype,
         )
         b_smem_layout_staged = cute.tile_to_shape(
@@ -995,56 +1940,52 @@ class HopperWgmmaGemmKernel:
             order=(0, 1, 2) if b_is_k_major else (1, 0, 2),
         )
 
-        d_smem_shape = epi_tile
-        d_major_mode_size = epi_tile[1] if d_layout.is_n_major_c() else epi_tile[0]
-        d_smem_layout_atom = warpgroup.make_smem_layout_atom(
-            sm90_utils.get_smem_layout_atom(
-                d_layout,
+        if d_dtype is not None:
+            d_smem_shape = epi_tile
+            d_major_mode_size = epi_tile[1] if d_layout.is_n_major_c() else epi_tile[0]
+            d_smem_layout_atom = warpgroup.make_smem_layout_atom(
+                sm90_utils.get_smem_layout_atom(d_layout, d_dtype, d_major_mode_size),
                 d_dtype,
-                d_major_mode_size,
-            ),
-            d_dtype,
-        )
-        epi_smem_layout_staged = cute.tile_to_shape(
-            d_smem_layout_atom,
-            cute.append(d_smem_shape, epi_stage),
-            order=(1, 0, 2) if d_layout.is_m_major_c() else (0, 1, 2),
-        )
-
-        return a_smem_layout_staged, b_smem_layout_staged, epi_smem_layout_staged
-
-    @staticmethod
-    def _compute_grid(
-        d: cute.Tensor,
-        tile_shape_mnk: Tuple[int, int, int],
-        cluster_shape_mnk: Tuple[int, int, int],
-    ) -> Tuple[int, int, int]:
-        """Compute grid shape for the output tensor C.
-
-        :param d: The output tensor C
-        :type d: cute.Tensor
-        :param tile_shape_mnk: The shape (M, N, K) of the CTA tile.
-        :type tile_shape_mnk: Tuple[int, int, int]
-        :param cluster_shape_mnk: Shape of each cluster in M, N, K dimensions.
-        :type cluster_shape_mnk: Tuple[int, int, int]
-
-        :return: Grid shape for kernel launch.
-        :rtype: Tuple[int, int, int]
-        """
+            )
+            epi_smem_layout_staged = cute.tile_to_shape(
+                d_smem_layout_atom,
+                cute.append(d_smem_shape, epi_stage),
+                order=(1, 0, 2) if d_layout.is_m_major_c() else (0, 1, 2),
+            )
+        else:
+            epi_smem_layout_staged = None
+
+        if c_dtype is not None:
+            assert c_layout is not None
+            c_smem_shape = epi_tile
+            c_major_mode_size = epi_tile[1] if c_layout.is_n_major_c() else epi_tile[0]
+            c_smem_layout_atom = warpgroup.make_smem_layout_atom(
+                sm90_utils.get_smem_layout_atom(c_layout, c_dtype, c_major_mode_size),
+                c_dtype,
+            )
+            epi_c_smem_layout_staged = cute.tile_to_shape(
+                c_smem_layout_atom,
+                cute.append(c_smem_shape, epi_c_stage),
+                order=(1, 0, 2) if c_layout.is_m_major_c() else (0, 1, 2),
+            )
+        else:
+            epi_c_smem_layout_staged = None
 
-        c_shape = (tile_shape_mnk[0], tile_shape_mnk[1])
-        gc = cute.zipped_divide(d, tiler=c_shape)
-        clusters = cute.ceil_div(cute.get(gc.layout, mode=[1]).shape, cluster_shape_mnk)
-        grid = tuple(x * y for x, y in zip(clusters, cluster_shape_mnk))
-        return grid
+        return (
+            a_smem_layout_staged,
+            b_smem_layout_staged,
+            epi_smem_layout_staged,
+            epi_c_smem_layout_staged,
+        )
 
     @staticmethod
-    def _make_tma_store_atoms_and_tensors(
+    def _make_tma_epi_atoms_and_tensors(
         tensor_d: cute.Tensor,
         epi_smem_layout_staged: cute.ComposedLayout,
         epi_tile: Tuple[int, int],
+        store_or_load: str,
     ) -> Tuple[cute.CopyAtom, cute.Tensor]:
-        """Create TMA atoms and tensors for C tensor storage.
+        """Create TMA atoms and tensors for storing D or loading C.
 
         :param tensor_d: Output tensor D
         :type tensor_d: cute.Tensor
@@ -1056,15 +1997,17 @@ class HopperWgmmaGemmKernel:
         :return: TMA atom and tensor for C
         :rtype: Tuple[cute.CopyAtom, cute.Tensor]
         """
+        assert store_or_load in ["load", "store"]
         epi_smem_layout = cute.slice_(epi_smem_layout_staged, (None, None, 0))
-        c_cta_v_layout = cute.composition(cute.make_identity_layout(tensor_d.shape), epi_tile)
+        d_cta_v_layout = cute.composition(cute.make_identity_layout(tensor_d.shape), epi_tile)
+        op = (
+            cpasync.CopyBulkTensorTileG2SOp()
+            if store_or_load == "load"
+            else cpasync.CopyBulkTensorTileS2GOp()
+        )
         tma_atom_d, tma_tensor_d = cpasync.make_tiled_tma_atom(
-            cpasync.CopyBulkTensorTileS2GOp(),
-            tensor_d,
-            epi_smem_layout,
-            c_cta_v_layout,
+            op, tensor_d, epi_smem_layout, d_cta_v_layout
         )
-
         return tma_atom_d, tma_tensor_d
 
     @staticmethod
@@ -1104,12 +2047,37 @@ class HopperWgmmaGemmKernel:
         )
         return tma_atom, tma_tensor
 
+    def _make_gmem_tiled_copy_A(self, dtype, major_mode, num_threads, copy_bits=128):
+        atom_async_copy = cute.make_copy_atom(
+            cpasync.CopyG2SOp(cache_mode=cpasync.LoadCacheMode.GLOBAL),
+            dtype,
+            num_bits_per_copy=copy_bits,
+        )
+        copy_elems = copy_bits // dtype.width
+        shape_dim_1 = cute.size(self.tile_shape_mnk[2]) // copy_elems
+        # thread layout for copy
+        thread_layout = cute.make_layout(
+            (num_threads // shape_dim_1, shape_dim_1), stride=(shape_dim_1, 1)
+        )
+        if major_mode != LayoutEnum.ROW_MAJOR:
+            shape_dim_0 = cute.size(self.tile_shape_mnk[0]) // copy_elems
+            thread_layout = cute.make_layout(
+                (shape_dim_0, num_threads // shape_dim_0), stride=(1, shape_dim_0)
+            )
+        # Value layout for copy
+        value_layout = (
+            cute.make_layout((1, copy_elems))
+            if major_mode == LayoutEnum.ROW_MAJOR
+            else cute.make_layout((copy_elems, 1))
+        )
+        return cute.make_tiled_copy_tv(atom_async_copy, thread_layout, value_layout)
+
     @staticmethod
     def is_valid_dtypes(
         a_dtype: Type[cutlass.Numeric],
         b_dtype: Type[cutlass.Numeric],
         acc_dtype: Type[cutlass.Numeric],
-        d_dtype: Type[cutlass.Numeric],
+        d_dtype: Optional[Type[cutlass.Numeric]],
         a_major: str,
         b_major: str,
     ) -> bool:
@@ -1133,7 +2101,6 @@ class HopperWgmmaGemmKernel:
         :rtype: bool
         """
         is_valid = True
-        # tested a_dtype
         if a_dtype not in {
             cutlass.Float16,
             cutlass.BFloat16,
@@ -1149,11 +2116,11 @@ class HopperWgmmaGemmKernel:
             cutlass.Float8E5M2,
         }:
             is_valid = False
-        # tested acc_dtype
         if acc_dtype not in {cutlass.Float32, cutlass.Float16}:
             is_valid = False
         # tested d_dtype
         if d_dtype not in {
+            None,
             cutlass.Float32,
             cutlass.Float16,
             cutlass.BFloat16,
@@ -1171,260 +2138,108 @@ class HopperWgmmaGemmKernel:
         # for Float8 types, this implementation only supports k-major layout
         if (a_dtype.width == 8 and a_major != "k") or (b_dtype.width == 8 and b_major != "k"):
             is_valid = False
-
         return is_valid
 
 
-def run(
-    mnkl: Tuple[int, int, int, int],
-    a_dtype: Type[cutlass.Numeric],
-    b_dtype: Type[cutlass.Numeric],
-    d_dtype: Type[cutlass.Numeric],
-    acc_dtype: Type[cutlass.Numeric],
-    a_major: str,
-    b_major: str,
-    d_major: str,
-    tile_shape_mnk: Tuple[int, int, int],
-    cluster_shape_mn: Tuple[int, int],
-    tolerance: float,
-    warmup_iterations: int,
-    iterations: int,
-    skip_ref_check: bool,
-    use_cold_l2: bool = False,
-    **kwargs,
-):
-    """
-    Prepare A/B/C tensors, launch GPU kernel, and reference checking.
-
-    :param mnkl: Problem size (M, N, K, L)
-    :type mnkl: Tuple[int, int, int, int]
-    :param a_dtype: Data type for input tensor A
-    :type a_dtype: Type[cutlass.Numeric]
-    :param b_dtype: Data type for input tensor B
-    :type b_dtype: Type[cutlass.Numeric]
-    :param d_dtype: Data type for output tensor C
-    :type d_dtype: Type[cutlass.Numeric]
-    :param acc_dtype: Data type for accumulation during matrix multiplication
-    :type acc_dtype: Type[cutlass.Numeric]
-    :param a_major/b_major/d_major: Memory layout of tensor A/B/C
-    :type a_major/b_major/d_major: str
-    :param tile_shape_mnk: CTA tile shape (M, N, K)
-    :type tile_shape_mnk: Tuple[int, int, int]
-    :param cluster_shape_mn: Cluster shape (M, N)
-    :type cluster_shape_mn: Tuple[int, int]
-    :param tolerance: Tolerance value for reference validation comparison
-    :type tolerance: float
-    :param warmup_iterations: Number of warmup iterations before benchmarking, defaults to 0
-    :type warmup_iterations: int, optional
-    :param iterations: Number of benchmark iterations to run, defaults to 1
-    :type iterations: int, optional
-    :param skip_ref_check: Whether to skip reference result validation, defaults to False
-    :type skip_ref_check: bool, optional
-    :param use_cold_l2: Whether to use circular buffer strategy to ensure cold L2 cache, defaults to False
-    :type use_cold_l2: bool, optional
-    :return: Execution time of the GEMM kernel in microseconds
-    :rtype: float
-    """
-
-    print("Running Hopper Dense GEMM with:")
-    print(f"mnkl: {mnkl}")
-    print(f"A dtype: {a_dtype}, B dtype: {b_dtype}, C dtype: {d_dtype}, Acc dtype: {acc_dtype}")
-    print(f"Matrix majors - A: {a_major}, B: {b_major}, C: {d_major}")
-    print(f"Tile Shape: {tile_shape_mnk}, Cluster Shape: {cluster_shape_mn}")
-    print(f"Tolerance: {tolerance}")
-    print(f"Warmup iterations: {warmup_iterations}")
-    print(f"Iterations: {iterations}")
-    print(f"Skip reference checking: {skip_ref_check}")
-    print(f"Use cold L2: {use_cold_l2}")
-
-    # Unpack parameters
-    m, n, k, l = mnkl
-    cluster_shape_mnk = (*cluster_shape_mn, 1)
-
-    # Skip unsupported types
-    if not HopperWgmmaGemmKernel.is_valid_dtypes(
-        a_dtype, b_dtype, acc_dtype, d_dtype, a_major, b_major
+def gemm_sm90(
+    A: Tensor,  # (l, m, k)
+    B: Tensor,  # (l, n, k)
+    D: Tensor,  # (l, m, n)
+    C: Optional[Tensor],  # (l, m, n)
+    tile_count_semaphore: Optional[Tensor],  # (1,)
+    tile_M: int,
+    tile_N: int,
+    cluster_M: int,
+    cluster_N: int,
+    pingpong: bool = False,
+    persistent: bool = True,
+    alpha: float | Tensor = 1.0,
+    beta: float | Tensor = 1.0,
+) -> None:
+    L, M, K, N, tensor_infos = GemmWrapperBase.validate_and_prepare_tensors(A, B, D, C)
+    GemmWrapperBase.permute_tensors(tensor_infos)
+    GemmWrapperBase.extract_dtypes(tensor_infos)
+    major_configs = {
+        "A": ("m", "k", "l"),
+        "B": ("n", "k", "l"),
+        "D": ("m", "n", "l"),
+        "C": ("m", "n", "l"),
+    }
+    GemmWrapperBase.determine_major_orders(tensor_infos, major_configs)
+
+    acc_dtype = cutlass.Float32
+    tile_shape_mn = (tile_M, tile_N)
+    cluster_shape_mnk = (cluster_M, cluster_N, 1)
+    if not GemmSm90.is_valid_dtypes(
+        tensor_infos["A"].dtype,
+        tensor_infos["B"].dtype,
+        acc_dtype,
+        tensor_infos["D"].dtype,
+        tensor_infos["A"].major,
+        tensor_infos["B"].major,
     ):
-        raise TypeError(
-            f"Skipping due to unsupported combination of types and majors: {a_dtype}, {b_dtype}, {acc_dtype}, {d_dtype}, {a_major=}, {b_major=}"
-        )
+        raise TypeError("Skipping due to unsupported combination of types and majors")
 
-    # Prepare pytorch tensors: A, B (random from 0 to 2) and C (all zero)
-    if not torch.cuda.is_available():
-        raise RuntimeError("GPU is required to run this example!")
-
-    torch.manual_seed(1111)
-
-    # Create and permute tensor A/B/C
-    def create_and_permute_tensor(l, mode0, mode1, is_mode0_major, dtype, is_dynamic_layout=True):
-        # is_mode0_major: (l, mode1, mode0) -> (mode0, mode1, l)
-        # else : (l, mode0, mode1) -> (mode0, mode1, l)
-        shape = (l, mode1, mode0) if is_mode0_major else (l, mode0, mode1)
-        permute_order = (2, 1, 0) if is_mode0_major else (1, 2, 0)
-        is_unsigned = dtype in {cutlass.Uint8}
-        # Temporarily use uint8 as torch does not support fp8 type
-        torch_dtype = (
-            cutlass_torch.dtype(dtype)
-            if dtype not in {cutlass.Float8E5M2, cutlass.Float8E4M3FN}
-            else torch.uint8
-        )
-
-        # Create dtype torch tensor (cpu)
-        torch_tensor_cpu = cutlass.torch.create_and_permute_torch_tensor(
-            shape,
-            torch_dtype,
-            permute_order=permute_order,
-            # init_type=cutlass.torch.TensorInitType.RANDOM,
-            # init_config=cutlass.torch.RandomInitConfig(
-            #     min_val=0 if is_unsigned else -2, max_val=4 if is_unsigned else 2
-            # ),
-            init_type=cutlass.torch.TensorInitType.GAUSSIAN,
-            init_config=cutlass.torch.GaussianInitConfig(std=k ** (-0.5), scale=1),
-        )
-        # Create dtype torch tensor (gpu)
-        torch_tensor = torch_tensor_cpu.cuda()
-
-        # Create f32 torch tensor (cpu)
-        f32_torch_tensor = torch_tensor_cpu.to(dtype=torch.float32)
-
-        # Create dtype cute tensor (gpu)
-        cute_tensor = from_dlpack(torch_tensor, assumed_align=16)
-        cute_tensor.element_type = dtype
-        if is_dynamic_layout:
-            cute_tensor = cute_tensor.mark_layout_dynamic(leading_dim=(0 if is_mode0_major else 1))
-        cute_tensor = cutlass.torch.convert_cute_tensor(
-            f32_torch_tensor,
-            cute_tensor,
-            dtype,
-            is_dynamic_layout=is_dynamic_layout,
-        )
+    max_active_clusters = get_max_active_clusters(cluster_M * cluster_N) if persistent else 0
+    GemmWrapperBase.create_cute_tensors(tensor_infos, major_configs)
 
-        return f32_torch_tensor, cute_tensor, torch_tensor
-
-    a, mA, a_torch = create_and_permute_tensor(l, m, k, a_major == "m", a_dtype)
-    b, mB, b_torch = create_and_permute_tensor(l, n, k, b_major == "n", b_dtype)
-    c, mC, c_torch = create_and_permute_tensor(l, m, n, d_major == "m", d_dtype)
-
-    gemm = HopperWgmmaGemmKernel(acc_dtype, tile_shape_mnk, cluster_shape_mnk)
-
-    torch_stream = torch.cuda.Stream()
-    stream = cuda.CUstream(torch_stream.cuda_stream)
-    # compile gemm kernel
-    compiled_gemm = cute.compile(gemm, mA, mB, mC, stream)
-
-    if not skip_ref_check:
-        # execution
-        compiled_gemm(mA, mB, mC, stream)
-
-        torch.cuda.synchronize()
-
-        # Ref check
-        ref = (torch.einsum("mkl,nkl->mnl", a, b)).cpu()
-
-        if d_dtype in (cutlass.Float8E4M3FN, cutlass.Float8E5M2):
-            # m major: (l, n, m) -> (m, n, l)
-            # n major: (l, m, n) -> (m, n, l)
-            permute_order = (1, 2, 0) if d_major == "n" else (2, 1, 0)
-            shape = (l, m, n) if d_major == "n" else (l, n, m)
-            f8_torch_tensor = cutlass_torch.create_and_permute_torch_tensor(
-                shape,
-                torch.uint8,
-                permute_order=permute_order,
-                init_type=cutlass_torch.TensorInitType.SKIP,
-            ).cuda()
-            # Create dtype cute tensor (gpu)
-            ref_c_tensor = from_dlpack(f8_torch_tensor, assumed_align=16).mark_layout_dynamic(
-                leading_dim=(1 if d_major == "n" else 0)
-            )
-            ref_c_tensor.element_type = d_dtype
-            ref_c_tensor = cutlass_torch.convert_cute_tensor(
-                ref,
-                ref_c_tensor,
-                d_dtype,
-                is_dynamic_layout=True,
-            )
-            ref_c = f8_torch_tensor.cpu()
+    def scalar_arg(scalar: float | Tensor):
+        if isinstance(scalar, float):
+            return Float32(scalar) if scalar != 1.0 else None
         else:
-            ref_c = ref.to(cutlass_torch.dtype(d_dtype))
-
-        torch.testing.assert_close(c_torch.cpu(), ref_c, atol=tolerance, rtol=1e-03)
-
-    def generate_tensors():
-        _, mA_workspace, _ = create_and_permute_tensor(l, m, k, a_major == "m", a_dtype)
-        _, mB_workspace, _ = create_and_permute_tensor(l, n, k, b_major == "n", b_dtype)
-        _, mC_workspace, _ = create_and_permute_tensor(l, m, n, d_major == "m", d_dtype)
-        return testing.JitArguments(mA_workspace, mB_workspace, mC_workspace, stream)
-
-    workspace_count = 1
-    if use_cold_l2:
-        one_workspace_bytes = (
-            a_torch.numel() * a_torch.element_size()
-            + b_torch.numel() * b_torch.element_size()
-            + c_torch.numel() * c_torch.element_size()
+            assert isinstance(scalar, Tensor)
+            return make_ptr(Float32, scalar.data_ptr(), cute.AddressSpace.gmem, assumed_align=4)
+
+    epi_args = GemmSm90.EpilogueArguments(scalar_arg(alpha), scalar_arg(beta))
+    scheduler_args = GemmWrapperBase.create_scheduler_args(
+        max_active_clusters, tile_count_semaphore
+    )
+    current_stream = cutlass_torch.current_stream()
+    compile_key = GemmWrapperBase.get_compile_key(
+        tensor_infos,
+        None,
+        tile_shape_mn,
+        cluster_shape_mnk,
+        pingpong,
+        persistent,
+        tile_count_semaphore is not None,
+        2 if isinstance(alpha, Tensor) else (1 if alpha == 1.0 else 0),
+        2 if isinstance(beta, Tensor) else (1 if beta == 1.0 else 0),
+        key_tensor_names=("A", "B", "D", "C"),
+    )
+    cache = gemm_sm90.compile_cache
+    if compile_key not in cache:
+        gemm = GemmSm90(
+            acc_dtype,
+            tensor_infos["A"].dtype,
+            tile_shape_mn,
+            cluster_shape_mnk,
+            pingpong=pingpong,
+            is_persistent=persistent,
         )
-        workspace_count = testing.get_workspace_count(
-            one_workspace_bytes, warmup_iterations, iterations
+        cache[compile_key] = cute.compile(
+            gemm,
+            tensor_infos["A"].cute_tensor,
+            tensor_infos["B"].cute_tensor,
+            tensor_infos["D"].cute_tensor,
+            tensor_infos["C"].cute_tensor,
+            epi_args,
+            scheduler_args,
+            None,  # varlen_args
+            None,  # mAIdx
+            current_stream,
         )
-
-    exec_time = testing.benchmark(
-        compiled_gemm,
-        workspace_generator=generate_tensors,
-        workspace_count=workspace_count,
-        stream=stream,
-        warmup_iterations=warmup_iterations,
-        iterations=iterations,
+    cache[compile_key](
+        tensor_infos["A"].cute_tensor,
+        tensor_infos["B"].cute_tensor,
+        tensor_infos["D"].cute_tensor,
+        tensor_infos["C"].cute_tensor,
+        epi_args,
+        scheduler_args,
+        None,
+        None,
+        current_stream,
     )
 
-    from triton.testing import do_bench
-
-    current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
-
-    flops = 2 * m * n * k * l
-
-    repeats = 30
-    # repeats = 1
-    warmup = 5
-
-    import time
-
-    time.sleep(0.5)
-    fn = lambda: torch.matmul(a_torch.permute(2, 0, 1), b_torch.permute(2, 0, 1).mT)
-    timing_cublas = do_bench(fn, warmup=warmup, rep=repeats)
-    tflops_cublas = flops / (timing_cublas * 1e9)  # Convert to TFlops
-    print(f"CuBLAS Average time: {timing_cublas:.3f} ms, TFLOPS: {tflops_cublas:.1f}")
-
-    time.sleep(0.5)
-    fn = lambda: compiled_gemm(mA, mB, mC, current_stream)
-    timing = do_bench(fn, warmup=warmup, rep=repeats)
-    tflops = flops / (timing * 1e9)  # Convert to TFlops
-    print(f"Cute-DSL Average time: {timing:.3f} ms, TFLOPS: {tflops:.1f}")
-
-    time.sleep(0.5)
-    fn = lambda: torch.matmul(a_torch.permute(2, 0, 1), b_torch.permute(2, 0, 1).mT)
-    timing_cublas = do_bench(fn, warmup=warmup, rep=repeats)
-    tflops_cublas = flops / (timing_cublas * 1e9)  # Convert to TFlops
-    print(f"CuBLAS Average time: {timing_cublas:.3f} ms, TFLOPS: {tflops_cublas:.1f}")
-
-    return exec_time  # Return execution time in microseconds
-
-
-if __name__ == "__main__":
-    args = parse_arguments()
-    run(
-        args.mnkl,
-        args.a_dtype,
-        args.b_dtype,
-        args.d_dtype,
-        args.acc_dtype,
-        args.a_major,
-        args.b_major,
-        args.d_major,
-        args.tile_shape_mnk,
-        args.cluster_shape_mn,
-        args.tolerance,
-        args.warmup_iterations,
-        args.iterations,
-        args.skip_ref_check,
-        args.use_cold_l2,
-    )
-    print("PASS")
+
+gemm_sm90.compile_cache = {}