quack-kernels 0.1.1__py3-none-any.whl → 0.1.3__py3-none-any.whl

quack/softmax.py

@@ -1,176 +1,186 @@
 import math
 import torch
-import operator
-from typing import Callable
+from typing import Type
 
 import cuda.bindings.driver as cuda
 
 import cutlass
 import cutlass.cute as cute
 from cutlass.cute.runtime import from_dlpack
-import cutlass.torch as cutlass_torch
 
 import quack.utils as utils
+from quack.reduction_base import ReductionBase, torch2cute_dtype_map
 
 
-@cute.kernel
-def softmax_kernel(
-    gX: cute.Tensor,
-    gO: cute.Tensor,
-    cX: cute.Tensor,  # coordinate tensor
-    shape: cute.Shape,
-    tv_layout: cute.Layout,
-    tiler_mn: cute.Shape,
-    cluster_n: cutlass.Constexpr = 1,
-):
-    tidx, _, _ = cute.arch.thread_idx()
-    bidx, cluster_y, _ = cute.arch.block_idx()
-    gdim, _, _ = cute.arch.grid_dim()
-
-    # slice for CTAs
-    # logical id -> address
-    blkX, blkOut, blkCrd = [gT[(None, None), bidx if cluster_n == 1 else (bidx, cluster_y)] for gT in (gX, gO, cX)]
-
-    # declare the atoms which will be used later for memory copy
-    copy_atom_load_X = cute.make_copy_atom(cute.nvgpu.CopyUniversalOp(), gX.element_type, num_bits_per_copy=128)
-    copy_atom_load_X_async = cute.make_copy_atom(cute.nvgpu.cpasync.CopyG2SOp(), gX.element_type, num_bits_per_copy=128)
-    copy_atom_store_O = cute.make_copy_atom(cute.nvgpu.CopyUniversalOp(), gO.element_type, num_bits_per_copy=128)
-
-    thr_copy_X = cute.make_tiled_copy(copy_atom_load_X, tv_layout, tiler_mn).get_slice(tidx)
-    thr_copy_X_async = cute.make_tiled_copy(copy_atom_load_X_async, tv_layout, tiler_mn).get_slice(tidx)
-    thr_copy_O = cute.make_tiled_copy(copy_atom_store_O, tv_layout, tiler_mn).get_slice(tidx)
-
-    smem = cutlass.utils.SmemAllocator()
-    # Don't use blkX.layout here, because the stride is N, not N_rounded
-    sX = smem.allocate_tensor(gX.element_type, cute.make_ordered_layout(blkX.shape, order=(1, 0)), byte_alignment=16)
-    num_warps = cute.size(tv_layout, mode=[0]) // cute.arch.WARP_SIZE
-    warps_per_row = utils.max_constexpr(tv_layout.shape[0][0] // cute.arch.WARP_SIZE, 1)
-
-    reduction_buffer_layout = cute.make_ordered_layout(
+class Softmax(ReductionBase):
+    def __init__(self, dtype: Type[cutlass.Numeric], N: int, online_softmax: bool = True):
         # 2 stages: 1 for max, 1 for sum
-        (num_warps // warps_per_row, warps_per_row if cluster_n == 1 else (warps_per_row, cluster_n), 2),
-        order=(1, 0, 2)
-    )
-    reduction_buffer = smem.allocate_tensor(cutlass.Float32, reduction_buffer_layout, byte_alignment=4)
-    if cutlass.const_expr(cluster_n > 1):
-        # 1 mbar for max reduction, 1 mbar for sum reduction
-        mbar_ptr = smem.allocate_array(cutlass.Int64, num_elems=2)
-    else:
-        mbar_ptr = None
-
-    tXgX = thr_copy_X_async.partition_S(blkX)
-    tXsX = thr_copy_X_async.partition_S(sX)
-    tXgO = thr_copy_O.partition_D(blkOut)
-    tXcX = thr_copy_X.partition_S(blkCrd)[(0, None), None, None]
-
-    # allocate fragments for gmem->rmem
-    tXrX, tXrO = [cute.make_fragment_like(thr) for thr in (tXgX, tXgO)]
-
-    if cluster_n > 1:
-        if tidx < 2:
-            cute.arch.mbarrier_init_arrive_cnt(mbar_ptr + tidx, 1)
-        cute.arch.mbarrier_init_fence()
-        if tidx < 2:
-            cute.arch.mbarrier_init_tx_bytes(mbar_ptr + tidx, num_warps * cluster_n * cutlass.Float32.width // 8)
-        # Cluster arrive after barrier init
-        cute.arch.cluster_arrive_relaxed()
-
-    tXpX = cute.make_fragment_like(tXgX[(0, None), None, None], cutlass.Boolean)
-    for i in range(cute.size(tXpX)):
-        tXpX[i] = cute.elem_less(tXcX[i][1], shape[1])
-
-    if tXcX[0][0] < shape[0]:
-        cute.copy(copy_atom_load_X_async, tXgX, tXsX, pred=tXpX)
-    cute.arch.cp_async_commit_group()
-    cute.arch.cp_async_wait_group(0)
-
-    cute.autovec_copy(tXsX, tXrX)
-    x = tXrX.load().to(cute.Float32)
-    max_x = utils.warp_reduce(
-        x.reduce(cute.ReductionOp.MAX, init_val=float('-inf'), reduction_profile=0),
-        cute.arch.fmax,
-        width=utils.min_constexpr(tv_layout.shape[0][0], cute.arch.WARP_SIZE),
-    )
-    if cutlass.const_expr(cluster_n > 1):
-        cute.arch.cluster_wait()
-    if cutlass.const_expr(warps_per_row > 1 or cluster_n > 1):
-        max_mbar_ptr = mbar_ptr + 0 if cluster_n > 1 else None
-        max_x = utils.block_or_cluster_reduce(
-            max_x, cute.arch.fmax, reduction_buffer[None, None, 0], max_mbar_ptr, init_val=-cutlass.Float32.inf
-        )
-    log2_e = math.log2(math.e)
-    exp_x = cute.math.exp2((x - max_x) * log2_e, fastmath=True)
-    denom = utils.warp_reduce(
-        exp_x.reduce(cute.ReductionOp.ADD, init_val=0.0, reduction_profile=0),
-        operator.add,
-        width=utils.min_constexpr(tv_layout.shape[0][0], cute.arch.WARP_SIZE),
-    )
-    if cutlass.const_expr(warps_per_row > 1 or cluster_n > 1):
-        sum_mbar_ptr = mbar_ptr + 1 if cluster_n > 1 else None
-        denom = utils.block_or_cluster_reduce(
-            denom, operator.add, reduction_buffer[None, None, 1], sum_mbar_ptr, init_val=0.0
-        )
-    inv = 1.0 / denom
-    y = exp_x * inv
-
-    tXrO.store(y.to(tXrO.element_type))
-    tOcX = thr_copy_O.partition_S(blkCrd)[(0, None), None, None]
-    tOpO = cute.make_fragment_like(tXgO[(0, None), None, None], cutlass.Boolean)
-    for i in range(cute.size(tOpO)):
-        tOpO[i] = cute.elem_less(tOcX[i][1], shape[1])
-    if tXcX[0][0] < shape[0]:
-        cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tOpO)
-
-
-@cute.jit
-def softmax_interface(
-    mX: cute.Tensor,
-    mOut: cute.Tensor,
-    stream: cuda.CUstream,
-    N: cutlass.Constexpr,
-    copy_bits: cutlass.Constexpr = 128
-):
-    vecsize = copy_bits // mX.element_type.width
-    assert N % vecsize == 0, f"Input N {N} is not divisible by vector size {vecsize}"
-    num_threads = 128 if N <= 16384 else 256
-    num_warps = num_threads // cute.arch.WARP_SIZE
-    assert num_threads % cute.arch.WARP_SIZE == 0
-    threads_per_row = 8 if N <= 64 else (16 if N <= 128 else (32 if N <= 3072 else (64 if N <= 6144 else (128 if N <= 16384 else 256))))
-    if cutlass.const_expr(mX.element_type.width == 16):
-        cluster_n = 1 if N <= 16 * 1024 else (2 if N <= 32 * 1024 else (4 if N <= 64 * 1024 else (8 if N <= 128 * 1024 else 16)))
-    else:  # fp32
-        cluster_n = 1 if N <= 32 * 1024 else (2 if N <= 64 * 1024 else (4 if N <= 128 * 1024 else (8 if N <= 256 * 1024 else 16)))
-
-    num_blocks_N = cute.ceil_div(N // vecsize, threads_per_row * cluster_n)
-    cols_per_block = num_threads // threads_per_row
-    tiler_mn = (cols_per_block, vecsize * num_blocks_N * threads_per_row)  # This rounds up N
-    tv_layout = cute.make_layout(
-        ((threads_per_row, cols_per_block), (vecsize, num_blocks_N)),
-        stride=((vecsize * cols_per_block, 1), (cols_per_block, cols_per_block * vecsize * threads_per_row))
-    )
+        super().__init__(
+            dtype,
+            N,
+            stage=2 if not online_softmax else 1,
+            reduction_dtype=cutlass.Float32 if not online_softmax else cutlass.Int64,
+        )
+        self.online_softmax = online_softmax
 
-    idX = cute.make_identity_tensor(mX.shape)
-    gX, gO, cX = [cute.zipped_divide(mT, tiler_mn) for mT in (mX, mOut, idX)]  # ((TileM,TileN),(RestM,RestN))
-
-    smem_allocated = cute.size_in_bytes(mX.element_type, cute.make_layout(gX.shape[0])) + 2 * num_warps * cluster_n * (cutlass.Float32.width // 8) + 2 * (cutlass.Int64.width // 8)
-    softmax_kernel(gX, gO, cX, mX.shape, tv_layout, tiler_mn, cluster_n).launch(
-        grid=[cute.size(gX, mode=[1, 0]), cluster_n, 1],
-        block=[cute.size(tv_layout, mode=[0]), 1, 1],
-        # Launching with cluster=[1, 1, 1] instead of None slows down the kernel by ~8us
-        cluster=[1, cluster_n, 1] if cluster_n > 1 else None,
-        smem=smem_allocated,
-        stream=stream,
-    )
+    def _calculate_threads_per_row(self):
+        N = self.N
+        return (
+            8
+            if N <= 64
+            else (
+                16
+                if N <= 128
+                else (32 if N <= 3072 else (64 if N <= 6144 else (128 if N <= 16384 else 256)))
+            )
+        )
 
+    def _set_cluster_n(self):
+        N = self.N
+        if cutlass.const_expr(self.dtype.width == 16):
+            cluster_n = (
+                1
+                if N <= 16 * 1024
+                else (
+                    2
+                    if N <= 32 * 1024
+                    else (4 if N <= 64 * 1024 else (8 if N <= 128 * 1024 else 16))
+                )
+            )
+        else:  # fp32
+            cluster_n = (
+                1
+                if N <= 32 * 1024
+                else (
+                    2
+                    if N <= 64 * 1024
+                    else (4 if N <= 128 * 1024 else (8 if N <= 256 * 1024 else 16))
+                )
+            )
+        self.cluster_n = cluster_n
 
-torch2cute_dtype_map = {
-    torch.float16: cutlass.Float16,
-    torch.bfloat16: cutlass.BFloat16,
-    torch.float32: cutlass.Float32,
-}
+    @cute.jit
+    def __call__(
+        self,
+        mX: cute.Tensor,
+        mO: cute.Tensor,
+        stream: cuda.CUstream,
+    ):
+        assert mX.element_type == self.dtype
+        assert mO.element_type == self.dtype
+        self._set_cluster_n()
+        tiler_mn, tv_layout = self._get_tv_layout()
+        num_threads = cute.size(tv_layout, mode=[0])
+        num_warps = num_threads // cute.arch.WARP_SIZE
+        self.kernel(mX, mO, tv_layout, tiler_mn).launch(
+            grid=[cute.ceil_div(mX.shape[0], tiler_mn[0]), self.cluster_n, 1],
+            block=[num_threads, 1, 1],
+            cluster=[1, self.cluster_n, 1] if self.cluster_n > 1 else None,
+            smem=self._smem_size_in_bytes(tiler_mn, num_warps),
+            stream=stream,
+        )
 
+    @cute.kernel
+    def kernel(
+        self,
+        mX: cute.Tensor,
+        mO: cute.Tensor,
+        tv_layout: cute.Layout,
+        tiler_mn: cute.Shape,
+    ):
+        tidx, _, _ = cute.arch.thread_idx()
+        bidx, cluster_y, _ = cute.arch.block_idx()
 
-def softmax(x: torch.Tensor) -> torch.Tensor:
+        shape = mX.shape
+        idX = cute.make_identity_tensor(shape)
+        # slice for CTAs
+        gX, gO, cX = [
+            cute.local_tile(mT, tiler_mn, (bidx, 0 if self.cluster_n == 1 else cluster_y))
+            for mT in (mX, mO, idX)
+        ]
+
+        smem = cutlass.utils.SmemAllocator()
+        sX = smem.allocate_tensor(
+            mX.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+        )
+        reduction_buffer, mbar_ptr = self._allocate_reduction_buffer_and_mbar(smem, tv_layout)
+
+        # declare the atoms which will be used later for memory copy
+        copy_atom_load_X = cute.make_copy_atom(
+            cute.nvgpu.cpasync.CopyG2SOp(), mX.element_type, num_bits_per_copy=128
+        )
+        copy_atom_store_O = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), gO.element_type, num_bits_per_copy=128
+        )
+
+        thr_copy_X = cute.make_tiled_copy(copy_atom_load_X, tv_layout, tiler_mn).get_slice(tidx)
+        thr_copy_O = cute.make_tiled_copy(copy_atom_store_O, tv_layout, tiler_mn).get_slice(tidx)
+
+        tXgX = thr_copy_X.partition_S(gX)
+        tXsX = thr_copy_X.partition_D(sX)
+        tXgO = thr_copy_O.partition_D(gO)
+        tXcX = thr_copy_X.partition_S(cX)[(0, None), None, None]
+
+        # allocate fragments for gmem->rmem
+        tXrX, tXrO = [cute.make_fragment_like(thr) for thr in (tXgX, tXgO)]
+
+        num_warps = cute.size(tv_layout, mode=[0]) // cute.arch.WARP_SIZE
+        self._initialize_cluster(tidx, mbar_ptr, num_warps)
+
+        is_even_N = cutlass.const_expr(shape[1] == tiler_mn[1] * self.cluster_n)
+        tXpX = (
+            utils.predicate_k(thr_copy_X.partition_S(cX), limit=shape[1]) if not is_even_N else None
+        )
+        if tXcX[0][0] < shape[0]:
+            cute.copy(copy_atom_load_X, tXgX, tXsX, pred=tXpX)
+        cute.arch.cp_async_commit_group()
+        cute.arch.cp_async_wait_group(0)
+        # Fill OOB values with -inf
+        if cutlass.const_expr(not is_even_N):
+            utils.fill_oob(tXsX, tXpX, -tXsX.element_type.inf)
+
+        cute.autovec_copy(tXsX, tXrX)
+        x = tXrX.load().to(cute.Float32)
+        threads_per_row = tv_layout.shape[0][0]
+        if cutlass.const_expr(not self.online_softmax):
+            max_x = utils.row_reduce(
+                x,
+                cute.ReductionOp.MAX,
+                threads_per_row,
+                reduction_buffer[None, None, 0],
+                mbar_ptr + 0 if self.cluster_n > 1 else None,
+                init_val=-cutlass.Float32.inf,
+                hook_fn=cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None,
+            )
+            log2_e = math.log2(math.e)
+            exp_x = cute.math.exp2((x - max_x) * log2_e, fastmath=True)
+            denom = utils.row_reduce(
+                exp_x,
+                cute.ReductionOp.ADD,
+                threads_per_row,
+                reduction_buffer[None, None, 1],
+                mbar_ptr + 1 if self.cluster_n > 1 else None,
+                init_val=0.0,
+            )
+        else:
+            max_x, denom, exp_x = utils.online_softmax_reduce(
+                x,
+                threads_per_row,
+                reduction_buffer[None, None, 0],
+                mbar_ptr,
+                hook_fn=cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None,
+                return_exp_x=True,
+            )
+        y = exp_x * (1.0 / denom)
+        tXrO.store(y.to(tXrO.element_type))
+        tOpO = (
+            utils.predicate_k(thr_copy_O.partition_S(cX), limit=shape[1]) if not is_even_N else None
+        )
+        if tXcX[0][0] < shape[0]:
+            cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tOpO)
+
+
+def _softmax_fwd(x: torch.Tensor) -> torch.Tensor:
     """Softmax forward pass.
     Args:
         x: Input tensor of shape (M, N)
@@ -181,22 +191,258 @@ def softmax(x: torch.Tensor) -> torch.Tensor:
     assert x.is_cuda, "Tensor must be on CUDA device"
     assert x.dtype in [torch.float16, torch.bfloat16, torch.float32], "Unsupported dtype"
     M, N = x.shape
-    device = x.device
     out = torch.empty_like(x)
     dtype = torch2cute_dtype_map[x.dtype]
     convert_from_dlpack = lambda tensor: (
-        from_dlpack(tensor.detach(), assumed_align=16)
-        .mark_compact_shape_dynamic(mode=0, stride_order=(0, 1))
+        from_dlpack(tensor.detach(), assumed_align=16).mark_compact_shape_dynamic(
+            mode=0, stride_order=(0, 1)
+        )
     )
     x_tensor, out_tensor = [convert_from_dlpack(tensor) for tensor in (x, out)]
     current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
     compile_key = (dtype, N)
-    if compile_key not in softmax.compile_cache:
-        softmax.compile_cache[compile_key] = cute.compile(
-            softmax_interface, x_tensor, out_tensor, current_stream, N
+    if compile_key not in _softmax_fwd.compile_cache:
+        softmax_op = Softmax(dtype, N)
+        _softmax_fwd.compile_cache[compile_key] = cute.compile(
+            softmax_op, x_tensor, out_tensor, current_stream
         )
-    softmax.compile_cache[compile_key](x_tensor, out_tensor, current_stream)
+    _softmax_fwd.compile_cache[compile_key](x_tensor, out_tensor, current_stream)
     return out
 
 
-softmax.compile_cache = {}
+_softmax_fwd.compile_cache = {}
+
+
+class SoftmaxBackward(ReductionBase):
+    def __init__(self, dtype: Type[cutlass.Numeric], N: int):
+        # 1 stage for computing dot product
+        super().__init__(dtype, N, stage=1, reduction_dtype=cutlass.Float32)
+
+    def _calculate_threads_per_row(self):
+        N = self.N
+        return (
+            8
+            if N <= 64
+            else (
+                16
+                if N <= 128
+                else (32 if N <= 3072 else (64 if N <= 6144 else (128 if N <= 8192 else 256)))
+            )
+        )
+
+    def _set_cluster_n(self):
+        N = self.N
+        if cutlass.const_expr(self.dtype.width == 16):
+            cluster_n = (
+                1
+                if N <= 16 * 1024
+                else (
+                    2
+                    if N <= 32 * 1024
+                    else (4 if N <= 64 * 1024 else (8 if N <= 128 * 1024 else 16))
+                )
+            )
+        else:  # fp32
+            cluster_n = (
+                1
+                if N <= 16 * 1024
+                else (
+                    2
+                    if N <= 32 * 1024
+                    else (4 if N <= 64 * 1024 else (8 if N <= 128 * 1024 else 16))
+                )
+            )
+        self.cluster_n = cluster_n
+
+    def _get_num_threads(self):
+        return 128 if self.N <= 8192 else 256
+
+    def _smem_size_in_bytes(self, tiler_mn, num_warps):
+        return (
+            # Multiply by 2 since we need space for Y and dY
+            cute.size_in_bytes(self.dtype, cute.make_layout(tiler_mn)) * 2
+            + self.stage * num_warps * self.cluster_n * (self.reduction_dtype.width // 8)
+            + self.stage * (cutlass.Int64.width // 8)
+        )
+
+    @cute.jit
+    def __call__(
+        self,
+        mdY: cute.Tensor,
+        mY: cute.Tensor,
+        mdX: cute.Tensor,
+        stream: cuda.CUstream,
+    ):
+        assert mdY.element_type == self.dtype
+        assert mY.element_type == self.dtype
+        assert mdX.element_type == self.dtype
+        self._set_cluster_n()
+        tiler_mn, tv_layout = self._get_tv_layout()
+        num_threads = cute.size(tv_layout, mode=[0])
+        num_warps = num_threads // cute.arch.WARP_SIZE
+        self.kernel(mdY, mY, mdX, tv_layout, tiler_mn).launch(
+            grid=[cute.ceil_div(mdY.shape[0], tiler_mn[0]), self.cluster_n, 1],
+            block=[num_threads, 1, 1],
+            cluster=[1, self.cluster_n, 1] if self.cluster_n > 1 else None,
+            smem=self._smem_size_in_bytes(tiler_mn, num_warps),
+            stream=stream,
+        )
+
+    @cute.kernel
+    def kernel(
+        self,
+        mdY: cute.Tensor,
+        mY: cute.Tensor,
+        mdX: cute.Tensor,
+        tv_layout: cute.Layout,
+        tiler_mn: cute.Shape,
+    ):
+        tidx, _, _ = cute.arch.thread_idx()
+        bidx, cluster_y, _ = cute.arch.block_idx()
+
+        shape = mdY.shape
+        idX = cute.make_identity_tensor(shape)
+        # slice for CTAs
+        gdY, gY, gdX, cX = [
+            cute.local_tile(mT, tiler_mn, (bidx, 0 if self.cluster_n == 1 else cluster_y))
+            for mT in (mdY, mY, mdX, idX)
+        ]
+
+        smem = cutlass.utils.SmemAllocator()
+        sdY = smem.allocate_tensor(
+            mdY.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+        )
+        sY = smem.allocate_tensor(
+            mY.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+        )
+        reduction_buffer, mbar_ptr = self._allocate_reduction_buffer_and_mbar(smem, tv_layout)
+
+        # declare the atoms which will be used later for memory copy
+        copy_atom_load = cute.make_copy_atom(
+            cute.nvgpu.cpasync.CopyG2SOp(), mdY.element_type, num_bits_per_copy=128
+        )
+        copy_atom_store = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), gdX.element_type, num_bits_per_copy=128
+        )
+
+        thr_copy_load = cute.make_tiled_copy(copy_atom_load, tv_layout, tiler_mn).get_slice(tidx)
+        thr_copy_store = cute.make_tiled_copy(copy_atom_store, tv_layout, tiler_mn).get_slice(tidx)
+
+        tdYgdY = thr_copy_load.partition_S(gdY)
+        tdYsdY = thr_copy_load.partition_D(sdY)
+        tYgY = thr_copy_load.partition_S(gY)
+        tYsY = thr_copy_load.partition_D(sY)
+        tdXgdX = thr_copy_store.partition_D(gdX)
+        tXcX = thr_copy_load.partition_S(cX)[(0, None), None, None]
+
+        # allocate fragments for gmem->rmem
+        tdYrdY, tYrY, tdXrdX = [cute.make_fragment_like(thr) for thr in (tdYgdY, tYgY, tdXgdX)]
+
+        num_warps = cute.size(tv_layout, mode=[0]) // cute.arch.WARP_SIZE
+        self._initialize_cluster(tidx, mbar_ptr, num_warps)
+
+        is_even_N = cutlass.const_expr(shape[1] == tiler_mn[1] * self.cluster_n)
+        tdYpdY = (
+            utils.predicate_k(thr_copy_load.partition_S(cX), limit=shape[1])
+            if not is_even_N
+            else None
+        )
+
+        if tXcX[0][0] < shape[0]:
+            cute.copy(copy_atom_load, tdYgdY, tdYsdY, pred=tdYpdY)
+            cute.copy(copy_atom_load, tYgY, tYsY, pred=tdYpdY)
+        cute.arch.cp_async_commit_group()
+        cute.arch.cp_async_wait_group(0)
+
+        cute.autovec_copy(tdYsdY, tdYrdY)
+        cute.autovec_copy(tYsY, tYrY)
+        dy = tdYrdY.load().to(cute.Float32)
+        y = tYrY.load().to(cute.Float32)
+
+        # Compute dot product: dot = Σⱼ dy_j × y_j
+        threads_per_row = tv_layout.shape[0][0]
+        dot = utils.row_reduce(
+            dy * y,
+            cute.ReductionOp.ADD,
+            threads_per_row,
+            reduction_buffer[None, None, 0],
+            mbar_ptr if self.cluster_n > 1 else None,
+            init_val=0.0,
+            hook_fn=cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None,
+        )
+
+        # Compute gradient: dx_i = y_i × (dy_i - dot)
+        dx = y * (dy - dot)
+        tdXrdX.store(dx.to(tdXrdX.element_type))
+        tdXpdX = (
+            utils.predicate_k(thr_copy_store.partition_S(cX), limit=shape[1])
+            if not is_even_N
+            else None
+        )
+        if tXcX[0][0] < shape[0]:
+            cute.copy(copy_atom_store, tdXrdX, tdXgdX, pred=tdXpdX)
+
+
+def _softmax_backward(dy: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+    """Softmax backward pass.
+    Args:
+        dy: Upstream gradients tensor of shape (M, N)
+        y: Softmax output tensor of shape (M, N)
+    Returns:
+        Input gradients tensor of same shape as dy and y
+    """
+    assert dy.dim() == 2, "dy must be 2D"
+    assert y.dim() == 2, "y must be 2D"
+    assert dy.shape == y.shape, "dy and y must have same shape"
+    assert dy.is_cuda and y.is_cuda, "Tensors must be on CUDA device"
+    assert dy.dtype in [torch.float16, torch.bfloat16, torch.float32], "Unsupported dtype"
+    assert y.dtype == dy.dtype, "dy and y must have same dtype"
+
+    M, N = dy.shape
+    dx = torch.empty_like(dy)
+    dtype = torch2cute_dtype_map[dy.dtype]
+    convert_from_dlpack = lambda tensor: (
+        from_dlpack(tensor.detach(), assumed_align=16).mark_compact_shape_dynamic(
+            mode=0, stride_order=(0, 1)
+        )
+    )
+    dy_tensor, y_tensor, dx_tensor = [convert_from_dlpack(tensor) for tensor in (dy, y, dx)]
+    current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
+
+    compile_key = (dtype, N)
+    if compile_key not in _softmax_backward.compile_cache:
+        softmax_backward_op = SoftmaxBackward(dtype, N)
+        _softmax_backward.compile_cache[compile_key] = cute.compile(
+            softmax_backward_op, dy_tensor, y_tensor, dx_tensor, current_stream
+        )
+    _softmax_backward.compile_cache[compile_key](dy_tensor, y_tensor, dx_tensor, current_stream)
+    return dx
+
+
+_softmax_backward.compile_cache = {}
+
+
+class SoftmaxFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x):
+        y = _softmax_fwd(x)
+        ctx.save_for_backward(y)
+        return y
+
+    @staticmethod
+    def backward(ctx, dy):
+        (y,) = ctx.saved_tensors
+        dx = _softmax_backward(dy, y)
+        return dx
+
+
+def softmax(x: torch.Tensor) -> torch.Tensor:
+    """Softmax forward pass with automatic differentiation support.
+
+    Args:
+        x: Input tensor of shape (M, N)
+
+    Returns:
+        Softmax output tensor of same shape as x
+    """
+    return SoftmaxFunction.apply(x)