quack-kernels 0.1.4__py3-none-any.whl → 0.1.6__py3-none-any.whl

quack/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.1.4"
+__version__ = "0.1.6"
 
 from quack.rmsnorm import rmsnorm
 from quack.softmax import softmax

quack/cross_entropy.py

@@ -1,3 +1,5 @@
+# Copyright (c) 2025, Wentao Guo, Ted Zadouri, Tri Dao.
+
 import math
 import torch
 from typing import Optional, Type
@@ -102,7 +104,10 @@ class CrossEntropy(ReductionBase):
         shape: cute.Shape = mX.shape
         idX = cute.make_identity_tensor(shape)
         # slice for CTAs
-        gX, cX = [cute.local_tile(mT, tiler_mn, (bidx, cluster_y)) for mT in (mX, idX)]
+        # We use domain_offset_i64 to deal with tensors larger than 2^31 elements
+        mX_off = utils.domain_offset_i64((bidx * tiler_mn[0], 0), mX)
+        gX = cute.local_tile(mX_off, tiler_mn, (0, cluster_y))
+        cX = cute.local_tile(idX, tiler_mn, (bidx, cluster_y))
 
         smem = cutlass.utils.SmemAllocator()
         sX = smem.allocate_tensor(
@@ -148,7 +153,9 @@ class CrossEntropy(ReductionBase):
 
         target_logit = cute.Float32.zero
         if row < shape[0] and tXcX[0][1] == 0:
-            target_logit = cute.Float32(mX[row, target])
+            # Use Int64 for indexing to deal with large tensors
+            mX_off = utils.domain_offset_i64((row, 0), mX)
+            target_logit = cute.Float32(mX_off[0, target])
 
         threads_per_row = tv_layout.shape[0][0]
         if cutlass.const_expr(not self.online_softmax):
@@ -200,7 +207,7 @@ class CrossEntropy(ReductionBase):
                 mLSE[row] = lse
 
 
-def cross_entropy(
+def _cross_entropy(
     x: torch.Tensor,
     target: torch.Tensor,
     return_lse: bool = False,
@@ -241,15 +248,299 @@ def cross_entropy(
     stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
 
     compile_key = (dtype, N, lse is not None)
-    if compile_key not in cross_entropy.compile_cache:
+    if compile_key not in _cross_entropy.compile_cache:
         cross_entropy_op = CrossEntropy(dtype, N)
-        cross_entropy.compile_cache[compile_key] = cute.compile(
+        _cross_entropy.compile_cache[compile_key] = cute.compile(
             cross_entropy_op, x_tensor, target_tensor, loss_tensor, lse_tensor, stream
         )
-    cross_entropy.compile_cache[compile_key](
+    _cross_entropy.compile_cache[compile_key](
         x_tensor, target_tensor, loss_tensor, lse_tensor, stream
     )
     return loss if not return_lse else (loss, lse)
 
 
-cross_entropy.compile_cache = {}
+_cross_entropy.compile_cache = {}
+
+
+class CrossEntropyBackward:
+    def __init__(self, dtype: Type[cutlass.Numeric], N: int):
+        self.dtype = dtype
+        self.N = N
+        self.vecsize = 128 // dtype.width
+
+    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 _get_tv_layout(self):
+        N = self.N
+        vecsize = self.vecsize
+        num_threads = 128 if N <= 16384 else 256
+        threads_per_row = self._calculate_threads_per_row()
+        cols_per_block = num_threads // threads_per_row
+        num_blocks_N = cute.ceil_div(min(N, 16384) // vecsize, threads_per_row)
+        tiler_mn = (cols_per_block, vecsize * num_blocks_N * threads_per_row)
+        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),
+            ),
+        )
+        return tiler_mn, tv_layout
+
+    @cute.jit
+    def __call__(
+        self,
+        mX: cute.Tensor,
+        mTarget: cute.Tensor,
+        mDLoss: cute.Tensor,
+        mdX: cute.Tensor,
+        mLSE: cute.Tensor,
+        stream: cuda.CUstream,
+    ):
+        assert mX.element_type == self.dtype
+        assert mdX.element_type == self.dtype
+
+        tiler_mn, tv_layout = self._get_tv_layout()
+        num_threads = cute.size(tv_layout, mode=[0])
+
+        mDLoss = cute.make_tensor(
+            mDLoss.iterator, cute.append(mDLoss.layout, cute.make_layout((self.N,), stride=(0,)))
+        )
+        mTarget = cute.make_tensor(
+            mTarget.iterator, cute.append(mTarget.layout, cute.make_layout((self.N,), stride=(0,)))
+        )
+        mLSE = cute.make_tensor(
+            mLSE.iterator, cute.append(mLSE.layout, cute.make_layout((self.N,), stride=(0,)))
+        )
+
+        smem_size = cute.size_in_bytes(
+            mX.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0))
+        )
+
+        self.kernel(
+            mX,
+            mTarget,
+            mDLoss,
+            mdX,
+            mLSE,
+            mX.shape,
+            tv_layout,
+            tiler_mn,
+        ).launch(
+            grid=[
+                cute.ceil_div(mX.shape[0], tiler_mn[0]),
+                cute.ceil_div(mX.shape[1], tiler_mn[1]),
+                1,
+            ],
+            block=[num_threads, 1, 1],
+            smem=smem_size,
+            stream=stream,
+        )
+
+    @cute.kernel
+    def kernel(
+        self,
+        mX: cute.Tensor,  # (M, N)
+        mTarget: cute.Tensor,  # (M,)
+        mDLoss: cute.Tensor,  # (M,)
+        mdX: cute.Tensor,  # (M, N)
+        mLSE: cute.Tensor,  # (M,)
+        shape: cute.Shape,
+        tv_layout: cute.Layout,
+        tiler_mn: cute.Shape,
+    ):
+        tidx, _, _ = cute.arch.thread_idx()
+        bidx, bidy, _ = cute.arch.block_idx()
+
+        smem = cutlass.utils.SmemAllocator()
+        sX = smem.allocate_tensor(
+            mX.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+        )
+
+        idX = cute.make_identity_tensor(shape)
+        # We use domain_offset_i64 to deal with tensors larger than 2^31 elements
+        mX, mdX = [utils.domain_offset_i64((bidx * tiler_mn[0], 0), mT) for mT in (mX, mdX)]
+        gX, gdX = [cute.local_tile(mT, tiler_mn, (0, bidy)) for mT in (mX, mdX)]
+        cX = cute.local_tile(idX, tiler_mn, (bidx, bidy))
+
+        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(), gdX.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)
+
+        #### Thread View
+        tXgX = thr_copy_X_async.partition_S(gX)
+        tXsX = thr_copy_X_async.partition_S(sX)
+
+        tXcX = thr_copy_X.partition_S(cX)[(0, None), None, None]
+        tXcFull = thr_copy_X.partition_S(cX)  # improve
+
+        tXgO = thr_copy_O.partition_D(gdX)
+
+        # allocate fragments for gmem->rmem
+        tXrX, tXrO = [cute.make_fragment_like(thr) for thr in (tXgX, tXgO)]
+
+        is_even_N = cutlass.const_expr(shape[1] % tiler_mn[1] == 0)
+        row = tXcX[0][0]
+
+        tXpX = (
+            utils.predicate_k(thr_copy_X_async.partition_S(cX), limit=shape[1])
+            if not is_even_N
+            else None
+        )
+
+        if row < 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)
+        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)
+
+        label = cute.Int32.zero
+        dloss = cute.Float32.zero
+        lse = cute.Float32.zero
+        if row < shape[0]:
+            label = cute.Int32(mTarget[row])
+            dloss = cute.Float32(mDLoss[row])
+            lse = cute.Float32(mLSE[row])
+
+        log2_e = math.log2(math.e)
+        probs = utils.exp2f((x - lse) * log2_e)
+        prob_shifted = probs - 1.0
+
+        mask = cute.make_fragment_like(tXrX, cutlass.Boolean)
+        for i in cutlass.range_constexpr(cute.size(tXcFull)):
+            mask[i] = tXcFull[i][1] == label
+
+        mask = mask.load()
+        grad = cute.where(mask, prob_shifted, probs)
+        grad = grad * dloss
+
+        tXrO.store(grad.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 row < shape[0]:
+            cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tOpO)
+
+
+def _cross_entropy_backward(
+    x: torch.Tensor,
+    target: torch.Tensor,
+    dloss: torch.Tensor,
+    lse: torch.Tensor,
+    inplace_backward: bool = False,
+) -> torch.Tensor:
+    """Cross entropy backward pass.
+    Args:
+        x: Input logits tensor of shape (M, N)
+        target: Target class indices tensor of shape (M,)
+        dloss: Upstream gradients tensor of shape (M,)
+        lse: Log-sum-exp values tensor of shape (M,)
+    Returns:
+        Input gradients tensor of shape (M, N)
+    """
+    assert x.dim() == 2, "Input must be 2D"
+    assert target.dim() == 1, "Target must be 1D"
+    assert dloss.dim() == 1, "dloss must be 1D"
+    assert lse.dim() == 1, "lse must be 1D"
+    assert x.shape[0] == target.shape[0], "Batch dimensions must match"
+    assert x.shape[0] == dloss.shape[0], "Batch dimensions must match"
+    assert x.shape[0] == lse.shape[0], "Batch dimensions must match"
+    assert (
+        x.is_cuda and target.is_cuda and dloss.is_cuda and lse.is_cuda
+    ), "Tensors must be on CUDA device"
+    assert x.dtype in [torch.float16, torch.bfloat16, torch.float32], "Unsupported input dtype"
+    assert target.dtype in [torch.int32, torch.int64], "Target must be int32 or int64"
+
+    M, N = x.shape
+    dx = torch.empty_like(x) if not inplace_backward else 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)
+        )
+    )
+    x_tensor = convert_from_dlpack(x)
+    dx_tensor = convert_from_dlpack(dx)
+    dloss_tensor = from_dlpack(dloss.detach(), assumed_align=16).mark_compact_shape_dynamic(mode=0)
+    lse_tensor = from_dlpack(lse.detach(), assumed_align=16).mark_compact_shape_dynamic(mode=0)
+    target_tensor = from_dlpack(target.detach(), assumed_align=32).mark_compact_shape_dynamic(
+        mode=0
+    )
+    stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
+
+    compile_key = (dtype, N)
+    if compile_key not in _cross_entropy_backward.compile_cache:
+        cross_entropy_backward_op = CrossEntropyBackward(dtype, N)
+        _cross_entropy_backward.compile_cache[compile_key] = cute.compile(
+            cross_entropy_backward_op,
+            x_tensor,
+            target_tensor,
+            dloss_tensor,
+            dx_tensor,
+            lse_tensor,
+            stream,
+        )
+    _cross_entropy_backward.compile_cache[compile_key](
+        x_tensor, target_tensor, dloss_tensor, dx_tensor, lse_tensor, stream
+    )
+    return dx
+
+
+_cross_entropy_backward.compile_cache = {}
+
+
+class CrossEntropyFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, target, inplace_backward=False):
+        loss, lse = _cross_entropy(x, target, return_lse=True)
+        ctx.save_for_backward(x, target, lse)
+        ctx.inplace_backward = inplace_backward
+        return loss
+
+    @staticmethod
+    def backward(ctx, dloss):
+        x, target, lse = ctx.saved_tensors
+        dx = _cross_entropy_backward(x, target, dloss, lse, inplace_backward=ctx.inplace_backward)
+        return dx, None, None
+
+
+def cross_entropy(
+    x: torch.Tensor, target: torch.Tensor, inplace_backward: bool = False
+) -> torch.Tensor:
+    """Cross entropy loss with automatic differentiation support.
+
+    Args:
+        x: Input logits tensor of shape (M, N)
+        target: Target class indices tensor of shape (M,)
+
+    Returns:
+        Cross entropy loss tensor of shape (M,)
+    """
+    return CrossEntropyFunction.apply(x, target, inplace_backward)

quack/layernorm.py

@@ -0,0 +1,351 @@
+# Copyright (c) 2025, Wentao Guo, Ted Zadouri, Tri Dao.
+
+
+import torch
+from typing import Optional
+
+import cuda.bindings.driver as cuda
+
+import cutlass
+import cutlass.cute as cute
+from cutlass.cute.runtime import from_dlpack
+import quack.utils as utils
+from quack.reduction_base import ReductionBase, torch2cute_dtype_map
+
+
+class LayerNorm(ReductionBase):
+    def __init__(self, dtype: cutlass.Numeric, N: int):
+        super().__init__(dtype, N, stage=2)  # 2 stages for mean and var
+        self.reload_from = None if N <= 16384 else "smem"
+        self.delay_w_load = False
+
+    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
+        # cluster_n = 4 is faster and cluster_n = 2 for N=64k for some reason
+        # Similarly cluster_n = 8 is faster for N=128k
+        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
+
+    @cute.jit
+    def __call__(
+        self,
+        mX: cute.Tensor,
+        mW: cute.Tensor,
+        mO: cute.Tensor,
+        mRstd: Optional[cute.Tensor],
+        mMean: Optional[cute.Tensor],
+        stream: cuda.CUstream,
+        eps: cutlass.Float32 = 1e-6,
+    ):
+        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
+        mW_expanded_layout = cute.prepend(mW.layout, cute.make_layout((tiler_mn[0],), stride=(0,)))
+        mW = cute.make_tensor(mW.iterator, mW_expanded_layout)
+        if cutlass.const_expr(mRstd is not None):
+            mRstd_expanded_layout = cute.append(
+                mRstd.layout, cute.make_layout((self.N,), stride=(0,))
+            )
+            mRstd = cute.make_tensor(mRstd.iterator, mRstd_expanded_layout)
+        if cutlass.const_expr(mMean is not None):
+            mMean_expanded_layout = cute.append(
+                mMean.layout, cute.make_layout((self.N,), stride=(0,))
+            )
+            mMean = cute.make_tensor(mMean.iterator, mMean_expanded_layout)
+        self.kernel(mX, mW, mO, mRstd, mMean, eps, tv_layout, tiler_mn, self.reload_from).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 cutlass.const_expr(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,
+        mW: cute.Tensor,
+        mO: cute.Tensor,
+        mRstd: Optional[cute.Tensor],
+        mMean: Optional[cute.Tensor],
+        eps: cute.Float32,
+        tv_layout: cute.Layout,
+        tiler_mn: cute.Shape,
+        reload_from: cutlass.Constexpr = None,
+        delay_w_load: cutlass.Constexpr = False,
+    ):
+        tidx, _, _ = cute.arch.thread_idx()
+        bidx, _, _ = cute.arch.block_idx()
+        if cutlass.const_expr(self.cluster_n > 1):
+            cluster_y = cute.arch.block_idx()[1]
+        else:
+            cluster_y = cutlass.const_expr(0)
+
+        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)
+
+        shape = mX.shape
+        idX = cute.make_identity_tensor(shape)
+        # slice for CTAs
+        # We use domain_offset_i64 to deal with tensors larger than 2^31 elements
+        mX, mO = [utils.domain_offset_i64((bidx * tiler_mn[0], 0), mT) for mT in (mX, mO)]
+        gX, gO = [cute.local_tile(mT, tiler_mn, (0, cluster_y)) for mT in (mX, mO)]
+        cX = cute.local_tile(idX, tiler_mn, (bidx, cluster_y))
+        gW = cute.local_tile(mW, tiler_mn, (0, cluster_y))
+        gRstd = (
+            cute.local_tile(mRstd, tiler_mn, (bidx, cluster_y))
+            if cutlass.const_expr(mRstd is not None)
+            else None
+        )
+        gMean = (
+            cute.local_tile(mMean, tiler_mn, (bidx, cluster_y))
+            if cutlass.const_expr(mMean is not None)
+            else None
+        )
+
+        # declare the atoms which will be used later for memory copy
+        copy_atom_load_X = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mX.element_type, num_bits_per_copy=128
+        )
+        copy_atom_load_X_async = cute.make_copy_atom(
+            cute.nvgpu.cpasync.CopyG2SOp(), mX.element_type, num_bits_per_copy=128
+        )
+        copy_atom_load_W = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mW.element_type, num_bits_per_copy=128
+        )
+        copy_atom_store_O = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mO.element_type, num_bits_per_copy=128
+        )
+
+        thr_copy_X = cute.make_tiled_copy(copy_atom_load_X_async, tv_layout, tiler_mn).get_slice(
+            tidx
+        )
+        thr_copy_W = cute.make_tiled_copy(copy_atom_load_W, 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)
+
+        tWgW = thr_copy_W.partition_S(gW)
+        tXgX = thr_copy_X.partition_S(gX)
+        tXsX = thr_copy_X.partition_D(sX)
+        tXgO = thr_copy_O.partition_D(gO)
+        tXrRstd = thr_copy_O.partition_D(gRstd) if cutlass.const_expr(mRstd is not None) else None
+        tXrMean = thr_copy_O.partition_D(gMean) if cutlass.const_expr(mMean is not None) else None
+        tXcX = thr_copy_X.partition_S(cX)[(0, None), None, None]
+
+        # allocate fragments for gmem->rmem
+        tWrW = cute.make_fragment_like(tWgW)
+        tXrW = thr_copy_X.retile(tWrW)
+        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)
+
+        tXpX = utils.predicate_k(thr_copy_X.partition_S(cX), limit=shape[1])
+        row = tXcX[0][0]
+        if row < shape[0]:
+            cute.copy(copy_atom_load_X_async, tXgX, tXsX, pred=tXpX)
+        cute.arch.cp_async_commit_group()
+
+        tWpW = utils.predicate_k(thr_copy_W.partition_S(cX), limit=shape[1])
+        if cutlass.const_expr(not delay_w_load):
+            cute.copy(copy_atom_load_W, tWgW, tWrW, pred=tWpW)
+
+        cute.arch.cp_async_wait_group(0)
+        cute.autovec_copy(tXsX, tXrX)
+        x = tXrX.load().to(cute.Float32)
+        threads_per_row = tv_layout.shape[0][0]
+        sum_x = utils.row_reduce(
+            x,
+            cute.ReductionOp.ADD,
+            threads_per_row,
+            reduction_buffer[None, None, 0],
+            mbar_ptr + 0 if cutlass.const_expr(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,
+        )
+        mean = sum_x / shape[1]
+        if cutlass.const_expr(reload_from == "smem"):
+            cute.autovec_copy(tXsX, tXrX)
+            x = tXrX.load().to(cute.Float32)
+        elif cutlass.const_expr(reload_from == "gmem"):
+            cute.copy(copy_atom_load_X, tXgX, tXrX, pred=tXpX)
+            x = tXrX.load().to(cute.Float32)
+
+        sum_sq_x_sub_mean = utils.row_reduce(
+            (x - mean) * (x - mean),
+            cute.ReductionOp.ADD,
+            threads_per_row,
+            reduction_buffer[None, None, 1],
+            mbar_ptr + 1 if cutlass.const_expr(self.cluster_n > 1) else None,
+            init_val=0.0,
+        )
+        rstd = utils.rsqrt(sum_sq_x_sub_mean / shape[1] + eps)
+        if cutlass.const_expr(mRstd is not None):
+            # Only the thread corresponding to column 0 writes out the rstd to gmem
+            if (
+                tXcX[0][1] == 0
+                and row < shape[0]
+                and (self.cluster_n == 1 or cute.arch.block_idx_in_cluster() == 0)
+            ):
+                tXrRstd[0] = rstd
+        if cutlass.const_expr(mMean is not None):
+            # Only the thread corresponding to column 0 writes out the mean to gmem
+            if (
+                tXcX[0][1] == 0
+                and row < shape[0]
+                and (self.cluster_n == 1 or cute.arch.block_idx_in_cluster() == 0)
+            ):
+                tXrMean[0] = mean
+        if cutlass.const_expr(delay_w_load):
+            cute.copy(copy_atom_load_W, tWgW, tWrW, pred=tWpW)
+        if cutlass.const_expr(reload_from == "smem"):
+            cute.autovec_copy(tXsX, tXrX)
+            x = tXrX.load().to(cute.Float32)
+        elif cutlass.const_expr(reload_from == "gmem"):
+            cute.copy(copy_atom_load_X, tXgX, tXrX, pred=tXpX)
+            x = tXrX.load().to(cute.Float32)
+        x_hat = (x - mean) * rstd
+        w = tXrW.load().to(cute.Float32)
+        y = x_hat * w
+        tXrO.store(y.to(tXrO.element_type))
+        tOpO = utils.predicate_k(thr_copy_O.partition_S(cX), limit=shape[1])
+        if row < shape[0]:
+            cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tOpO)
+
+
+def layernorm(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    eps: float = 1e-6,
+    return_rstd: bool = False,
+    return_mean: bool = False,
+) -> torch.Tensor:
+    """LayerNorm forward pass.
+
+    Args:
+        x: Input tensor of shape (M, N)
+        weight: Weight tensor of shape (N,)
+        eps: Small value for numerical stability
+        return_rstd: Whether to return the reciprocal standard deviation
+        return_mean: Whether to return the mean
+
+    Returns:
+        Normalized output tensor of same shape as x
+        If return_rstd is True, also returns rstd tensor of shape (M,)
+        If return_mean is True, also returns mean tensor of shape (M,)
+    """
+    assert x.dim() == 2, "Input must be 2D"
+    assert weight.dim() == 1, "Weight must be 1D"
+    assert x.shape[-1] == weight.shape[0], "Last dimension of input must match weight dimension"
+    assert x.is_cuda and weight.is_cuda, "Tensors must be on CUDA device"
+    assert x.dtype in [torch.float16, torch.bfloat16, torch.float32], "Unsupported dtype"
+    assert weight.dtype == torch.float32, "Weight must be float32"
+    M, N = x.shape
+    device = x.device
+    out = torch.empty_like(x)
+    rstd = torch.empty(M, device=device, dtype=torch.float32) if return_rstd else None
+    mean = torch.empty(M, device=device, dtype=torch.float32) if return_mean else None
+    dtype = torch2cute_dtype_map[x.dtype]
+    convert_from_dlpack = lambda x: (
+        from_dlpack(x.detach(), assumed_align=16).mark_compact_shape_dynamic(
+            mode=0, stride_order=(0, 1)
+        )
+    )
+    x_tensor, out_tensor = [
+        # utils.convert_from_dlpack(t, leading_dim=t.ndim - 1, divisibility=128 // dtype.width)
+        convert_from_dlpack(t)
+        for t in (x, out)
+    ]
+    weight_tensor = utils.convert_from_dlpack(
+        weight.detach(), leading_dim=0, divisibility=128 // cutlass.Float32.width
+    )
+    rstd_tensor = (
+        from_dlpack(rstd.detach(), assumed_align=4).mark_layout_dynamic(leading_dim=0)
+        if rstd is not None
+        else None
+    )
+    mean_tensor = (
+        from_dlpack(mean.detach(), assumed_align=4).mark_layout_dynamic(leading_dim=0)
+        if mean is not None
+        else None
+    )
+    current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
+    compile_key = (dtype, N, rstd is not None, mean is not None)
+    if compile_key not in layernorm.compile_cache:
+        rmsnorm_op = LayerNorm(dtype, N)
+        layernorm.compile_cache[compile_key] = cute.compile(
+            rmsnorm_op,
+            x_tensor,
+            weight_tensor,
+            out_tensor,
+            rstd_tensor,
+            mean_tensor,
+            current_stream,
+        )
+    layernorm.compile_cache[compile_key](
+        x_tensor, weight_tensor, out_tensor, rstd_tensor, mean_tensor, current_stream, eps
+    )
+    return (
+        (out, rstd, mean)
+        if return_mean and return_rstd
+        else (
+            (out, rstd)
+            if return_rstd and not return_mean
+            else ((out, mean) if return_mean and not return_rstd else (out))
+        )
+    )
+
+
+layernorm.compile_cache = {}
+
+
+def layernorm_ref(x: torch.Tensor, w: torch.Tensor, eps: float = 1e-6):
+    x_f32 = x.float()
+    return torch.nn.functional.layer_norm(x_f32, w.shape, w, None, eps).to(x.dtype)
+
+
+def rstd_ref(x: torch.Tensor, eps: float = 1e-6):
+    x_f32 = x.float()
+    mean = x_f32.mean(dim=-1, keepdim=True)
+    var = ((x_f32 - mean) ** 2).mean(dim=-1)
+    return 1.0 / torch.sqrt(var + eps)
+
+
+def mean_ref(x: torch.Tensor) -> torch.Tensor:
+    return x.float().mean(dim=-1)

quack/reduction_base.py

@@ -6,8 +6,6 @@ from typing import Type, Tuple, Optional
 import cutlass
 import cutlass.cute as cute
 
-import quack.utils as utils
-
 
 torch2cute_dtype_map = {
     torch.float16: cutlass.Float16,
@@ -39,7 +37,6 @@ class ReductionBase:
         vecsize = copy_bits // self.dtype.width
         assert self.N % vecsize == 0, f"Input N {self.N} is not divisible by vector size {vecsize}"
         num_threads = self._get_num_threads()
-        num_warps = num_threads // cute.arch.WARP_SIZE
         assert num_threads % cute.arch.WARP_SIZE == 0
 
         threads_per_row = self._calculate_threads_per_row()
@@ -64,7 +61,7 @@ class ReductionBase:
 
     def _get_reduction_buffer_layout(self, tv_layout: cute.Layout, cluster_n: int):
         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)
+        warps_per_row = max(tv_layout.shape[0][0] // cute.arch.WARP_SIZE, 1)
         return cute.make_ordered_layout(
             (num_warps // warps_per_row, (warps_per_row, cluster_n), self.stage),
             order=(1, 0, 2),

quack/rmsnorm.py

@@ -9,7 +9,6 @@ import cuda.bindings.driver as cuda
 import cutlass
 import cutlass.cute as cute
 from cutlass.cute.runtime import from_dlpack
-
 import quack.utils as utils
 from quack.reduction_base import ReductionBase, torch2cute_dtype_map
 
@@ -118,7 +117,10 @@ class RMSNorm(ReductionBase):
         shape = mX.shape
         idX = cute.make_identity_tensor(shape)
         # slice for CTAs
-        gX, gO, cX = [cute.local_tile(mT, tiler_mn, (bidx, cluster_y)) for mT in (mX, mO, idX)]
+        # We use domain_offset_i64 to deal with tensors larger than 2^31 elements
+        mX, mO = [utils.domain_offset_i64((bidx * tiler_mn[0], 0), mT) for mT in (mX, mO)]
+        gX, gO = [cute.local_tile(mT, tiler_mn, (0, cluster_y)) for mT in (mX, mO)]
+        cX = cute.local_tile(idX, tiler_mn, (bidx, cluster_y))
         gW = cute.local_tile(mW, tiler_mn, (0, cluster_y))
         gRstd = (
             cute.local_tile(mRstd, tiler_mn, (bidx, cluster_y))
@@ -210,20 +212,18 @@ class RMSNorm(ReductionBase):
             cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tOpO)
 
 
-def rmsnorm(
+def _rmsnorm_fwd(
     x: torch.Tensor,
     weight: torch.Tensor,
     eps: float = 1e-6,
     return_rstd: bool = False,
 ) -> torch.Tensor:
     """RMSNorm forward pass.
-
     Args:
         x: Input tensor of shape (M, N)
         weight: Weight tensor of shape (N,)
         eps: Small value for numerical stability
         return_rstd: Whether to return the reciprocal standard deviation
-
     Returns:
         Normalized output tensor of same shape as x
         If return_rstd is True, also returns rstd tensor of shape (M,)
@@ -259,18 +259,18 @@ def rmsnorm(
     )
     current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
     compile_key = (dtype, N, rstd is not None)
-    if compile_key not in rmsnorm.compile_cache:
+    if compile_key not in _rmsnorm_fwd.compile_cache:
         rmsnorm_op = RMSNorm(dtype, N)
-        rmsnorm.compile_cache[compile_key] = cute.compile(
+        _rmsnorm_fwd.compile_cache[compile_key] = cute.compile(
             rmsnorm_op, x_tensor, weight_tensor, out_tensor, rstd_tensor, current_stream
         )
-    rmsnorm.compile_cache[compile_key](
+    _rmsnorm_fwd.compile_cache[compile_key](
         x_tensor, weight_tensor, out_tensor, rstd_tensor, current_stream, eps
     )
     return (out, rstd) if return_rstd else out
 
 
-rmsnorm.compile_cache = {}
+_rmsnorm_fwd.compile_cache = {}
 
 
 def rmsnorm_ref(x, w, eps=1e-6):
@@ -283,3 +283,383 @@ def rmsnorm_ref(x, w, eps=1e-6):
 def rstd_ref(x, eps=1e-6):
     x_f32 = x.float()
     return 1.0 / torch.sqrt(torch.mean(x_f32 * x_f32, dim=-1) + eps)
+
+
+def rmsnorm_bwd_ref(x, w, dout, rstd, eps=1e-6):
+    """Reference implementation for RMSNorm backward pass."""
+    x_f32 = x.float()
+    x_hat = x_f32 * rstd.unsqueeze(1)
+    wdy = dout * w
+    c1 = (x_hat * wdy).mean(dim=-1, keepdim=True)
+    dx = (wdy - x_hat * c1) * rstd.unsqueeze(1)
+
+    # dL/dW
+    dw = (dout * x_hat).sum(dim=0)
+    return dx.to(x.dtype), dw.to(w.dtype)
+
+
+class RMSNormBackward(ReductionBase):
+    def __init__(self, dtype: cutlass.Numeric, N: int):
+        # 1 stage for computing mean of x_hat * wdy
+        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 <= 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
+
+    @cute.jit
+    def __call__(
+        self,
+        mX: cute.Tensor,
+        mW: cute.Tensor,
+        mDout: cute.Tensor,
+        mRstd: cute.Tensor,
+        mDx: cute.Tensor,
+        mDw: cute.Tensor,
+        sm_count: cutlass.Constexpr,
+        stream: cuda.CUstream,
+    ):
+        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
+
+        mW_expanded_layout = cute.prepend(mW.layout, cute.make_layout((tiler_mn[0],), stride=(0,)))
+        mW = cute.make_tensor(mW.iterator, mW_expanded_layout)
+
+        mRstd_expanded_layout = cute.append(mRstd.layout, cute.make_layout((self.N,), stride=(0,)))
+        mRstd = cute.make_tensor(mRstd.iterator, mRstd_expanded_layout)
+
+        num_blocks = (
+            sm_count if tiler_mn[0] == 1 else min(sm_count, cute.ceil_div(1024, tiler_mn[0]))
+        )
+
+        self.kernel(mX, mW, mDout, mRstd, mDx, mDw, sm_count, tv_layout, tiler_mn).launch(
+            grid=[num_blocks, 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,
+        mW: cute.Tensor,
+        mDout: cute.Tensor,
+        mRstd: cute.Tensor,
+        mDx: cute.Tensor,
+        mDw: cute.Tensor,
+        sm_count: cutlass.Constexpr,
+        tv_layout: cute.Layout,
+        tiler_mn: cute.Shape,
+    ):
+        tidx, _, _ = cute.arch.thread_idx()
+        bidx, cluster_y, _ = cute.arch.block_idx()
+        gdim, _, _ = cute.arch.grid_dim()
+
+        shape = mX.shape
+        M, N = shape[0], shape[1]
+
+        idX = cute.make_identity_tensor(shape)
+
+        smem = cutlass.utils.SmemAllocator()
+        reduction_buffer, mbar_ptr = self._allocate_reduction_buffer_and_mbar(smem, tv_layout)
+
+        copy_atom_load_X = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mX.element_type, num_bits_per_copy=128
+        )
+
+        copy_atom_load_W = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mW.element_type, num_bits_per_copy=128
+        )
+
+        copy_atom_store_dX = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mDx.element_type, num_bits_per_copy=128
+        )
+
+        copy_atom_dw = cute.make_copy_atom(
+            cute.nvgpu.CopyUniversalOp(), mDw.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_W = cute.make_tiled_copy(copy_atom_load_W, tv_layout, tiler_mn).get_slice(tidx)
+        thr_copy_dw = cute.make_tiled_copy(copy_atom_dw, tv_layout, tiler_mn).get_slice(tidx)
+        thr_store_dx = cute.make_tiled_copy(copy_atom_store_dX, tv_layout, tiler_mn).get_slice(tidx)
+
+        gW = cute.local_tile(mW, tiler_mn, (bidx, 0 if self.cluster_n == 1 else cluster_y))
+        tWgW = thr_copy_W.partition_S(gW)
+        tWrW = cute.make_fragment_like(tWgW)
+        tXrW = thr_copy_X.retile(tWrW)
+
+        gW_coord = cute.local_tile(idX, tiler_mn, (0, 0 if self.cluster_n == 1 else cluster_y))
+
+        tWpW = utils.predicate_k(thr_copy_W.partition_S(gW_coord), limit=shape[1])
+        cute.copy(copy_atom_load_W, tWgW, tWrW, pred=tWpW)
+        weight = tXrW.load().to(cute.Float32)
+
+        num_warps = cute.size(tv_layout, mode=[0]) // cute.arch.WARP_SIZE
+
+        self._initialize_cluster(tidx, mbar_ptr, num_warps)
+
+        dw_coord = cute.local_tile(idX, tiler_mn, (0, 0 if self.cluster_n == 1 else cluster_y))
+        tDwpDw = utils.predicate_k(thr_copy_dw.partition_S(dw_coord), limit=shape[1])
+
+        gDw = cute.local_tile(mDw, tiler_mn, (bidx, 0 if self.cluster_n == 1 else cluster_y))
+        tDwgDw = thr_copy_dw.partition_D(gDw)
+        tDwrDw = cute.make_fragment_like(tDwgDw)
+        dw_accumulator = thr_copy_X.retile(tDwrDw)
+        dw_accumulator.fill(0.0)
+
+        M_pad = ((M + sm_count - 1) // sm_count) * sm_count
+
+        jump = sm_count if tiler_mn[0] == 1 else min(sm_count, cute.ceil_div(1024, tiler_mn[0]))
+
+        if cutlass.const_expr(self.cluster_n > 1):
+            cute.arch.cluster_arrive()
+            cute.arch.cluster_wait()
+
+        ## need to update range_dynamic since it will be deprecated soon
+        for row_offset in cutlass.range_dynamic(bidx, M_pad, jump):
+            gX = cute.local_tile(
+                mX, tiler_mn, (row_offset, 0 if self.cluster_n == 1 else cluster_y)
+            )
+            gDout = cute.local_tile(
+                mDout, tiler_mn, (row_offset, 0 if self.cluster_n == 1 else cluster_y)
+            )
+            gRstd = cute.local_tile(
+                mRstd, tiler_mn, (row_offset, 0 if self.cluster_n == 1 else cluster_y)
+            )
+            gDx = cute.local_tile(
+                mDx, tiler_mn, (row_offset, 0 if self.cluster_n == 1 else cluster_y)
+            )
+            cX = cute.local_tile(
+                idX, tiler_mn, (row_offset, 0 if self.cluster_n == 1 else cluster_y)
+            )
+
+            tXgX = thr_copy_X.partition_S(gX)
+            thrDout = thr_copy_X.partition_S(gDout)
+            tXrRstd = thr_copy_W.partition_S(gRstd)
+            thrDx = thr_store_dx.partition_D(gDx)
+            tXcX = thr_copy_X.partition_S(cX)[(0, None), None, None]
+
+            tXrX, frgDout, frgDx = [cute.make_fragment_like(thr) for thr in (tXgX, thrDout, thrDx)]
+
+            tXpX = utils.predicate_k(thr_copy_X.partition_S(cX), limit=shape[1])
+
+            if tXcX[0][0] < shape[0]:
+                cute.copy(copy_atom_load_X, tXgX, tXrX, pred=tXpX)
+                cute.copy(copy_atom_load_X, thrDout, frgDout, pred=tXpX)
+
+            x = tXrX.load().to(cute.Float32)
+            dout = frgDout.load().to(cute.Float32)
+
+            rstd = tXrRstd[0]
+            x_hat = x * rstd
+            wdy = dout * weight
+
+            threads_per_row = tv_layout.shape[0][0]
+
+            row = tXcX[0][0]
+            if cutlass.const_expr(self.cluster_n > 1):
+                cute.arch.cluster_arrive()
+                cute.arch.cluster_wait()
+            else:
+                cute.arch.barrier()
+
+            mean_xhat_wdy = (
+                utils.row_reduce(
+                    x_hat * wdy,
+                    cute.ReductionOp.ADD,
+                    threads_per_row,
+                    reduction_buffer[None, None, 0],
+                    mbar_ptr + 0 if cutlass.const_expr(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,
+                )
+                / shape[1]
+            )
+
+            dx = (wdy - x_hat * mean_xhat_wdy) * rstd
+            frgDx.store(dx.to(frgDout.element_type))
+
+            if row < M:
+                cute.copy(copy_atom_store_dX, frgDx, thrDx, pred=tXpX)
+
+            if cutlass.const_expr(self.cluster_n > 1):
+                cute.arch.cluster_arrive()
+                cute.arch.cluster_wait()
+            else:
+                cute.arch.barrier()
+
+            if row < M:
+                dw_row = dout * x_hat
+                current_dw = dw_accumulator.load().to(cute.Float32)
+                updated_dw = current_dw + dw_row
+                dw_accumulator.store(updated_dw.to(dw_accumulator.element_type))
+
+            """ 
+            if cutlass.const_expr(self.cluster_n > 1):
+                cute.arch.cluster_arrive()
+                cute.arch.cluster_wait()
+            else:
+                cute.arch.barrier()
+            """
+        """ 
+        if cutlass.const_expr(self.cluster_n > 1):
+            cute.arch.cluster_arrive()
+            cute.arch.cluster_wait()
+        else:
+            cute.arch.barrier()
+        """
+
+        cute.autovec_copy(dw_accumulator, tDwrDw)
+        cute.copy(copy_atom_dw, tDwrDw, tDwgDw, pred=tDwpDw)
+
+
+def _rmsnorm_backward(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    dout: torch.Tensor,
+    rstd: torch.Tensor,
+) -> (torch.Tensor, torch.Tensor):
+    """RMSNorm backward pass.
+    Args:
+        x: Input tensor of shape (M, N)
+        weight: Weight tensor of shape (N,)
+        dout: Upstream gradients tensor of shape (M, N)
+        rstd: Reciprocal standard deviation tensor of shape (M,)
+    Returns:
+        Tuple of (dx, dw) where:
+        - dx: Input gradients tensor of same shape as x
+        - dw: Weight gradients tensor of same shape as weight
+    """
+    assert x.dim() == 2, "Input must be 2D"
+    assert weight.dim() == 1, "Weight must be 1D"
+    assert x.shape[-1] == weight.shape[0], "Last dimension of input must match weight dimension"
+    assert x.is_cuda and weight.is_cuda, "Tensors must be on CUDA device"
+    assert x.dtype in [torch.float16, torch.bfloat16, torch.float32], "Unsupported dtype"
+    assert weight.dtype == torch.float32, "Weight must be float32"
+
+    M, N = x.shape
+    dx = torch.empty_like(x)
+
+    device = x.device
+
+    sm_count = torch.cuda.get_device_properties(device).multi_processor_count * 8
+    dw_partial = torch.zeros((sm_count, N), device=device, dtype=weight.dtype)
+
+    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)
+        )
+    )
+
+    x_tensor, dout_tensor, dx_tensor = [convert_from_dlpack(tensor) for tensor in (x, dout, dx)]
+
+    weight_tensor = utils.convert_from_dlpack(
+        weight.detach(), leading_dim=0, divisibility=128 // cutlass.Float32.width
+    )
+
+    dw_partial_tensor = convert_from_dlpack(dw_partial)
+    rstd_tensor = from_dlpack(rstd.detach(), assumed_align=4).mark_layout_dynamic(leading_dim=0)
+
+    current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
+
+    compile_key = (dtype, N)
+    if compile_key not in _rmsnorm_backward.compile_cache:
+        rmsnorm_backward_op = RMSNormBackward(dtype, N)
+        _rmsnorm_backward.compile_cache[compile_key] = cute.compile(
+            rmsnorm_backward_op,
+            x_tensor,
+            weight_tensor,
+            dout_tensor,
+            rstd_tensor,
+            dx_tensor,
+            dw_partial_tensor,
+            sm_count,
+            current_stream,
+        )
+
+    _rmsnorm_backward.compile_cache[compile_key](
+        x_tensor,
+        weight_tensor,
+        dout_tensor,
+        rstd_tensor,
+        dx_tensor,
+        dw_partial_tensor,
+        current_stream,
+    )
+
+    dw = dw_partial.sum(dim=0).to(weight.dtype)
+    return dx, dw
+
+
+_rmsnorm_backward.compile_cache = {}
+
+
+class RMSNormFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, x, weight, eps):
+        out, rstd = _rmsnorm_fwd(x, weight, eps, return_rstd=True)
+        ctx.save_for_backward(x, weight, rstd)
+        ctx.eps = eps
+        return out
+
+    @staticmethod
+    def backward(ctx, dout):
+        x, weight, rstd = ctx.saved_tensors
+        dx, dw = _rmsnorm_backward(x, weight, dout, rstd)
+        # dw is returned for weight gradient, None for eps gradient
+        return dx, dw, None
+
+
+def rmsnorm(x: torch.Tensor, weight: torch.Tensor, eps: float = 1e-6) -> torch.Tensor:
+    """RMSNorm forward pass with automatic differentiation support.
+
+    Args:
+        x: Input tensor of shape (M, N)
+        weight: Weight tensor of shape (N,)
+        eps: Small value for numerical stability
+
+    Returns:
+        Normalized output tensor of same shape as x
+    """
+    return RMSNormFunction.apply(x, weight, eps)

quack/softmax.py

@@ -98,7 +98,10 @@ class Softmax(ReductionBase):
         shape = mX.shape
         idX = cute.make_identity_tensor(shape)
         # slice for CTAs
-        gX, gO, cX = [cute.local_tile(mT, tiler_mn, (bidx, cluster_y)) for mT in (mX, mO, idX)]
+        # We use domain_offset_i64 to deal with tensors larger than 2^31 elements
+        mX, mO = [utils.domain_offset_i64((bidx * tiler_mn[0], 0), mT) for mT in (mX, mO)]
+        gX, gO = [cute.local_tile(mT, tiler_mn, (0, cluster_y)) for mT in (mX, mO)]
+        cX = cute.local_tile(idX, tiler_mn, (bidx, cluster_y))
 
         smem = cutlass.utils.SmemAllocator()
         sX = smem.allocate_tensor(
@@ -312,9 +315,11 @@ class SoftmaxBackward(ReductionBase):
         shape = mdY.shape
         idX = cute.make_identity_tensor(shape)
         # slice for CTAs
-        gdY, gY, gdX, cX = [
-            cute.local_tile(mT, tiler_mn, (bidx, cluster_y)) for mT in (mdY, mY, mdX, idX)
+        mdY, mY, mdX = [
+            utils.domain_offset_i64((bidx * tiler_mn[0], 0), mT) for mT in (mdY, mY, mdX)
         ]
+        gdY, gY, gdX = [cute.local_tile(mT, tiler_mn, (0, cluster_y)) for mT in (mdY, mY, mdX)]
+        cX = cute.local_tile(idX, tiler_mn, (bidx, cluster_y))
 
         smem = cutlass.utils.SmemAllocator()
         sdY = smem.allocate_tensor(

quack/utils.py

@@ -23,20 +23,6 @@ def convert_from_dlpack(x, leading_dim, alignment=16, divisibility=1) -> cute.Te
     )
 
 
-@cute.jit
-def max_constexpr(
-    a: cutlass.Constexpr[cute.Numeric], b: cutlass.Constexpr[cute.Numeric]
-) -> cutlass.Constexpr[cute.Numeric]:
-    return a if a > b else b
-
-
-@cute.jit
-def min_constexpr(
-    a: cutlass.Constexpr[cute.Numeric], b: cutlass.Constexpr[cute.Numeric]
-) -> cutlass.Constexpr[cute.Numeric]:
-    return a if a < b else b
-
-
 @cute.jit
 def warp_reduce(
     val: cute.TensorSSA | cute.Numeric,
@@ -196,7 +182,7 @@ def row_reduce(
     val = warp_reduce(
         val,
         warp_op,
-        width=min_constexpr(threads_per_row, cute.arch.WARP_SIZE),
+        width=min(threads_per_row, cute.arch.WARP_SIZE),
     )
     if cutlass.const_expr(hook_fn is not None):
         hook_fn()
@@ -226,7 +212,7 @@ def online_softmax_reduce(
     max_x = warp_reduce(
         x.reduce(cute.ReductionOp.MAX, init_val=-Float32.inf, reduction_profile=0),
         cute.arch.fmax,
-        width=min_constexpr(threads_per_row, cute.arch.WARP_SIZE),
+        width=min(threads_per_row, cute.arch.WARP_SIZE),
     )
     log2_e = math.log2(math.e)
     exp_x = exp2f(x * log2_e - (max_x * log2_e))
@@ -234,7 +220,7 @@ def online_softmax_reduce(
     sum_exp_x = warp_reduce(
         exp_x.reduce(cute.ReductionOp.ADD, init_val=0.0, reduction_profile=0),
         operator.add,
-        width=min_constexpr(threads_per_row, cute.arch.WARP_SIZE),
+        width=min(threads_per_row, cute.arch.WARP_SIZE),
     )
     if cutlass.const_expr(hook_fn is not None):
         hook_fn()
@@ -303,7 +289,6 @@ def online_softmax_reduce(
 @cute.jit
 def exp2f(x: cute.TensorSSA | Float32) -> cute.TensorSSA | Float32:
     """exp2f calculation for both vector and scalar.
-
     :param x: input value
     :type x: cute.TensorSSA or Float32
     :return: exp2 value
@@ -405,3 +390,19 @@ def i64_to_f32x2(c: cutlass.Int64, *, loc=None, ip=None) -> Tuple[Float32, Float
         vector.extract(vec_f32x2, dynamic_position=[], static_position=[1], loc=loc, ip=ip)
     )
     return res0, res1
+
+
+@dsl_user_op
+def domain_offset_i64(coord: cute.Coord, tensor: cute.Tensor, *, loc=None, ip=None) -> cute.Tensor:
+    flat_coord_i64 = tuple(cutlass.Int64(c) for c in cute.flatten(coord))
+    flat_stride = cute.flatten_to_tuple(tensor.stride)
+    offset = sum(c * s for c, s in zip(flat_coord_i64, flat_stride))
+    assert isinstance(tensor.iterator, cute.Pointer)
+    # HACK: we assume that applying the offset does not change the pointer alignment
+    new_ptr = cute.make_ptr(
+        tensor.element_type,
+        tensor.iterator.toint() + offset * tensor.element_type.width // 8,
+        tensor.memspace,
+        assumed_align=tensor.iterator.max_alignment,
+    )
+    return cute.make_tensor(new_ptr, tensor.layout)

{quack_kernels-0.1.4.dist-info → quack_kernels-0.1.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quack-kernels
-Version: 0.1.4
+Version: 0.1.6
 Requires-Python: >=3.9
 License-File: LICENSE
 Requires-Dist: nvidia-cutlass-dsl==4.1.0.dev0

quack_kernels-0.1.6.dist-info/RECORD

@@ -0,0 +1,12 @@
+quack/__init__.py,sha256=EV_43VfcxQUsFu_Nfq0944ImloZ2T9X94-8T9IQM0I8,203
+quack/cross_entropy.py,sha256=bg66wECki5I71SMPIRUa-6-oFJ93aIKpK1jqT__SCBM,19775
+quack/layernorm.py,sha256=1WUspbr6ktPZ25O00kKs-FK_lm_Fejat72BMV8tBSfw,13504
+quack/reduction_base.py,sha256=fFuGXPR3lDq2yw_m86ujmkni6R51jzNAzy_r9R6C8tA,3563
+quack/rmsnorm.py,sha256=rbRP_O-EJYvrvxYPFwfy0yCbG7Qs_DgbzgacxTLvih4,24159
+quack/softmax.py,sha256=b-QEQiGjOEPidolE2--K21kwEaLJ-3wQoGIz_BsEcSI,16742
+quack/utils.py,sha256=laz_lqeggiIOY_xCs3c3VvCPP8bJmSKBfjFpFR1Al80,15946
+quack_kernels-0.1.6.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+quack_kernels-0.1.6.dist-info/METADATA,sha256=NKhSoudW9lNdYryNiMkyjKXYl5n37cuFMbyLv3zr5js,289
+quack_kernels-0.1.6.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+quack_kernels-0.1.6.dist-info/top_level.txt,sha256=6e4Jr_vNJbZTYwlO_Ahf_sDeHDE0zcqcf7Le11FKxxo,6
+quack_kernels-0.1.6.dist-info/RECORD,,

quack_kernels-0.1.4.dist-info/RECORD

@@ -1,11 +0,0 @@
-quack/__init__.py,sha256=cFLxO6nA_faFqHf4N-Fy7G0j8ykuYPB1uOt9uoJ2dkQ,203
-quack/cross_entropy.py,sha256=HnF2OErEzb10SWxY6HoYE42lnvlw2DsWCks7mylPwnI,9511
-quack/reduction_base.py,sha256=Rsj9ZeSHcKAXGn1p7mY1vrrBqxevi4feLjY0JJhKnmY,3663
-quack/rmsnorm.py,sha256=TkOZsXJwcsoZMLnmEWQ-pEF0r-iiZhGrCNLSFCXfv6s,10676
-quack/softmax.py,sha256=VfhlC2huRuv7olFSVFgS8LF1yF8TFV64yjjjQxYX9yk,16364
-quack/utils.py,sha256=zVc9U-5No19trE585KqDdXx9chAruXPRIPMZdO7mkRg,15603
-quack_kernels-0.1.4.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-quack_kernels-0.1.4.dist-info/METADATA,sha256=xl62C5WFgiUbnOICAzjldsljJ9j1Fb_JxZVksHLCI8I,289
-quack_kernels-0.1.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-quack_kernels-0.1.4.dist-info/top_level.txt,sha256=6e4Jr_vNJbZTYwlO_Ahf_sDeHDE0zcqcf7Le11FKxxo,6
-quack_kernels-0.1.4.dist-info/RECORD,,