quack-kernels 0.1.5__py3-none-any.whl → 0.1.6__py3-none-any.whl

quack/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.1.5"
+__version__ = "0.1.6"
 
 from quack.rmsnorm import rmsnorm
 from quack.softmax import softmax

quack/cross_entropy.py

@@ -104,7 +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(
@@ -150,7 +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):
@@ -363,11 +368,10 @@ class CrossEntropyBackward:
         )
 
         idX = cute.make_identity_tensor(shape)
-
-        gX, gdX, cX, gTarget, gDLoss, gLse = [
-            cute.local_tile(mT, tiler_mn, (bidx, bidy))
-            for mT in (mX, mdX, idX, mTarget, mDLoss, mLSE)
-        ]
+        # 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

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/rmsnorm.py

@@ -117,7 +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))

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

@@ -390,3 +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-0.1.5.dist-info → quack_kernels-0.1.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quack-kernels
-Version: 0.1.5
+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-0.1.5.dist-info/RECORD

@@ -1,11 +0,0 @@
-quack/__init__.py,sha256=GPoImcynY5-OkMep5RhQhXrnZyxgqZG3RoHhsYQFSL4,203
-quack/cross_entropy.py,sha256=WkngPY8uk4RCjCFtHtB7h9GF_8xt4NnyvDzvw73gIL4,19320
-quack/reduction_base.py,sha256=fFuGXPR3lDq2yw_m86ujmkni6R51jzNAzy_r9R6C8tA,3563
-quack/rmsnorm.py,sha256=N9NavrR85ws4cZgkfpeRLjYkVSq2yfyzJQWvfKf98pY,23935
-quack/softmax.py,sha256=VfhlC2huRuv7olFSVFgS8LF1yF8TFV64yjjjQxYX9yk,16364
-quack/utils.py,sha256=6EyWgf0z3wcbhGUivHmWB8hVBnEzMyOhmAuZ2Te82k0,15226
-quack_kernels-0.1.5.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-quack_kernels-0.1.5.dist-info/METADATA,sha256=WI-2CP1mRH05V9Fjdx7HsErNOkrc6fUhheoH4ynlo-U,289
-quack_kernels-0.1.5.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-quack_kernels-0.1.5.dist-info/top_level.txt,sha256=6e4Jr_vNJbZTYwlO_Ahf_sDeHDE0zcqcf7Le11FKxxo,6
-quack_kernels-0.1.5.dist-info/RECORD,,