quack-kernels 0.1.7__py3-none-any.whl → 0.1.9__py3-none-any.whl

quack/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.1.7"
+__version__ = "0.1.9"
 
 from quack.rmsnorm import rmsnorm
 from quack.softmax import softmax

quack/cross_entropy.py

@@ -1,16 +1,16 @@
 # Copyright (c) 2025, Wentao Guo, Ted Zadouri, Tri Dao.
 
 import math
-import torch
 from typing import Optional, Type
 
 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
+import torch
+from cutlass.cute.runtime import from_dlpack
 from quack.reduction_base import ReductionBase, torch2cute_dtype_map
 
 
@@ -79,7 +79,7 @@ class CrossEntropy(ReductionBase):
         self.kernel(mX, mTarget, mLoss, mLSE, 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 cutlass.const_expr(self.cluster_n > 1) else None,
+            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,
         )
@@ -111,7 +111,9 @@ class CrossEntropy(ReductionBase):
 
         smem = cutlass.utils.SmemAllocator()
         sX = smem.allocate_tensor(
-            mX.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+            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)
 
@@ -166,7 +168,9 @@ class CrossEntropy(ReductionBase):
                 reduction_buffer[None, None, 0],
                 mbar_ptr + 0 if cutlass.const_expr(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,
+                hook_fn=(
+                    cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None
+                ),
             )
             if cutlass.const_expr(self.reload_from == "smem"):
                 cute.autovec_copy(tXsX, tXrX)
@@ -191,7 +195,9 @@ class CrossEntropy(ReductionBase):
                 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,
+                hook_fn=(
+                    cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None
+                ),
             )
 
         if (
@@ -225,7 +231,11 @@ def _cross_entropy(
     assert target.dim() == 1, "Target must be 1D"
     assert x.shape[0] == target.shape[0], "Batch dimensions must match"
     assert x.is_cuda and target.is_cuda, "Tensors must be on CUDA device"
-    assert x.dtype in [torch.float16, torch.bfloat16, torch.float32], "Unsupported input dtype"
+    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
     device = x.device
@@ -314,13 +324,16 @@ class CrossEntropyBackward:
         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,)))
+            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,)))
+            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,)))
+            mLSE.iterator,
+            cute.append(mLSE.layout, cute.make_layout((self.N,), stride=(0,))),
         )
 
         smem_size = cute.size_in_bytes(
@@ -364,7 +377,9 @@ class CrossEntropyBackward:
 
         smem = cutlass.utils.SmemAllocator()
         sX = smem.allocate_tensor(
-            mX.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+            mX.element_type,
+            cute.make_ordered_layout(tiler_mn, order=(1, 0)),
+            byte_alignment=16,
         )
 
         idX = cute.make_identity_tensor(shape)
@@ -474,7 +489,11 @@ def _cross_entropy_backward(
     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 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
@@ -532,15 +551,37 @@ class CrossEntropyFunction(torch.autograd.Function):
 
 
 def cross_entropy(
-    x: torch.Tensor, target: torch.Tensor, inplace_backward: bool = False
+    x: torch.Tensor,
+    target: torch.Tensor,
+    inplace_backward: bool = True,
+    reduction: str = "none",
 ) -> 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,)
+        inplace_backward: Whether to perform backward pass in-place
+        reduction: Specifies the reduction to apply to the output:
+            'none': no reduction will be applied (default)
+            'mean': the sum of the output will be divided by the number of elements
+            'sum': the output will be summed
 
     Returns:
-        Cross entropy loss tensor of shape (M,)
+        Cross entropy loss tensor:
+            - If reduction='none': tensor of shape (M,) with per-example losses
+            - If reduction='mean': scalar tensor with mean loss
+            - If reduction='sum': scalar tensor with sum of losses
     """
-    return CrossEntropyFunction.apply(x, target, inplace_backward)
+    loss = CrossEntropyFunction.apply(x, target, inplace_backward)
+
+    if reduction == "mean":
+        return loss.mean()
+    elif reduction == "sum":
+        return loss.sum()
+    elif reduction == "none":
+        return loss
+    else:
+        raise ValueError(
+            f"Invalid reduction mode: {reduction}. Expected one of 'none', 'mean', or 'sum'"
+        )

quack/rmsnorm.py

@@ -1,14 +1,16 @@
 # 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 import Float32, Int32
 from cutlass.cute.runtime import from_dlpack
+
 import quack.utils as utils
+import torch
 from quack.reduction_base import ReductionBase, torch2cute_dtype_map
 
 
@@ -19,41 +21,55 @@ class RMSNorm(ReductionBase):
         self.delay_w_load = False
 
     def _calculate_threads_per_row(self):
+        """Calculate the number of threads per row for the RMSNorm kernel."""
         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)))
-            )
-        )
+        if N <= 64:
+            return 8
+        elif N <= 128:
+            return 16
+        elif N <= 3072:
+            return 32
+        elif N <= 6144:
+            return 64
+        elif N <= 16384:
+            return 128
+        else:
+            return 256
 
     def _set_cluster_n(self):
+        """
+        Set the number of clusters for the RMSNorm kernel.
+        Stored in self.cluster_n.
+        """
         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))
-                )
-            )
+            # 16-bit types (fp16, bf16)
+            if N <= 16 * 1024:
+                cluster_n = 1
+            elif N <= 32 * 1024:
+                cluster_n = 2
+            elif N <= 64 * 1024:
+                cluster_n = 4
+            elif N <= 128 * 1024:
+                cluster_n = 8
+            else:
+                cluster_n = 16
+        else:
+            # 32-bit types (fp32)
+            if N <= 32 * 1024:
+                cluster_n = 1
+            elif N <= 64 * 1024:
+                cluster_n = 2
+            elif N <= 128 * 1024:
+                cluster_n = 4
+            elif N <= 256 * 1024:
+                cluster_n = 8
+            else:
+                cluster_n = 16
+
         self.cluster_n = cluster_n
 
     @cute.jit
@@ -64,8 +80,17 @@ class RMSNorm(ReductionBase):
         mO: cute.Tensor,
         mRstd: Optional[cute.Tensor],
         stream: cuda.CUstream,
-        eps: cutlass.Float32 = 1e-6,
+        eps: Float32 = 1e-6,
     ):
+        semistatic_shape = (*mX.shape[:-1], self.N)  # Set last dimension to be statically N
+        new_stride = lambda t: (
+            cute.assume(t.stride[0], divby=128 // t.element_type.width),
+            t.stride[1],
+        )
+        mX, mO = [
+            cute.make_tensor(t.iterator, cute.make_layout(semistatic_shape, stride=new_stride(t)))
+            for t in (mX, mO)
+        ]
         assert mX.element_type == self.dtype
         assert mO.element_type == self.dtype
         self._set_cluster_n()
@@ -82,7 +107,7 @@ class RMSNorm(ReductionBase):
         self.kernel(mX, mW, mO, mRstd, 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,
+            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,
         )
@@ -109,7 +134,9 @@ class RMSNorm(ReductionBase):
 
         smem = cutlass.utils.SmemAllocator()
         sX = smem.allocate_tensor(
-            mX.element_type, cute.make_ordered_layout(tiler_mn, order=(1, 0)), byte_alignment=16
+            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)
 
@@ -134,30 +161,33 @@ class RMSNorm(ReductionBase):
         copy_atom_load_X_async = cute.make_copy_atom(
             cute.nvgpu.cpasync.CopyG2SOp(), mX.element_type, num_bits_per_copy=128
         )
+        num_bits_per_copy_W = cutlass.const_expr(
+            min(128, 128 // mX.element_type.width * mW.element_type.width)
+        )
         copy_atom_load_W = cute.make_copy_atom(
-            cute.nvgpu.CopyUniversalOp(), mW.element_type, num_bits_per_copy=128
+            cute.nvgpu.CopyUniversalOp(), mW.element_type, num_bits_per_copy=num_bits_per_copy_W
+        )
+        num_bits_per_copy_O = cutlass.const_expr(
+            min(128, 128 // mX.element_type.width * mO.element_type.width)
         )
         copy_atom_store_O = cute.make_copy_atom(
-            cute.nvgpu.CopyUniversalOp(), mO.element_type, num_bits_per_copy=128
+            cute.nvgpu.CopyUniversalOp(), mO.element_type, num_bits_per_copy=num_bits_per_copy_O
         )
 
         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)
+        tXgW = thr_copy_X.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
+        tXgO = thr_copy_X.partition_D(gO)
+        tXrRstd = thr_copy_X.partition_D(gRstd) if cutlass.const_expr(mRstd 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)
-        tWrW.fill(0.0)
-        tXrW = thr_copy_X.retile(tWrW)
+        tXrW = cute.make_fragment_like(tXgW)
+        tXrW.fill(0.0)
         tXrX, tXrO = [cute.make_fragment_like(thr) for thr in (tXgX, tXgO)]
 
         num_warps = cute.size(tv_layout, mode=[0]) // cute.arch.WARP_SIZE
@@ -169,9 +199,9 @@ class RMSNorm(ReductionBase):
             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])
+        tXpW = utils.predicate_k(thr_copy_X.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.copy(copy_atom_load_W, tXgW, tXrW, pred=tXpW)
 
         cute.arch.cp_async_wait_group(0)
         cute.autovec_copy(tXsX, tXrX)
@@ -184,7 +214,7 @@ class RMSNorm(ReductionBase):
             reduction_buffer[None, None, 0],
             mbar_ptr,
             init_val=0.0,
-            hook_fn=cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None,
+            hook_fn=(cute.arch.cluster_wait if cutlass.const_expr(self.cluster_n > 1) else None),
         )
         rstd = utils.rsqrt(sum_sq_x / shape[1] + eps)
         if cutlass.const_expr(mRstd is not None):
@@ -196,7 +226,7 @@ class RMSNorm(ReductionBase):
             ):
                 tXrRstd[0] = rstd
         if cutlass.const_expr(delay_w_load):
-            cute.copy(copy_atom_load_W, tWgW, tWrW, pred=tWpW)
+            cute.copy(copy_atom_load_W, tXgW, tXrW, pred=tXpW)
         if cutlass.const_expr(reload_from == "smem"):
             cute.autovec_copy(tXsX, tXrX)
             x = tXrX.load().to(cute.Float32)
@@ -207,9 +237,9 @@ class RMSNorm(ReductionBase):
         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])
+        tXpO = utils.predicate_k(thr_copy_X.partition_S(cX), limit=shape[1])
         if row < shape[0]:
-            cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tOpO)
+            cute.copy(copy_atom_store_O, tXrO, tXgO, pred=tXpO)
 
 
 def _rmsnorm_fwd(
@@ -232,25 +262,36 @@ def _rmsnorm_fwd(
     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"
+    assert x.dtype in [
+        torch.float16,
+        torch.bfloat16,
+        torch.float32,
+    ], "Unsupported dtype"
+
+    assert weight.dtype in [
+        torch.float32,
+        torch.bfloat16,
+        torch.float16,
+    ], "Weight must be float32, float16 or bfloat16"
+
     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
     dtype = torch2cute_dtype_map[x.dtype]
+    # convert_from_dlpack = lambda x: (
+    #     from_dlpack(x.detach(), assumed_align=16).mark_compact_shape_dynamic(
+    #         mode=0, divisibility=128 // dtype.width
+    #     )
+    # )
     convert_from_dlpack = lambda x: (
-        from_dlpack(x.detach(), assumed_align=16).mark_compact_shape_dynamic(
-            mode=0, stride_order=(0, 1)
-        )
+        from_dlpack(x.detach(), assumed_align=16).mark_layout_dynamic(leading_dim=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)
-    ]
+    x_tensor, out_tensor = [convert_from_dlpack(t) for t in (x, out)]
+    # handle weight divisibility based on weight dtype
+    weight_dtype = torch2cute_dtype_map[weight.dtype]
     weight_tensor = utils.convert_from_dlpack(
-        weight.detach(), leading_dim=0, divisibility=128 // cutlass.Float32.width
+        weight.detach(), leading_dim=0, divisibility=128 // weight_dtype.width
     )
     rstd_tensor = (
         from_dlpack(rstd.detach(), assumed_align=4).mark_layout_dynamic(leading_dim=0)
@@ -258,7 +299,7 @@ def _rmsnorm_fwd(
         else None
     )
     current_stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
-    compile_key = (dtype, N, rstd is not None)
+    compile_key = (dtype, N, rstd is not None, weight.dtype)
     if compile_key not in _rmsnorm_fwd.compile_cache:
         rmsnorm_op = RMSNorm(dtype, N)
         _rmsnorm_fwd.compile_cache[compile_key] = cute.compile(
@@ -301,7 +342,8 @@ def rmsnorm_bwd_ref(x, w, dout, rstd, eps=1e-6):
 class RMSNormBackward(ReductionBase):
     def __init__(self, dtype: cutlass.Numeric, N: int):
         # 2 stages for double buffering when computing mean of x_hat * wdy
-        super().__init__(dtype, N, stage=2, reduction_dtype=cutlass.Float32)
+        super().__init__(dtype, N, stage=2, reduction_dtype=Float32)
+        self.reload_wdy = None if N <= 16 * 1024 else "smem"
         if self.N > 128 * 1024 and self.dtype.width >= 32:
             # Not enough smem
             raise ValueError("RMSNormBackward does not support N > 128k with dtype >= 32 bits")
@@ -348,9 +390,18 @@ class RMSNormBackward(ReductionBase):
         mRstd: cute.Tensor,
         mdX: cute.Tensor,
         mdW: cute.Tensor,
-        sm_count: cutlass.Int32,
+        sm_count: Int32,
         stream: cuda.CUstream,
     ):
+        semistatic_shape = (*mX.shape[:-1], self.N)  # Set last dimension to be statically N
+        new_stride = lambda t: (
+            cute.assume(t.stride[0], divby=128 // t.element_type.width),
+            t.stride[1],
+        )
+        mX, mdOut, mdX = [
+            cute.make_tensor(t.iterator, cute.make_layout(semistatic_shape, stride=new_stride(t)))
+            for t in (mX, mdOut, mdX)
+        ]
         self._set_cluster_n()
         tiler_mn, tv_layout = self._get_tv_layout()
         num_threads = cute.size(tv_layout, mode=[0])
@@ -412,39 +463,41 @@ class RMSNormBackward(ReductionBase):
         copy_atom_load_X_async = cute.make_copy_atom(
             cute.nvgpu.cpasync.CopyG2SOp(), mX.element_type, num_bits_per_copy=128
         )
+        num_bits_per_copy_W = cutlass.const_expr(
+            min(128, 128 // mX.element_type.width * mW.element_type.width)
+        )
         copy_atom_load_W = cute.make_copy_atom(
-            cute.nvgpu.CopyUniversalOp(), mW.element_type, num_bits_per_copy=128
+            cute.nvgpu.CopyUniversalOp(), mW.element_type, num_bits_per_copy=num_bits_per_copy_W
+        )
+        num_bits_per_copy_dX = cutlass.const_expr(
+            min(128, 128 // mX.element_type.width * mdX.element_type.width)
         )
         copy_atom_store_dX = cute.make_copy_atom(
-            cute.nvgpu.CopyUniversalOp(), mdX.element_type, num_bits_per_copy=128
+            cute.nvgpu.CopyUniversalOp(), mdX.element_type, num_bits_per_copy=num_bits_per_copy_dX
+        )
+        num_bits_per_copy_dW = cutlass.const_expr(
+            min(128, 128 // mX.element_type.width * mdW.element_type.width)
         )
         copy_atom_store_dW = cute.make_copy_atom(
-            cute.nvgpu.CopyUniversalOp(), mdW.element_type, num_bits_per_copy=128
+            cute.nvgpu.CopyUniversalOp(), mdW.element_type, num_bits_per_copy=num_bits_per_copy_dW
         )
 
         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_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_store_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, (0, cluster_y))
-        tWgW = thr_copy_W.partition_S(gW)
-        tWrW = cute.make_fragment_like(tWgW)
+        tXgW = thr_copy_X.partition_S(gW)
+        tXrW = cute.make_fragment_like(tXgW)
         # Need this, otherwise rW can have arbitrary values that changes the reduction
         if not is_even_N:
-            tWrW.fill(0.0)
-        tXrW = thr_copy_X.retile(tWrW)
+            tXrW.fill(0.0)
 
         gW_coord = cute.local_tile(idX, tiler_mn, (0, cluster_y))
-        tWpW = (
-            utils.predicate_k(thr_copy_W.partition_S(gW_coord), limit=shape[1])
+        tXpW = (
+            utils.predicate_k(thr_copy_X.partition_S(gW_coord), limit=shape[1])
             if not is_even_N
             else None
         )
-        cute.copy(copy_atom_load_W, tWgW, tWrW, pred=tWpW)
+        cute.copy(copy_atom_load_W, tXgW, tXrW, pred=tXpW)
         weight = tXrW.load().to(cute.Float32)
 
         num_warps = cute.size(tv_layout, mode=[0]) // cute.arch.WARP_SIZE
@@ -452,16 +505,16 @@ class RMSNormBackward(ReductionBase):
         self._initialize_cluster(tidx, mbar_ptr, num_warps, is_persistent=True)
 
         dw_coord = cute.local_tile(idX, tiler_mn, (0, cluster_y))
-        tdWpdW = (
-            utils.predicate_k(thr_copy_dW.partition_S(dw_coord), limit=shape[1])
+        tXpdW = (
+            utils.predicate_k(thr_copy_X.partition_S(dw_coord), limit=shape[1])
             if not is_even_N
             else None
         )
 
         gdW = cute.local_tile(mdW, (1, tiler_mn[1]), (bidx_start, cluster_y))
-        tdWgdW = thr_copy_dW.partition_D(gdW)
-        tdWrdW = cute.make_fragment_like(tdWgdW, cutlass.Float32)
-        tXrdW = thr_copy_X.retile(tdWrdW)
+        tXgdW = thr_copy_X.partition_S(gdW)
+        # Always compute partial weight gradients in fp32
+        tXrdW = cute.make_fragment_like(tXgdW, Float32)
 
         gX = cute.local_tile(mX, tiler_mn, (None, cluster_y))
         gdOut = cute.local_tile(mdOut, tiler_mn, (None, cluster_y))
@@ -471,7 +524,7 @@ class RMSNormBackward(ReductionBase):
         tXsX = thr_copy_X.partition_D(sX)
         tXgdOut = thr_copy_X.partition_S(gdOut)
         tXsdOut = thr_copy_X.partition_D(sdOut)
-        tXgdX = thr_store_dX.partition_D(gdX)
+        tXgdX = thr_copy_X.partition_D(gdX)
         tXcX = thr_copy_X.partition_S(cX)[(0, None), None, None, None]
         # This doesn't change across iterations
         tXpX = (
@@ -513,9 +566,9 @@ class RMSNormBackward(ReductionBase):
 
         threads_per_row = tv_layout.shape[0][0]
         tXrdW.fill(0.0)
-        stage = cutlass.Int32(0)
-        producer_phase = cutlass.Int32(1)
-        consumer_phase = cutlass.Int32(0)
+        stage = Int32(0)
+        producer_phase = Int32(1)
+        consumer_phase = Int32(0)
         for bidx in cutlass.range(bidx_start, cute.ceil_div(M, tiler_mn[0]), gdim):
             row = tXcX[None, None, None, bidx][0][0]
             rstd = cutlass.Float.zero
@@ -538,10 +591,14 @@ class RMSNormBackward(ReductionBase):
                 )
             elif tiler_mn[0] > 1:
                 utils.fill_oob(
-                    tXsX[None, None, None, stage ^ 1], None, fill_value=mX.element_type.zero
+                    tXsX[None, None, None, stage ^ 1],
+                    None,
+                    fill_value=mX.element_type.zero,
                 )
                 utils.fill_oob(
-                    tXsdOut[None, None, None, stage ^ 1], None, fill_value=mdOut.element_type.zero
+                    tXsdOut[None, None, None, stage ^ 1],
+                    None,
+                    fill_value=mdOut.element_type.zero,
                 )
             cute.arch.cp_async_commit_group()
             if row < M or tiler_mn[0] == 1:
@@ -561,12 +618,13 @@ class RMSNormBackward(ReductionBase):
                     cute.ReductionOp.ADD,
                     threads_per_row,
                     reduction_buffer[None, None, stage],
-                    mbar_full_ptr + stage if cutlass.const_expr(self.cluster_n > 1) else None,
+                    (mbar_full_ptr + stage if cutlass.const_expr(self.cluster_n > 1) else None),
                     phase=consumer_phase,
                     init_val=0.0,
                 )
                 / shape[1]
             )
+
             if cutlass.const_expr(self.cluster_n > 1):
                 # It's faster to have 1 lane per warp to signal the mbar, rather than all lanes
                 # Requires adjusting the thread_count when initializing the mbar
@@ -576,12 +634,20 @@ class RMSNormBackward(ReductionBase):
                     cute.arch.mbarrier_arrive(
                         mbar_empty_ptr + stage, peer_cta_rank_in_cluster=lane_idx
                     )
+
+            if cutlass.const_expr(self.reload_wdy == "smem"):
+                cute.autovec_copy(tXsdOut[None, None, None, stage], tXrdOut)
+                dout = tXrdOut.load().to(cute.Float32)
+                wdy = dout * weight
+
             dx = (wdy - x_hat * mean_xhat_wdy) * rstd
             tXrdX.store(dx.to(tXrdOut.element_type))
             if row < M or tiler_mn[0] == 1:
                 tXgdX_cur = utils.coord_offset_i64(bidx, tXgdX, dim=3)[None, None, None, 0]
                 cute.copy(copy_atom_store_dX, tXrdX, tXgdX_cur, pred=tXpX)
+            # Accumulate weight gradients in fp32
             tXrdW.store(tXrdW.load() + dout * x_hat)
+
             stage ^= 1
             if stage == 0:
                 consumer_phase ^= 1
@@ -608,9 +674,10 @@ class RMSNormBackward(ReductionBase):
                     tXsdW_other = cute.make_tensor(tXsdW.iterator + i * sdW.stride[0], tXsdW.layout)
                     cute.autovec_copy(tXsdW_other, tXrdW_other)
                     tXrdW.store(tXrdW.load() + tXrdW_other.load())
-                cute.copy(copy_atom_store_dW, tdWrdW, tdWgdW, pred=tdWpdW)
+                cute.copy(copy_atom_store_dW, tXrdW, tXgdW, pred=tXpdW)
         else:
-            cute.copy(copy_atom_store_dW, tdWrdW, tdWgdW, pred=tdWpdW)
+            # dw is already in fp32, so we can directly copy to global memory
+            cute.copy(copy_atom_store_dW, tXrdW, tXgdW, pred=tXpdW)
 
 
 def _rmsnorm_backward(
@@ -634,8 +701,17 @@ def _rmsnorm_backward(
     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"
+    assert x.dtype in [
+        torch.float16,
+        torch.bfloat16,
+        torch.float32,
+    ], "Unsupported dtype"
+
+    assert weight.dtype in [
+        torch.float32,
+        torch.bfloat16,
+        torch.float16,
+    ], "Weight must be float32, float16 or bfloat16"
 
     M, N = x.shape
     dx = torch.empty_like(x)
@@ -654,28 +730,29 @@ def _rmsnorm_backward(
     sm_count = (
         sm_count * sm_count_multiple if N <= 8192 else sm_count // 2 if N <= 16384 else sm_count * 2
     )
-    dw_partial = torch.empty(sm_count, N, device=device, dtype=weight.dtype)
+
+    # Always store partial gradients in fp32 for numerical accuracy
+    dw_partial = torch.empty(sm_count, N, device=device, dtype=torch.float32)
 
     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)
-        )
+    convert_from_dlpack = lambda x: (
+        from_dlpack(x.detach(), assumed_align=16).mark_layout_dynamic(leading_dim=1)
     )
-
     x_tensor, dout_tensor, dx_tensor = [convert_from_dlpack(tensor) for tensor in (x, dout, dx)]
 
+    # Handle weight div based on weight dtype
+    weight_dtype = torch2cute_dtype_map[weight.dtype]
     weight_tensor = utils.convert_from_dlpack(
-        weight.detach(), leading_dim=0, divisibility=128 // cutlass.Float32.width
+        weight.detach(), leading_dim=0, divisibility=128 // weight_dtype.width
     )
 
-    dw_partial_tensor = convert_from_dlpack(dw_partial)
+    dw_partial_tensor = from_dlpack(dw_partial, assumed_align=16).mark_compact_shape_dynamic(mode=0)
     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)
+    compile_key = (dtype, N, weight.dtype)
     if compile_key not in _rmsnorm_backward.compile_cache:
         rmsnorm_backward_op = RMSNormBackward(dtype, N)
         _rmsnorm_backward.compile_cache[compile_key] = cute.compile(
@@ -700,7 +777,7 @@ def _rmsnorm_backward(
         sm_count,
         current_stream,
     )
-
+    # we have summed the partial gradients in fp32, now we convert back to the weight dtype
     dw = dw_partial.sum(dim=0).to(weight.dtype)
     return dx, dw
 
@@ -711,16 +788,29 @@ _rmsnorm_backward.compile_cache = {}
 class RMSNormFunction(torch.autograd.Function):
     @staticmethod
     def forward(ctx, x, weight, eps):
+        x_shape_start = x.shape
+
+        # Flatten input
+        x = x.view(-1, x.shape[-1])
+
         out, rstd = _rmsnorm_fwd(x, weight, eps, return_rstd=True)
         ctx.save_for_backward(x, weight, rstd)
         ctx.eps = eps
-        return out
+        ctx.x_shape_start = x_shape_start
+
+        return out.reshape(x_shape_start)
 
     @staticmethod
     def backward(ctx, dout):
         x, weight, rstd = ctx.saved_tensors
+        x_shape_start = ctx.x_shape_start
+        # Reshape dout to match the flattened shape used in forward
+        dout = dout.view(-1, dout.shape[-1])
         dx, dw = _rmsnorm_backward(x, weight, dout, rstd)
-        # dw is returned for weight gradient, None for eps gradient
+        dx = dx.view(x_shape_start)
+        # dx is returned for input gradient,
+        # dw is returned for weight gradient,
+        # None for eps gradient
         return dx, dw, None
 
 
@@ -736,3 +826,39 @@ def rmsnorm(x: torch.Tensor, weight: torch.Tensor, eps: float = 1e-6) -> torch.T
         Normalized output tensor of same shape as x
     """
     return RMSNormFunction.apply(x, weight, eps)
+
+
+class QuackRMSNorm(torch.nn.Module):
+    """RMSNorm module that behaves like torch.nn.RMSNorm.
+
+    This class provides a drop-in replacement for torch.nn.RMSNorm that uses
+    the quack.rmsnorm implementation under the hood.
+
+    Args:
+        dim (int): The dimension to normalize over
+        eps (float, optional): A small constant for numerical stability. Default: 1e-6
+
+    Attributes:
+        weight (torch.nn.Parameter): The learnable weight parameter
+        eps (float): A small constant for numerical stability
+    """
+
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = torch.nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Apply RMSNorm to the input tensor.
+
+        Args:
+            x (torch.Tensor): Input tensor
+
+        Returns:
+            torch.Tensor: Normalized tensor
+        """
+        return rmsnorm(x, self.weight, self.eps)
+
+    def reset_parameters(self):
+        """Reset the weight parameter to ones."""
+        torch.nn.init.ones_(self.weight)

{quack_kernels-0.1.7.dist-info → quack_kernels-0.1.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: quack-kernels
-Version: 0.1.7
+Version: 0.1.9
 Requires-Python: >=3.9
 License-File: LICENSE
 Requires-Dist: nvidia-cutlass-dsl==4.1.0.dev0

quack_kernels-0.1.9.dist-info/RECORD

@@ -0,0 +1,12 @@
+quack/__init__.py,sha256=CT76CeRNh5bzQ9f13yVuRz9Sj7V3MvwzHH4fB1iQIf0,203
+quack/cross_entropy.py,sha256=VYSAd28GmtnMoKQwLrorvySDtJfRhoqVd-aeM52FmsI,20866
+quack/layernorm.py,sha256=1WUspbr6ktPZ25O00kKs-FK_lm_Fejat72BMV8tBSfw,13504
+quack/reduction_base.py,sha256=4nAzkZR1yoQVA4Lc-GpU0XMjS5ARAmvYdeE0Doy7UCU,3789
+quack/rmsnorm.py,sha256=bJEHqc8ila-LTGco-tNNCUyFBjJ2UdXeoMplYNJPXFI,32740
+quack/softmax.py,sha256=3-5P_ORBrfQ6JYTIzgDs9jwmV7Za73SogaX7q9M7GCM,16698
+quack/utils.py,sha256=aiyzBc9BEwq8s965elfiR331hAaLLBKL9kDHjuls86Q,17791
+quack_kernels-0.1.9.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+quack_kernels-0.1.9.dist-info/METADATA,sha256=vOnpbShNHRiUXKAnOUxzfRM7zkpW3RmjW4hIgvYda08,289
+quack_kernels-0.1.9.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+quack_kernels-0.1.9.dist-info/top_level.txt,sha256=6e4Jr_vNJbZTYwlO_Ahf_sDeHDE0zcqcf7Le11FKxxo,6
+quack_kernels-0.1.9.dist-info/RECORD,,

quack_kernels-0.1.7.dist-info/RECORD

@@ -1,12 +0,0 @@
-quack/__init__.py,sha256=R9cZd_vslI5oZjjS-ojfWAd9tCZAqsLUiFVqEbUaGnw,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=4nAzkZR1yoQVA4Lc-GpU0XMjS5ARAmvYdeE0Doy7UCU,3789
-quack/rmsnorm.py,sha256=3jiwWhVmaG0n5vuUnGGrpg3StAB4lnzziNF97QVMLGQ,28870
-quack/softmax.py,sha256=3-5P_ORBrfQ6JYTIzgDs9jwmV7Za73SogaX7q9M7GCM,16698
-quack/utils.py,sha256=aiyzBc9BEwq8s965elfiR331hAaLLBKL9kDHjuls86Q,17791
-quack_kernels-0.1.7.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-quack_kernels-0.1.7.dist-info/METADATA,sha256=9RlqUmX3-7BI2aZk88r84B8o2FzZkQgkfV1UxwN8GlE,289
-quack_kernels-0.1.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-quack_kernels-0.1.7.dist-info/top_level.txt,sha256=6e4Jr_vNJbZTYwlO_Ahf_sDeHDE0zcqcf7Le11FKxxo,6
-quack_kernels-0.1.7.dist-info/RECORD,,