liger-kernel-nightly 0.5.6.dev20250403190551__py3-none-any.whl → 0.6.4.dev20251212103629__py3-none-any.whl

liger_kernel/ops/rms_norm.py

@@ -21,8 +21,10 @@ from liger_kernel.ops.utils import calculate_settings
 from liger_kernel.ops.utils import compare_version
 from liger_kernel.ops.utils import ensure_contiguous
 from liger_kernel.ops.utils import torch_to_triton_dtype
+from liger_kernel.utils import get_npu_multi_processor_count
+from liger_kernel.utils import is_npu_available
 
-if compare_version("triton", operator.ge, "3.0.0"):
+if compare_version("triton", operator.ge, "3.0.0") and not is_npu_available():
     try:
         # typical import path with dispatch available
         from triton.language.extra.libdevice import rsqrt
@@ -63,7 +65,7 @@ def _rms_norm_forward_kernel(
     3. https://arxiv.org/pdf/1910.07467
     """
 
-    row_idx = tl.program_id(0)
+    row_idx = tl.program_id(0).to(tl.int64)
     col_offsets = tl.arange(0, BLOCK_SIZE)
     mask = col_offsets < n_cols
 
@@ -137,7 +139,7 @@ def _rms_norm_backward_kernel(
     dw = sum(dy * (x / RMS)). summation over BxT dimension
     """
 
-    row_block_id = tl.program_id(0)
+    row_block_id = tl.program_id(0).to(tl.int64)
     row_start = row_block_id * rows_per_program
     row_end = min((row_block_id + 1) * rows_per_program, n_rows)
     col_offsets = tl.arange(0, BLOCK_SIZE)
@@ -194,6 +196,176 @@ def _rms_norm_backward_kernel(
     tl.store(dW_ptr + row_block_id * dW_row_stride + col_offsets, dW_row, mask=mask)
 
 
+@triton.jit
+def _block_rms_norm_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    W_ptr,
+    W_row_stride,
+    RSTD_ptr,
+    RSTD_row_stride,
+    n_rows,
+    n_cols,
+    eps,
+    offset,
+    casting_mode: tl.constexpr,  # constexpr so the `if` blocks can be optimized out
+    BLOCK_SIZE: tl.constexpr,
+    BLOCK_ROW: tl.constexpr,
+):
+    """
+    y_i = (x_i / (RMS)) * (offset + wi), RMS = sqrt(sum(x_i^2) / N)
+
+    Reference:
+    1. https://triton-lang.org/main/getting-started/tutorials/05-layer-norm.html
+    2. https://github.com/unslothai/unsloth/blob/fd753fed99ed5f10ef8a9b7139588d9de9ddecfb/unsloth/kernels/rms_layernorm.py#L22
+    3. https://arxiv.org/pdf/1910.07467
+    """
+
+    row_idx = tl.program_id(0) * BLOCK_ROW + tl.arange(0, BLOCK_ROW)
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    row_mask = row_idx < n_rows
+    col_mask = col_offsets < n_cols
+
+    X_row = tl.load(
+        X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :],
+        mask=row_mask[:, None] & col_mask[None, :],
+        other=0,
+    )
+    X_row_dtype = X_row.dtype
+    W_row = tl.load(W_ptr + col_offsets, mask=col_mask, other=0)
+
+    # On Llama, only rstd is computed on fp32
+    if casting_mode == _CASTING_MODE_LLAMA:
+        X_row = X_row.to(tl.float32)
+
+    # Gemma computes everything on fp32, and then casts back the output to the original dtype
+    if casting_mode == _CASTING_MODE_GEMMA:
+        W_row = W_row.to(tl.float32)
+        X_row = X_row.to(tl.float32)
+
+    if casting_mode == _CASTING_MODE_NONE:
+        eps = eps.to(X_row_dtype)
+        offset = offset.to(X_row_dtype)
+
+    mean_square = tl.sum(X_row * X_row, axis=1) / n_cols
+    rstd = rsqrt(mean_square + eps)
+
+    # We can save time by caching rms with minimal memory overhead
+    # because rms is much smaller compared to X_row, as rms is for each row.
+    # However, on the computation side, it can save 4 operations (*, sum, /, sqrt).
+    tl.store(RSTD_ptr + row_idx * RSTD_row_stride, rstd, row_mask)
+
+    X_row = X_row * rstd[:, None]
+
+    # On Llama, the multiplication with the weight is done on the original dtype
+    if casting_mode == _CASTING_MODE_LLAMA:
+        X_row = X_row.to(X_row_dtype)
+
+    Y_row = X_row * (offset + W_row)[None, :]
+
+    if casting_mode == _CASTING_MODE_GEMMA:
+        Y_row = Y_row.to(X_row_dtype)
+
+    tl.store(
+        Y_ptr + row_idx[:, None] * Y_row_stride + col_offsets[None, :],
+        Y_row,
+        mask=row_mask[:, None] & col_mask[None, :],
+    )
+
+
+@triton.jit
+def _block_rms_norm_backward_kernel(
+    dY_ptr,
+    dY_row_stride,
+    dX_ptr,
+    dX_row_stride,
+    X_ptr,
+    X_row_stride,
+    X_dtype: tl.constexpr,
+    W_ptr,
+    W_row_stride,
+    RSTD_ptr,
+    RSTD_row_stride,
+    dW_ptr,
+    dW_row_stride,
+    n_rows,
+    n_cols,
+    offset,
+    rows_per_program: tl.constexpr,
+    casting_mode: tl.constexpr,
+    BLOCK_SIZE: tl.constexpr,
+    BLOCK_ROW: tl.constexpr,
+):
+    """
+    dx = (1 / RMS) * [dy * (w + offset - (1 / N) * (1 / RMS^2) * ((dy * (w + offset)) dot x) * x]. * means element-wise multiplication, whileas dot means dot product
+    dw = sum(dy * (x / RMS)). summation over BxT dimension
+    """
+
+    pid = tl.program_id(0).cast(tl.int64)
+    NUM_SMS = tl.num_programs(0)
+
+    col_offsets = tl.arange(0, BLOCK_SIZE)
+    col_mask = col_offsets < n_cols
+
+    dW_row = tl.zeros((BLOCK_SIZE,), dtype=tl.float32)
+
+    W_row = tl.load(W_ptr + col_offsets, mask=col_mask, other=0.0)
+    W_row = W_row + offset
+
+    for start in range(pid * BLOCK_ROW, n_rows, NUM_SMS * BLOCK_ROW):
+        row_idx = start + tl.arange(0, BLOCK_ROW)
+        row_mask = row_idx < n_rows
+        dY_row = tl.load(
+            dY_ptr + row_idx[:, None] * dY_row_stride + col_offsets[None, :],
+            mask=row_mask[:, None] & col_mask[None, :],
+            other=0.0,
+        )
+        X_row = tl.load(
+            X_ptr + row_idx[:, None] * X_row_stride + col_offsets[None, :],
+            mask=row_mask[:, None] & col_mask[None, :],
+            other=0.0,
+        )
+
+        # Get cached rms
+        rstd_row = tl.load(RSTD_ptr + row_idx * RSTD_row_stride, row_mask)
+
+        X_row = X_row.to(tl.float32)
+
+        # Different bacward graphs for different casting modes
+        if casting_mode == _CASTING_MODE_LLAMA:
+            m = (dY_row * W_row[None, :]).to(tl.float32)
+
+        elif casting_mode == _CASTING_MODE_GEMMA:
+            dY_row = dY_row.to(tl.float32)
+            m = dY_row * W_row[None, :]
+        else:
+            m = dY_row * W_row[None, :]
+
+        dX_row = rstd_row[:, None] * m
+
+        dX_row += (rstd_row[:, None]) * (
+            -(1 / n_cols) * (rstd_row * rstd_row * tl.sum(m * X_row, axis=1))[:, None] * X_row
+        )
+
+        # calculate the gradient of W
+        if casting_mode == _CASTING_MODE_LLAMA:
+            # TODO(tcc): use tl.sum(..., dtype=tl.float32) once we upgrade to triton>=3.3.0
+            dW_row += tl.sum((dY_row * (X_row * rstd_row[:, None]).to(X_dtype)).to(tl.float32), 0)
+        else:
+            # here X_row is already in fp32 (see previous if block)
+            dW_row += tl.sum(dY_row * (X_row * rstd_row[:, None]), 0)
+
+        tl.store(
+            dX_ptr + row_idx[:, None] * dX_row_stride + col_offsets[None, :],
+            dX_row,
+            mask=row_mask[:, None] & col_mask[None, :],
+        )
+
+    tl.store(dW_ptr + pid * dW_row_stride + col_offsets, dW_row, mask=col_mask)
+
+
 _str_to_casting_mode = {
     "llama": _CASTING_MODE_LLAMA.value,
     "gemma": _CASTING_MODE_GEMMA.value,
@@ -201,7 +373,7 @@ _str_to_casting_mode = {
 }
 
 
-def rms_norm_forward(X, W, eps, offset, casting_mode):
+def rms_norm_forward(X, W, eps, offset, casting_mode, row_mode):
     if not isinstance(casting_mode, int):
         assert casting_mode in _str_to_casting_mode, f"Invalid casting mode: {casting_mode}"
         casting_mode = _str_to_casting_mode[casting_mode]
@@ -223,26 +395,53 @@ def rms_norm_forward(X, W, eps, offset, casting_mode):
     # Check constraints.
     assert X.shape[1] == W.shape[0], "Incompatible hidden size dimension between tensor1.shape[1] and tensor2.shape[0]"
 
-    _rms_norm_forward_kernel[(n_rows,)](
-        Y,
-        Y.stride(0),
-        X,
-        X.stride(0),
-        W,
-        W.stride(0),
-        RSTD,
-        RSTD.stride(0),
-        n_cols,
-        eps,
-        offset,
-        casting_mode,
-        BLOCK_SIZE=BLOCK_SIZE,
-        num_warps=num_warps,
-    )
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+    if BLOCK_SIZE > 256 or n_rows < 4096 * 8 or row_mode:
+        _rms_norm_forward_kernel[(n_rows,)](
+            Y,
+            Y.stride(0),
+            X,
+            X.stride(0),
+            W,
+            W.stride(0),
+            RSTD,
+            RSTD.stride(0),
+            n_cols,
+            eps,
+            offset,
+            casting_mode,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
+    else:
+        BLOCK_ROW = 16
+        kernel_args["BLOCK_ROW"] = BLOCK_ROW
+        _block_rms_norm_forward_kernel[(triton.cdiv(n_rows, BLOCK_ROW),)](
+            Y,
+            Y.stride(0),
+            X,
+            X.stride(0),
+            W,
+            W.stride(0),
+            RSTD,
+            RSTD.stride(0),
+            n_rows,
+            n_cols,
+            eps,
+            offset,
+            casting_mode,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
     return Y.view(*shape), X, RSTD, BLOCK_SIZE, num_warps, casting_mode
 
 
-def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warps, in_place):
+def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warps, in_place, row_mode):
     shape = dY.shape
     dim = shape[-1]
     dY = dY.view(-1, dim)
@@ -252,7 +451,9 @@ def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warp
     if X.device.type == "cuda":
         sm_count = torch.cuda.get_device_properties(X.device).multi_processor_count
     elif X.device.type == "xpu":
-        sm_count = torch.xpu.get_device_properties(X.device).gpu_subslice_count
+        sm_count = torch.xpu.get_device_properties(X.device).gpu_eu_count
+    elif X.device.type == "npu":
+        sm_count = get_npu_multi_processor_count()
 
     # fp32 for numerical stability especially.
     _dW = torch.empty((sm_count, n_cols), dtype=torch.float32, device=W.device)
@@ -267,28 +468,61 @@ def rms_norm_backward(dY, X, W, RSTD, offset, casting_mode, BLOCK_SIZE, num_warp
     else:
         dX = torch.zeros_like(dY)
 
-    _rms_norm_backward_kernel[grid](
-        dY,
-        dY.stride(0),
-        dX,
-        dX.stride(0),
-        X,
-        X.stride(0),
-        torch_to_triton_dtype[X.dtype],
-        W,
-        W.stride(0),
-        RSTD,
-        RSTD.stride(0),
-        _dW,
-        _dW.stride(0),
-        n_rows,
-        n_cols,
-        offset,
-        rows_per_program,
-        casting_mode,
-        BLOCK_SIZE=BLOCK_SIZE,
-        num_warps=num_warps,
-    )
+    # XPU-specific optimization
+    kernel_args = {}
+    if X.device.type == "xpu":
+        kernel_args["grf_mode"] = "large"
+
+    if BLOCK_SIZE > 256 or n_rows < 4096 * 8 or row_mode:
+        _rms_norm_backward_kernel[grid](
+            dY,
+            dY.stride(0),
+            dX,
+            dX.stride(0),
+            X,
+            X.stride(0),
+            torch_to_triton_dtype[X.dtype],
+            W,
+            W.stride(0),
+            RSTD,
+            RSTD.stride(0),
+            _dW,
+            _dW.stride(0),
+            n_rows,
+            n_cols,
+            offset,
+            rows_per_program,
+            casting_mode,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
+    else:
+        BLOCK_ROW = 16
+        kernel_args["BLOCK_ROW"] = BLOCK_ROW
+        _block_rms_norm_backward_kernel[grid](
+            dY,
+            dY.stride(0),
+            dX,
+            dX.stride(0),
+            X,
+            X.stride(0),
+            torch_to_triton_dtype[X.dtype],
+            W,
+            W.stride(0),
+            RSTD,
+            RSTD.stride(0),
+            _dW,
+            _dW.stride(0),
+            n_rows,
+            n_cols,
+            offset,
+            rows_per_program,
+            casting_mode,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+            **kernel_args,  # XPU-specific optimization
+        )
     dX = dX.view(*shape)
     dW = _dW.sum(dim=0).to(W.dtype)
 
@@ -319,15 +553,16 @@ class LigerRMSNormFunction(torch.autograd.Function):
 
     @staticmethod
     @ensure_contiguous
-    def forward(ctx, X, W, eps, offset=0.0, casting_mode="llama", in_place=True):
+    def forward(ctx, X, W, eps, offset=0.0, casting_mode="llama", in_place=True, row_mode=None):
         """
         X: (B, T, H) or (BxT, H)
         W: (H,)
         """
-        Y, X, RSTD, BLOCK_SIZE, num_warps, casting_mode = rms_norm_forward(X, W, eps, offset, casting_mode)
+        Y, X, RSTD, BLOCK_SIZE, num_warps, casting_mode = rms_norm_forward(X, W, eps, offset, casting_mode, row_mode)
         ctx.offset = offset
         ctx.casting_mode = casting_mode
         ctx.in_place = in_place
+        ctx.row_mode = row_mode
         ctx.BLOCK_SIZE = BLOCK_SIZE
         ctx.num_warps = num_warps
         ctx.save_for_backward(X, W, RSTD)
@@ -341,14 +576,6 @@ class LigerRMSNormFunction(torch.autograd.Function):
         """
         X, W, RSTD = ctx.saved_tensors
         dX, dW = rms_norm_backward(
-            dY,
-            X,
-            W,
-            RSTD,
-            ctx.offset,
-            ctx.casting_mode,
-            ctx.BLOCK_SIZE,
-            ctx.num_warps,
-            ctx.in_place,
+            dY, X, W, RSTD, ctx.offset, ctx.casting_mode, ctx.BLOCK_SIZE, ctx.num_warps, ctx.in_place, ctx.row_mode
         )
-        return dX, dW, None, None, None, None
+        return dX, dW, None, None, None, None, None

liger_kernel/ops/rope.py

@@ -32,7 +32,7 @@ def _triton_rope(
 
     # cos size: (1, seq_len, head_dim) or (bsz, seq_len, head_dim)
     # stride: (seq_len * head_dim, head_dim, 1)
-    pid = tl.program_id(0)
+    pid = tl.program_id(0).to(tl.int64)
 
     # locate start address
     q_ptr = q_ptr + pid * q_row_stride

liger_kernel/ops/softmax.py

@@ -0,0 +1,201 @@
+from typing import Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from liger_kernel.ops.utils import calculate_settings
+from liger_kernel.ops.utils import ensure_contiguous
+
+
+@triton.jit
+def _softmax_single_block_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+    mask = offs < n_cols
+
+    x = tl.load(X_ptr + row_id * X_row_stride + offs, mask=mask, other=-float("inf"), cache_modifier=".ca")
+    m = tl.max(x, axis=0)
+    e = tl.exp(x - m)
+    d = tl.sum(e, axis=0)
+    y = e / d
+    tl.store(Y_ptr + row_id * Y_row_stride + offs, y, mask=mask, cache_modifier=".cs")
+
+
+@triton.jit
+def _softmax_multi_block_forward_kernel(
+    Y_ptr,
+    Y_row_stride,
+    X_ptr,
+    X_row_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+
+    m = tl.float32(-float("inf"))
+    d = tl.float32(0.0)
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        xblk = tl.load(X_ptr + row_id * X_row_stride + idx, mask=mask, other=-float("inf"), cache_modifier=".ca")
+        blk_max = tl.max(xblk, axis=0)
+        new_m = tl.max(m, blk_max)
+        d = d * tl.exp(m - new_m) + tl.sum(tl.exp(xblk - new_m), axis=0)
+        m = new_m
+
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        xblk = tl.load(X_ptr + row_id * X_row_stride + idx, mask=mask, other=-float("inf"), cache_modifier=".ca")
+        yblk = tl.exp(xblk - m) / d
+        tl.store(Y_ptr + row_id * Y_row_stride + idx, yblk, mask=mask, cache_modifier=".cs")
+
+
+@triton.jit
+def _softmax_single_block_backward_kernel(
+    dy_ptr,
+    dy_stride,
+    y_ptr,
+    y_stride,
+    dx_ptr,
+    dx_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+    mask = offs < n_cols
+
+    dy = tl.load(dy_ptr + row_id * dy_stride + offs, mask=mask, other=0.0)
+    y = tl.load(y_ptr + row_id * y_stride + offs, mask=mask, other=0.0, cache_modifier=".ca")
+    dot = tl.sum(dy * y, axis=0)
+    dx = y * (dy - dot)
+    tl.store(dx_ptr + row_id * dx_stride + offs, dx, mask=mask, cache_modifier=".wb")
+
+
+@triton.jit
+def _softmax_multi_block_backward_kernel(
+    dy_ptr,
+    dy_stride,
+    y_ptr,
+    y_stride,
+    dx_ptr,
+    dx_stride,
+    n_cols,
+    BLOCK_SIZE: tl.constexpr,
+):
+    row_id = tl.program_id(0)
+    offs = tl.arange(0, BLOCK_SIZE)
+    acc = tl.float32(0.0)
+
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        dy_blk = tl.load(dy_ptr + row_id * dy_stride + idx, mask=mask, other=0.0)
+        y_blk = tl.load(y_ptr + row_id * y_stride + idx, mask=mask, other=0.0, cache_modifier=".ca")
+        acc += tl.sum(dy_blk * y_blk, axis=0)
+
+    for start in tl.range(0, n_cols, BLOCK_SIZE):
+        idx = start + offs
+        mask = idx < n_cols
+        dy_blk = tl.load(dy_ptr + row_id * dy_stride + idx, mask=mask, other=0.0)
+        y_blk = tl.load(y_ptr + row_id * y_stride + idx, mask=mask, other=0.0, cache_modifier=".ca")
+        dx_blk = y_blk * (dy_blk - acc)
+        tl.store(dx_ptr + row_id * dx_stride + idx, dx_blk, mask=mask, cache_modifier=".wb")
+
+
+def _softmax_forward(x: torch.Tensor) -> Tuple[torch.Tensor, int, int, bool]:
+    *batch, n_cols = x.shape
+    x2d = x.contiguous().view(-1, n_cols)
+    n_rows = x2d.shape[0]
+
+    BLOCK_SIZE, num_warps = calculate_settings(n_cols)
+    y2d = torch.empty_like(x2d)
+
+    if n_cols <= BLOCK_SIZE:
+        _softmax_single_block_forward_kernel[(n_rows,)](
+            y2d, y2d.stride(0), x2d, x2d.stride(0), n_cols, BLOCK_SIZE=BLOCK_SIZE, num_warps=num_warps
+        )
+        multi_block_launch = False
+    else:
+        _softmax_multi_block_forward_kernel[(n_rows,)](
+            y2d, y2d.stride(0), x2d, x2d.stride(0), n_cols, BLOCK_SIZE=BLOCK_SIZE, num_warps=num_warps
+        )
+        multi_block_launch = True
+
+    return y2d.view(*batch, n_cols), BLOCK_SIZE, num_warps, multi_block_launch
+
+
+def _softmax_backward(
+    dy: torch.Tensor,
+    y: torch.Tensor,
+    BLOCK_SIZE: int,
+    num_warps: int,
+    multi_block_launch: bool,
+) -> torch.Tensor:
+    *batch, n_cols = dy.shape
+    dy2d = dy.contiguous().view(-1, n_cols)
+    y2d = y.contiguous().view(-1, n_cols)
+    n_rows = dy2d.shape[0]
+    dx2d = torch.empty_like(dy2d)
+
+    if not multi_block_launch and n_cols <= BLOCK_SIZE:
+        _softmax_single_block_backward_kernel[(n_rows,)](
+            dy2d,
+            dy2d.stride(0),
+            y2d,
+            y2d.stride(0),
+            dx2d,
+            dx2d.stride(0),
+            n_cols,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+        )
+    else:
+        _softmax_multi_block_backward_kernel[(n_rows,)](
+            dy2d,
+            dy2d.stride(0),
+            y2d,
+            y2d.stride(0),
+            dx2d,
+            dx2d.stride(0),
+            n_cols,
+            BLOCK_SIZE=BLOCK_SIZE,
+            num_warps=num_warps,
+        )
+
+    return dx2d.view(*batch, n_cols)
+
+
+class LigerSoftmaxFunction(torch.autograd.Function):
+    @staticmethod
+    @ensure_contiguous
+    def forward(ctx, input_: torch.Tensor):
+        y, BLOCK_SIZE, num_warps, multi_block_launch = _softmax_forward(input_)
+        ctx.save_for_backward(y)
+        ctx.BLOCK_SIZE = BLOCK_SIZE
+        ctx.num_warps = num_warps
+        ctx.multi_block_launch = multi_block_launch
+        return y
+
+    @staticmethod
+    @ensure_contiguous
+    def backward(ctx, grad_output):
+        (y,) = ctx.saved_tensors
+        dx = _softmax_backward(
+            grad_output,
+            y,
+            ctx.BLOCK_SIZE,
+            ctx.num_warps,
+            ctx.multi_block_launch,
+        )
+        return dx