Elementwise Operators

Every op on this page is used the same way: construct it once, then call it. The constructor takes what the kernel is compiled with; the call takes the tensors. Both are documented under each op — __init__ and forward, where forward is what runs when you call op(...).

Arithmetic

tileops.ops.elementwise.arithmetic.AddFwdOp

Element-wise addition with broadcast: y = input + alpha * other.

Conforms to torch.add(input, other, *, alpha=1). alpha is baked into the kernel, so non-default alpha runs through the same fast kernel as the default.

__init__

__init__(
    *,
    alpha=1,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.SubFwdOp

Element-wise subtraction with broadcast: y = input - alpha * other.

Conforms to torch.sub(input, other, *, alpha=1). alpha is baked into the kernel, so non-default alpha runs through the same fast kernel as the default.

__init__

__init__(
    *,
    alpha=1,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.MulFwdOp

Element-wise multiplication with broadcast: y = input * other.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.DivFwdOp

Element-wise division with broadcast: y = input / other.

Conforms to torch.div(input, other, *, rounding_mode=None). rounding_mode accepts None (true division), "trunc" (truncation toward zero), or "floor" (floor division); each value selects a dedicated kernel specialization. It is fixed for the instance, which is why it is not part of the memory key.

__init__

__init__(
    *,
    rounding_mode=None,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• rounding_mode (Optional[str], default: None ) –

  Manifest params.rounding_mode, str | None, default None.

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.FloorDivideFwdOp

Element-wise floor division with broadcast: y = floor(a / b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.RemainderFwdOp

Element-wise remainder with broadcast: y = a % b.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.PowFwdOp

Element-wise power with broadcast: y = input ** exponent.

Conforms to torch.pow(input, exponent): the second operand carries the manifest-declared name exponent rather than the generic other so the L1 signature check matches the manifest.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.MaximumFwdOp

Element-wise maximum with broadcast: y = max(a, b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.MinimumFwdOp

Element-wise minimum with broadcast: y = min(a, b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.LerpFwdOp

Element-wise lerp with broadcast: y = a + weight * (b - a).

Unlike torch.lerp(a, b, weight) where weight is a runtime parameter, here weight is a construction-time constant baked into the compiled kernel. This enables compile-time folding but means a new Op instance is needed for each distinct weight value.

__init__

__init__(
    *,
    weight=0.5,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• weight (float, default: 0.5 ) –

  Scalar interpolation weight, fixed at construction (manifest params.weight, default 0.5).

• target (Target, default: None ) –

  Which set of kernels serves this op.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.arithmetic.LerpTensorFwdOp

Tensor-weight lerp: out = input + weight * (end - input).

Conforms to the Tensor-weight overload of torch.lerp — torch.lerp(input, end, weight: Tensor) where weight is a Tensor that broadcasts together with input and end to the output shape. The scalar-weight overload is handled separately by LerpFwdOp.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    end,
    weight,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype float16 | bfloat16 | float32.

• end (Tensor) –

  Input tensor, dtype same_as(input).

• weight (Tensor) –

  Input tensor, dtype same_as(input).

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == broadcast_shapes(input.shape, end.shape, weight.shape).

Unary math

tileops.ops.elementwise.math_unary.ExpFwdOp

Element-wise exp(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.Expm1FwdOp

Element-wise exp(x) - 1.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.LogFwdOp

Element-wise log(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.Log1pFwdOp

Element-wise log(1 + x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.SqrtFwdOp

Element-wise sqrt(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.RsqrtFwdOp

Element-wise 1/sqrt(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.ReciprocalFwdOp

Element-wise 1/x.

Mirrors torch.reciprocal int-input promotion: the manifest declares the output as promote_int_to_float(input), and ReciprocalFwdKernel.specialize names float32 as the compute type for an integral input. The semantic dtype keys the specialization and drives roofline accounting — integer input bytes, float32 output bytes — while the kernel is built for the type it computes in. Floating inputs follow the standard same-dtype path.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.AbsFwdOp

Element-wise |x|.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.NegFwdOp

Element-wise -x.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.SignFwdOp

Element-wise sign(x): -1, 0, or +1.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.SinFwdOp

Element-wise sin(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.CosFwdOp

Element-wise cos(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.ErfFwdOp

Element-wise erf(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.FloorFwdOp

Element-wise floor(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.CeilFwdOp

Element-wise ceil(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.RoundFwdOp

Element-wise round(x) to decimals decimal places.

The shipped kernel performs banker's round-to-nearest-integer, matching torch.round for decimals=0. decimals is a manifest param, so it is fixed for the instance and handed to whichever kernel serves the op; in-tree, a non-zero value selects _RoundDecimalsCall.

__init__

__init__(
    *,
    decimals=0,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• decimals (int, default: 0 ) –

  Number of decimal places to round to (manifest params.decimals, default 0).

• target (Target, default: None ) –

  Which set of kernels serves this op.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.math_unary.TruncFwdOp

Element-wise trunc(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

Activations

tileops.ops.elementwise.activations.ReluFwdOp

ReLU activation: y = max(x, 0).

__init__

__init__(
    *,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.LeakyReluFwdOp

Leaky ReLU: y = x if x > 0 else negative_slope * x.

__init__

__init__(
    *,
    negative_slope=0.01,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• negative_slope (float, default: 0.01 ) –

  Slope for negative inputs (default 0.01).

• inplace (bool, default: False ) –

  When True, copy the result back into input and return input (preserving tensor identity). The kernel still computes into a fresh buffer; only the user-visible tensor is mutated, mirroring torch.nn.functional.leaky_relu.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune the kernel.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.EluFwdOp

ELU: y = x if x > 0 else alpha * (exp(x) - 1).

__init__

__init__(
    *,
    alpha=1.0,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• alpha (float, default: 1.0 ) –

  Scale for the negative part (default 1.0).

• inplace (bool, default: False ) –

  When True, copy the result back into input and return input (preserving tensor identity).

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune the kernel.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.SeluFwdOp

Element-wise SELU activation.

__init__

__init__(
    *,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.GeluFwdOp

Element-wise GELU honoring the manifest approximate contract.

Parameters:

• approximate (str, default: 'none' ) –

  Approximation mode. 'none' (default) routes to the erf-based GeluFwdKernel. 'tanh' routes to GeluTanhFwdKernel (the fused tanh approximation 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))).

• target (Target, default: None ) –

  Which set of kernels serves this op.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune the kernel.

__init__

__init__(
    *,
    approximate="none",
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.SiluFwdOp

Element-wise SiLU (Swish): y = x * sigmoid(x).

__init__

__init__(
    *,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.MishFwdOp

Element-wise Mish: y = x * tanh(softplus(x)).

__init__

__init__(
    *,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.SigmoidFwdOp

Element-wise sigmoid(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.TanhFwdOp

Element-wise tanh(x).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.HardsigmoidFwdOp

Element-wise HardSigmoid: y = clamp(x + 3, 0, 6) / 6.

__init__

__init__(
    *,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.HardswishFwdOp

Element-wise HardSwish: y = x * clamp(x + 3, 0, 6) / 6.

__init__

__init__(
    *,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Backend target to serve this op, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel override dict.

• tune (bool, default: False ) –

  Whether to autotune, applied when a kernel is first built.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.HardtanhFwdOp

Hardtanh: y = clamp(x, min_val, max_val).

__init__

__init__(
    *,
    min_val=-1.0,
    max_val=1.0,
    inplace=False,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• min_val (float, default: -1.0 ) –

  Lower bound (default -1.0).

• max_val (float, default: 1.0 ) –

  Upper bound (default 1.0).

• inplace (bool, default: False ) –

  When True, copy the result back into input and return input (preserving tensor identity).

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune the kernel.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.activations.SoftplusFwdOp

Softplus: y = log(1 + exp(xbeta))/beta if xbeta <= threshold else x.

__init__

__init__(
    *,
    beta=1.0,
    threshold=20.0,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• beta (float, default: 1.0 ) –

  Scaling factor (default 1.0).

• threshold (float, default: 20.0 ) –

  Linear regime threshold (default 20.0).

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune the kernel.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.prelu.PreluFwdOp

PReLU: y = x if x > 0 else weight[channel] * x.

Channel dimension follows PyTorch convention: dimension 1 for inputs with ndim >= 2, dimension 0 for 1-D inputs. Both the shape and the channel count arrive with the tensors.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional dispatch override mapping kernel keys to Kernel subclasses. Falls back to default_kernel_map.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    weight,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype float16 | bfloat16 | float32.

• weight (Tensor) –

  Input tensor, dtype same_as(input).

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == input.shape.

Gated activations

tileops.ops.elementwise.activations.SiluAndMulFwdOp

SiLU-and-Mul: y = silu(gate) * value.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    x,
)

Run the op on x.

tileops.ops.elementwise.activations.GeluAndMulFwdOp

GELU-and-Mul: y = gelu(gate) * value (exact GELU).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    x,
)

Run the op on x.

tileops.ops.elementwise.activations.GeluTanhAndMulFwdOp

GELU-Tanh-and-Mul: y = gelu_tanh(gate) * value (tanh approximation).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    x,
)

Run the op on x.

Comparison

tileops.ops.elementwise.comparison.EqFwdOp

Element-wise equality with broadcast: y = (a == b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.comparison.NeFwdOp

Element-wise not-equal with broadcast: y = (a != b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.comparison.GtFwdOp

Element-wise greater-than with broadcast: y = (a > b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.comparison.GeFwdOp

Element-wise greater-equal with broadcast: y = (a >= b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.comparison.LtFwdOp

Element-wise less-than with broadcast: y = (a < b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.comparison.LeFwdOp

Element-wise less-equal with broadcast: y = (a <= b).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.comparison.IsnanFwdOp

Element-wise isnan with bool output.

Always False on integer / bool input (no NaN representation in those dtypes).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.comparison.IsinfFwdOp

Element-wise isinf with bool output.

Always False on integer / bool input (no Inf representation in those dtypes).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.comparison.IsfiniteFwdOp

Element-wise isfinite with bool output.

Always True on integer / bool input (every value in those dtypes is finite).

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

Logical and bitwise

tileops.ops.elementwise.logical.LogicalAndFwdOp

Element-wise logical AND with broadcast using non-zero truthiness.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.logical.LogicalOrFwdOp

Element-wise logical OR with broadcast using non-zero truthiness.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.logical.LogicalNotFwdOp

Element-wise logical NOT with bool output.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

tileops.ops.elementwise.bitwise.BitwiseAndFwdOp

Element-wise bitwise AND with broadcast: y = a & b.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.bitwise.BitwiseOrFwdOp

Element-wise bitwise OR with broadcast: y = a | b.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.bitwise.BitwiseXorFwdOp

Element-wise bitwise XOR with broadcast: y = a ^ b.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    other,
)

Run the op on input and other.

tileops.ops.elementwise.bitwise.BitwiseNotFwdOp

Element-wise bitwise NOT (~x) for bool/integer inputs.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on input.

Selection and clamping

tileops.ops.elementwise.where.WhereFwdOp

Where: out = condition ? input : other (with full PyTorch broadcasting).

Conforms to torch.where(condition, input, other): condition is a bool tensor and input / other may broadcast with each other and with condition to produce the output. The three operand shapes arrive with the tensors; broadcasting them is the kernel's business.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional dispatch override mapping kernel keys to Kernel subclasses. Falls back to default_kernel_map.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    condition,
    input,
    other,
)

Run the op on the inputs the manifest declares.

Parameters:

• condition (Tensor) –

  Input tensor, dtype bool.

• input (Tensor) –

  Input tensor, dtype float16 | bfloat16 | float32.

• other (Tensor) –

  Input tensor, dtype same_as(input).

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == broadcast_shapes(condition.shape, input.shape, other.shape).

tileops.ops.elementwise.clamp.ClampFwdOp

Clamp with Tensor lower and/or upper bounds (broadcasting).

Conforms to torch.clamp(input, min, max) where min and max are each either a Tensor or None. At least one of the two bounds must be a Tensor. All Tensor operands broadcast together. A single bound is torch.clamp_min / torch.clamp_max.

Which bounds this call carries is read off the call, not settled at construction: the manifest declares both as optional: true, so presence is a fact of the call, and it reaches the kernel's cache key because it changes what gets built. One instance therefore serves clamp, clamp_min and clamp_max, one specialization each.

__init__

__init__(
    *,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
    min=None,
    max=None,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype float16 | bfloat16 | float32.

• min (Optional[Tensor], default: None ) –

  Input tensor, dtype same_as(input). Optional.

• max (Optional[Tensor], default: None ) –

  Input tensor, dtype same_as(input). Optional.

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: min is None or max is None or output.shape == broadcast_shapes(input.shape, min.shape, max.shape); min is None or max is not None or output.shape == broadcast_shapes(input.shape, min.shape); max is None or min is not None or output.shape == broadcast_shapes(input.shape, max.shape).

tileops.ops.elementwise.clamp.ClampScalarFwdOp

Scalar-bound clamp (torch.clamp(input, min: Number|None, max: Number|None)).

__init__

__init__(
    *,
    min=None,
    max=None,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• min (Optional[float], default: None ) –

  Lower bound (Number or None).

• max (Optional[float], default: None ) –

  Upper bound (Number or None).

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune.

forward

forward(
    input,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype float16 | bfloat16 | float32.

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == input.shape.

tileops.ops.elementwise.masked_fill.MaskedFillFwdOp

MaskedFill with 0-dim Tensor value (torch.Tensor.masked_fill(mask, value: Tensor)).

Output shape is the bidirectional broadcast of input and mask; value must be a 0-dim Tensor. The kernel reads value at forward time, which is consistent with the 0-dim semantics.

__init__

__init__(
    *,
    target=None,
    kernel_map=None
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional dispatch override mapping kernel keys to Kernel subclasses. Falls back to default_kernel_map.

forward

forward(
    input,
    mask,
    value,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype bool | uint8 | int8 | int16 | int32 | int64 | float16 | bfloat16 | float32.

• mask (Tensor) –

  Input tensor, dtype bool.

• value (Tensor) –

  Input tensor, dtype same_as(input).

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == broadcast_shapes(input.shape, mask.shape).

tileops.ops.elementwise.masked_fill.MaskedFillScalarFwdOp

MaskedFill with Number (scalar) value.

Conforms to torch.Tensor.masked_fill(mask, value: Number). Output shape follows the bidirectional broadcast of input and mask.

The manifest declares the PyTorch dtype union (bool | uint8 | int8 | int16 | int32 | int64 | float16 | bfloat16 | float32); every union member dispatches to a real kernel. A bool operand is served by whatever storage the selected kernel requires; the op passes and receives semantic bool either way.

__init__

__init__(
    *,
    value=0,
    target=None,
    kernel_map=None
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• value (bool | int | float, default: 0 ) –

  Scalar fill value (bool / int / float). Range-validated against the element type of the call with PyTorch Tensor.masked_fill coercion: bool reduces non-zero to True; integer dtypes range-check the real value against torch.iinfo and truncate floats toward zero (1.5 -> 1); torch.uint8 additionally wraps Python ints in [-255, 0) via two's complement.

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional dispatch override mapping kernel keys to Kernel subclasses. Falls back to default_kernel_map.

forward

forward(
    input,
    mask,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype bool | uint8 | int8 | int16 | int32 | int64 | float16 | bfloat16 | float32.

• mask (Tensor) –

  Input tensor, dtype bool.

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == broadcast_shapes(input.shape, mask.shape).

tileops.ops.elementwise.nan_to_num.NanToNumFwdOp

NanToNum: replace NaN, +Inf, -Inf with specified values.

__init__

__init__(
    *,
    nan=0.0,
    posinf=None,
    neginf=None,
    target=None,
    kernel_map=None,
    tune=False
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• nan (float, default: 0.0 ) –

  Replacement for NaN (default 0.0).

• posinf (Optional[float], default: None ) –

  Replacement for +Inf. Manifest default None resolves to the largest finite value representable in the element type of the call (matches torch.nan_to_num). Explicit values must also be representable in that dtype end-to-end; values that fit only in the kernel's intermediate dtype (e.g. fp16 for fp8_e5m2) are rejected so the post-cast cannot resurface them as Inf.

• neginf (Optional[float], default: None ) –

  Replacement for -Inf. Manifest default None resolves to the smallest (most negative) finite value representable in the element type of the call.

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None to decide from the input device.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional kernel dispatch override.

• tune (bool, default: False ) –

  Whether to autotune the kernel.

forward

forward(
    input,
)

Run the op on the inputs the manifest declares.

Parameters:

• input (Tensor) –

  Input tensor, dtype float16 | bfloat16 | float32.

Returns:

• Tensor –

  output, as the manifest declares. Shape rules: output.shape == input.shape.

Positional encodings

tileops.ops.elementwise.alibi.AlibiFwdOp

ALiBi position encoding: bias[h, i, j] = -slope_h * |i - j|.

Generates the full (num_heads, seq_len, seq_len) bias tensor.

Note

Eager-only. Unlike the other elementwise ops in this package, AlibiFwdOp is not registered as a torch.library.custom_op, so torch.compile graph capture is not supported. The op has zero tensor inputs and constructs its output entirely from __init__ parameters; no compile-time wrapping is needed.

__init__

__init__(
    *,
    seq_len,
    num_heads,
    dtype,
    target=None,
    kernel_map=None
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• seq_len (int) –

  Sequence length.

• num_heads (int) –

  Number of attention heads.

• dtype (dtype) –

  Torch dtype.

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None. Nothing is probed: with no tensor input there is no device to detect, so the in-tree kernels serve unless a target is named.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional dispatch override mapping kernel keys to Kernel subclasses. Falls back to default_kernel_map.

forward

forward()

Run the op on the inputs the manifest declares.

Returns:

• Tensor –

  output, as the manifest declares.

tileops.ops.elementwise.sinusoidal.SinusoidalFwdOp

Sinusoidal positional encoding from "Attention Is All You Need".

Generates the full (seq_len, d_model) encoding tensor.

Note

Eager-only. Unlike the other elementwise ops in this package, SinusoidalFwdOp is not registered as a torch.library.custom_op, so torch.compile graph capture is not supported. The op has zero tensor inputs and constructs its output entirely from __init__ parameters; no compile-time wrapping is needed.

__init__

__init__(
    *,
    seq_len,
    d_model,
    dtype,
    target=None,
    kernel_map=None
)

Build the op. Shapes and dtype are taken from the first call.

Parameters:

• seq_len (int) –

  Sequence length.

• d_model (int) –

  Model dimension.

• dtype (dtype) –

  Torch dtype.

• target (Target, default: None ) –

  Which set of kernels serves this op — a target name, BUILTIN for the in-tree kernels, or None. Nothing is probed: with no tensor input there is no device to detect, so the in-tree kernels serve unless a target is named.

• kernel_map (Optional[Dict[str, Kernel]], default: None ) –

  Optional dispatch override mapping kernel keys to Kernel subclasses. Falls back to default_kernel_map.

forward

forward()

Run the op on the inputs the manifest declares.

Returns:

• Tensor –

  output, as the manifest declares.