`constant`¶

Emit a tensor of a fixed shape filled with one constant value.

Category: arithmetic · Identity: constant@1

Shape¶

x[B, ...]:any -> out

Relation: out == [B, *shape]; x contributes only its batch extent and device

Port	Direction	Pattern	dtype
`x`	input	`x[B, ...]:any`	any
`out`	output	`out`	compute

Arguments¶

Name	Type	Default	Constraints	Description
`shape` (positional)	ints	`()`	>= 1; <= 2147483647	Non-batch dimensions of the constant. Omit it to read them from the output contract.
`value`	float	`0.0`	—	Value every element is filled with.

Positional values fill shape.

Description¶

out = full([B, *shape], value): a source of fixed values with no parameters and no buffers.

out[b, ...] = value

Give the non-batch dimensions positionally, constant(3, 8, 8), or as shape=(3, 8, 8); omit them entirely and the plan reads them from the output contract, the way reshape() does. value defaults to 0.0, so constant(16) is a zero vector — a null conditioning input — and constant(16, value=1.0) is a vector of ones.

The incoming tensor x supplies only the runtime batch extent and the device; its own shape, rank, and values are ignored, so the input port accepts any dtype including an integer graph input such as token ids. The result always carries the plan's compute dtype and it is created fresh on every call, detached from the autograd graph: it requires no gradient and no gradient reaches x. The operation is deterministic and identical in train and eval mode.

Because the constant tells the solver nothing about x, it does not propagate shapes backward: the operation before it must have its shape fixed from the input side or by its own arguments.

Examples¶

Example 1¶

A null conditioning vector of 16 zeros appended to the features.

z = constant(16)
concat(x, z)
linear(8)

Input ['B', 4] → output ['B', 8].

Network: [B, 4] -> [B, 8]  dtype=float32
index  name  operation  input shapes           output shapes
0      n0    constant   x=[B, 4]               out=[B, 16]
1      n1    concat     x0=[B, 4], x1=[B, 16]  out=[B, 20]
2      n2    linear     x=[B, 20]              out=[B, 8]

Parameters: 168

Example 2¶

A constant image plane: every pixel of every channel is centered by 0.5.

c = constant(3, 8, 8, value=0.5)
sub(x, c)
conv(4, kernel_size=3, padding=1)

Input ['B', 3, 8, 8] → output ['B', 4, 8, 8].

Network: [B, 3, 8, 8] -> [B, 4, 8, 8]  dtype=float32
index  name  operation  input shapes                    output shapes
0      n0    constant   x=[B, 3, 8, 8]                  out=[B, 3, 8, 8]
1      n1    sub        a=[B, 3, 8, 8], b=[B, 3, 8, 8]  out=[B, 3, 8, 8]
2      n2    conv       x=[B, 3, 8, 8]                  out=[B, 4, 8, 8]

Parameters: 112

Example 3¶

The omitted shape is read from the [B, 8] contract the add imposes.

h = linear(8)
z = constant(value=1.0)
add(h, z)