Python module

ops

Implements ops used when staging a graph.

Although the following modules provide a lot of the ops you want when building a graph, you can also use functions in Graph to add constant values, such as constant(), vector(), and scalar().

The TensorValue type (returned by most ops) also implements various dunder methods to support operations between TensorValues, such as + add, * multiply, and @ matmul, plus convenience methods such as reshape() and swapaxes().

Casting

broadcast_to()

max.graph.ops.broadcast_to(x: TensorValue, shape: Iterable[int | str | Dim | integer]) → TensorValue

Broadcasts a symbolic tensor.

Broadcasts the input tensor to the specified shape. Dimensions in the input must be one or match the target dimension.

• Parameters:

  • x – The input symbolic tensor to broadcast. This tensor may not contain any dynamic dimensions.
  • shape – The new shape as a list of dimensions. Dynamic dimensions are not allowed.
  • location – An optional location for a more specific error message.

• Returns:

  A symbolic tensor with the same elements as the original tensor, but in a new shape. Its symbolic shape is the same as shape.

cast()

max.graph.ops.cast(x: Value | Value | ndarray, dtype: DType) → TensorValue

Casts a symbolic tensor to a different data type.

• Parameters:

  • v – The input tensor to cast.
  • dtype – The target dtype to which the tensor is cast.

• Returns:

  A new symbolic tensor with the same shape as the input and the specified dtype.

rebind()

max.graph.ops.rebind(x: Value | Value | ndarray, shape: Iterable[int | str | Dim | integer], message: str) → TensorValue

Rebinds a symbolic tensor to a specified set of dimensions.

This does not mutate the symbolic tensor passed in, but instead adds a runtime assert that the input symbolic shape is equivalent to out_dims shape. For example, if the input tensor shape has dynamic/unknown sizes, this will assert a fixed sizes that may be required for a subsequent operation.

• Parameters:

  • x – The input symbolic tensor to rebind.
  • shape – The symbolic shape to assert for x, as a list of Dim values.
  • message – The message printed if the rebind fails at runtime.

• Returns:

  A symbolic tensor with the same elements and shape as the given tensor, but with the symbolic shape asserted to out_dims.

reshape()

max.graph.ops.reshape(x: Value | Value | ndarray, shape: Iterable[int | str | Dim | integer]) → TensorValue

Reshapes a symbolic tensor.

The number and order of the elements in the tensor is unchanged. In other words, if you were to iterate over elements in the tensor by major dimension to minor dimension, the iteration order would stay the same.

If a value of -1 is present in the shape, that dimension becomes an automatically calculated dimension collecting all unspecified dimensions. Its length becomes the number of elements in the original tensor divided by the product of elements of the reshape.

• Parameters:

  • x – The input symbolic tensor to reshape. This tensor may not contain any dynamic dimensions.
  • shape – The new shape as a list of dimensions. Dynamic dimensions are not allowed. A single dimension may be -1.

• Returns:

  A symbolic tensor with the same elements as the original tensor, but in a new shape. Its symbolic shape is the same as shape.

shape_to_tensor()

max.graph.ops.shape_to_tensor(shape: Iterable[int | str | Dim | integer]) → TensorValue

squeeze()

max.graph.ops.squeeze(x: Value | Value | ndarray, axis: int) → TensorValue

Removes a size-1 dimension from a symbolic tensor.

• Parameters:

  • x – The input symbolic tensor to squeeze.
  • axis – The dimension to remove from the input’s shape. If negative, this indexes from the end of the tensor. For example, squeeze(v, -1) squeezes the last dimension.

• Returns:

  A symbolic tensor with the same number of elements as the input tensor, and whose rank is 1 less than the rank of the input tensor.

transpose()

max.graph.ops.transpose(x: Value | Value | ndarray, dim_1: int, dim_2: int) → TensorValue

Transposes two dimensions of a symbolic tensor.

• Parameters:

  • input – The input symbolic tensor to transpose.
  • x – One of the two dimensions to transpose. If negative, this indexes from the end of the tensor. For example, transpose(v, -1, -2) transposes the last two dimensions.
  • y – The other dimension to transpose. May also be negative to index from the end of the tensor.

• Returns:

  A new symbolic tensor with the two specified dimensions transposed. It has the same elements and dtype, but the order of the elements is different according to the transposition.

unsqueeze()

max.graph.ops.unsqueeze(x: Value | Value | ndarray, axis: int) → TensorValue

Inserts a size-1 dimension into a symbolic tensor.

• Parameters:

  • x – The input symbolic tensor to unsqueeze.
  • axis – The index at which to insert a new dimension into the input’s shape. Elements at that index or higher are shifted back. If negative, it indexes relative 1 plus the rank of the tensor. For example, unsqueeze(v, -1) adds a new dimension at the end, and unsqueeze(v, -2) inserts the dimension immediately before the last dimension.

• Returns:

  A symbolic tensor with the same number of elements as the input tensor, whose rank is 1 larger than the rank of the input tensor. The result’s shape at the axis dimension is a static dimension of size 1.

Complex

as_interleaved_complex()

max.graph.ops.as_interleaved_complex(x: Value | Value | ndarray) → TensorValue

Reshapes the input symbolic tensor as complex from alternating (real, imag).

• Parameters:

  interleaved – A symbolic tensor representing complex numbers as alternating pairs of (real, imag) real-valued numbers. Its last dimension must have an even size.

• Returns:

  A symbolic tensor representing the complex-valued tensor, but with the values pulled out as complex numbers. The result has the same dimensions for all dimensions except the last dimension, which is halved, and then a final dimension of size 2 representing the complex value.

Constant

constant()

max.graph.ops.constant(value: ndarray) → TensorValue

Adds a node representing a constant operation.

The value of this constant will have the type TensorType with the same shape and dtype as value.

• Parameters:

  • dtype – The constant tensor’s element type.
  • value – The constant’s value.

• Returns:

  A graph value containing the constant data as an attribute.

scalar()

max.graph.ops.scalar(value: int | float, dtype: DType, rank: int = 0) → TensorValue

Creates a scalar in the graph and returns the value.

A scalar is a tensor with a single element. Generally scalars are of rank 0, but that is configurable.

Elementwise

abs()

max.graph.ops.abs(x: TensorValue) → TensorValue

Computes the elementwise absolute value of a symbolic tensor.

Creates a new op node to compute the elementwise absolute value of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the absolute value computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

add()

max.graph.ops.add(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Adds two symbolic tensors.

Creates a new op node to compute the addition of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according : to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the : same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.
  • location – An optional location for a more specific error message.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

cos()

max.graph.ops.cos(x: TensorValue) → TensorValue

Computes the elementwise cosine of a symbolic tensor.

Creates a new op node to compute the elementwise cosine of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the cos computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

div()

max.graph.ops.div(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Divides two symbolic tensors.

Creates a new op node to compute the division of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according : to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the : same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.
  • location – An optional location for a more specific error message.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

equal()

max.graph.ops.equal(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise equality comparison between two symbolic tensors.

Creates a new op node to compute the equality comparison of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

erf()

max.graph.ops.erf(x: TensorValue) → TensorValue

Computes the elementwise error function of a symbolic tensor.

Creates a new op node to compute the elementwise error function of a symbolic tensor and adds it to the graph, returning the symbolic result.

The error function erf is defined as the probability that a randomly sampled normal distribution falls within a given range.

• Parameters:

  value – The symbolic tensor to use as the input to the error function computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

exp()

max.graph.ops.exp(x: TensorValue) → TensorValue

Computes the elementwise exp function of a symbolic tensor.

Creates a new op node to compute the elementwise exp function of a symbolic tensor and adds it to the graph, returning the symbolic result.

exp is defined as exp(x) = e^x, where e is Euler’s number.

• Parameters:

  value – The symbolic tensor to use as the input to the exp function computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

floor()

max.graph.ops.floor(x: TensorValue) → TensorValue

Computes the elementwise floor of a symbolic tensor.

Creates a new op node to compute the elementwise floor of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the floor computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

gelu()

max.graph.ops.gelu(x: TensorValue) → TensorValue

Computes the elementwise gelu function of a symbolic tensor.

Creates a new op node to compute the elementwise gelu function of a symbolic tensor and adds it to the graph, returning the symbolic result.

gelu is defined as $$gelu(x) = x \Phi(x)$$ where $$\Phi$$ is the cumulative distribution function of the Gaussian distribution.

• Parameters:

  value – The symbolic tensor to use as the input to the gelu function computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

greater()

max.graph.ops.greater(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise greater than comparison between two symbolic tensors.

Creates a new op node to compute the greater than comparison of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

greater_equal()

max.graph.ops.greater_equal(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise greater-or-equal comparison between two symbolic tensors.

Creates a new op node to compute the equality comparison of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

is_inf()

max.graph.ops.is_inf(x: TensorValue) → TensorValue

Computes the elementwise is_inf of a symbolic tensor.

Creates a new op node to compute the elementwise is_inf of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the is_inf computation.

• Returns:

  • element type bool, true if the element at a given position : is plus or minus infinity, false otherwise
  • the same shape as the input value.

• Return type:

  The result will have

• Raises:

  Raises – If the symbol doesn’t represent a tensor value.

is_nan()

max.graph.ops.is_nan(x: TensorValue) → TensorValue

Computes the elementwise is_nan of a symbolic tensor.

Creates a new op node to compute the elementwise is_nan of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the is_nan computation.

• Returns:

  • element type bool, true if the element at a given position : is NaN, false otherwise
  • the same shape as the input value.

• Return type:

  The result will have

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

log()

max.graph.ops.log(x: TensorValue) → TensorValue

Computes the elementwise natural logarithm of a symbolic tensor.

Creates a new op node to compute the elementwise natural logarithm of a symbolic tensor and adds it to the graph, returning the symbolic result.

The natural logarithm function log is defined as the inverse of the exponential function exp(). In other words, it computes the value y in the equation x = e^y where e is Euler’s number.

log(x) is undefined for x <= 0 for real numbers. Complex numbers are currently unsupported.

• Parameters:

  value – The symbolic tensor to use as the input to the natural logarithm computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

log1p()

max.graph.ops.log1p(x: TensorValue) → TensorValue

Computes the elementwise logarithm of 1 plus a symbolic tensor.

Creates a new op node to compute the elementwise log1p of a symbolic tensor and adds it to the graph, returning the symbolic result.

The log1p function is defined as log1p(x) = log(1 + x), where log() is the natural logarithm.

Using log1p(x) rather than computing log(1 + x) can give greater numerical precision results.

log(x) is undefined for x <= 0 for real numbers. Complex numbers are currently unsupported.

• Parameters:

  value – The symbolic tensor to use as the input to the log1p computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

logical_not()

max.graph.ops.logical_not(x: TensorValue) → TensorValue

Computes the elementwise logical_not of a symbolic tensor.

Creates a new op node to compute the elementwise logical_not of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the logical_not computation.

• Returns:

  • element type bool, true if the element at a given position : is plus or minus infinity, false otherwise
  • the same shape as the input value.

• Return type:

  The result will have

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

logsoftmax()

max.graph.ops.logsoftmax(x: TensorValue) → TensorValue

Computes the elementwise logsoftmax of a symbolic tensor.

Creates a new op node to compute the elementwise logsoftmax of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the logsoftmax computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

max()

max.graph.ops.max(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise maximum of two symbolic tensors.

Creates a new op node to compute the maximum of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

min()

max.graph.ops.min(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise minimum of two symbolic tensors.

Creates a new op node to compute the minimum of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

mod()

max.graph.ops.mod(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise maximum of two symbolic tensors.

Creates a new op node to compute the maximum of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

mul()

max.graph.ops.mul(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise multiplication of two symbolic tensors.

Creates a new op node to compute the multiplication of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

negate()

max.graph.ops.negate(x: TensorValue) → TensorValue

Computes the elementwise negate of a symbolic tensor.

Creates a new op node to compute the elementwise negate of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the negate computation.

• Returns:

  • element type bool, true if the element at a given position : is plus or minus infinity, false otherwise
  • the same shape as the input value.

• Return type:

  The result will have

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

not_equal()

max.graph.ops.not_equal(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise inequality comparison between two symbolic tensors.

Creates a new op node to compute the inequality comparison of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

pow()

max.graph.ops.pow(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise exponentiation of two symbolic tensors.

Creates a new op node to compute the exponentiation of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

relu()

max.graph.ops.relu(x: TensorValue) → TensorValue

Computes the elementwise relu of a symbolic tensor.

Creates a new op node to compute the elementwise relu of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the relu computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

round()

max.graph.ops.round(x: TensorValue) → TensorValue

Computes the elementwise round of a symbolic tensor.

Creates a new op node to compute the elementwise round of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the round computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

roundeven()

max.graph.ops.roundeven(x: TensorValue) → TensorValue

Computes the elementwise roundeven of a symbolic tensor.

Creates a new op node to compute the elementwise roundeven of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the roundeven computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

rsqrt()

max.graph.ops.rsqrt(x: TensorValue) → TensorValue

Computes the elementwise inverse-square-root of a symbolic tensor.

Creates a new op node to compute the elementwise rsqrt of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the rsqrt computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

sigmoid()

max.graph.ops.sigmoid(x: TensorValue) → TensorValue

Computes the elementwise sigmoid of a symbolic tensor.

Creates a new op node to compute the elementwise sigmoid of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the sigmoid computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

silu()

max.graph.ops.silu(x: TensorValue)

Computes the elementwise silu of a symbolic tensor.

Creates a new op node to compute the elementwise silu of a symbolic tensor and adds it to the graph, returning the symbolic result.

silu is defined as silu(x) = x * sigmoid(x).

• Parameters:

  value – The symbolic tensor to use as the input to the silu computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

sin()

max.graph.ops.sin(x: TensorValue) → TensorValue

Computes the elementwise sine of a symbolic tensor.

Creates a new op node to compute the elementwise sine of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the sin computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

softmax()

max.graph.ops.softmax(x: TensorValue) → TensorValue

Computes the elementwise softmax of a symbolic tensor.

Creates a new op node to compute the elementwise softmax of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the softmax computation.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

sqrt()

max.graph.ops.sqrt(x: TensorValue) → TensorValue

Computes the elementwise sqrt of a symbolic tensor.

Creates a new op node to compute the elementwise sqrt of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the sqrt computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

sub()

max.graph.ops.sub(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the elementwise subtraction of two symbolic tensors.

Creates a new op node to compute the subtraction of two symbol tensor values and adds it to the graph, returning the symbolic result.

• • If lhs and rhs have different dtypes, they will be promoted according to the dtype promotion rules before the operation.
  • If lhs and rhs have different shapes, they will be broadcast to the same shape according to broadcasting rules` before the operation.

• Parameters:

  • lhs – The symbol to use as left side of the addition.
  • rhs – The symbol to use as right side of the addition.

• Returns:

  A symbolic tensor value representing the output of the addition. The result will have:

  • the same dtype as the type-promotion of the two input dtypes
  • the same shape as the broadcast of the two input shapes.

• Raises:

  • Error – If the input values’ shapes are not compatible for broadcasting.
  • Error – If one of the input values has an unsupported dtype.
  • Error – If the two symbols are parts of different graphs.

tanh()

max.graph.ops.tanh(x: TensorValue) → TensorValue

Computes the elementwise tanh of a symbolic tensor.

Creates a new op node to compute the elementwise tanh of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the tanh computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

trunc()

max.graph.ops.trunc(x: TensorValue) → TensorValue

Computes the elementwise truncation of a symbolic tensor.

Creates a new op node to compute the elementwise truncation of a symbolic tensor and adds it to the graph, returning the symbolic result.

• Parameters:

  value – The symbolic tensor to use as the input to the truncation computation. If it’s not a floating-point DType, an exception will be raised.

• Returns:

  A new symbolic tensor value representing the output of the absolute : value computation.

• Raises:

  Error – If the symbol doesn’t represent a tensor value.

Linalg

band_part()

max.graph.ops.band_part(x: Value | Value | ndarray, num_lower: int, num_upper: int, exclude: bool = False) → TensorValue

Masks out everything except a diagonal band of an input matrix.

Copies a tensor setting everything outside the central diagonal band of the matricies to zero, where all but the last two axes are effectively batches, and the last two axes define sub matricies.

Assumes the input has dimensions [I, J, …, M, N], then the output tensor has the same shape as the input, and the values are given by

out[i, j, ..., m, n] = in_band(m, n) * input[i, j,  ..., m, n].

with the indicator function:

in_band(m, n) = ((num_lower < 0 || (m - n) <= num_lower)) &&
                (num_upper < 0 || (n - m) <= num_upper))

• Parameters:

  • input – The input to mask out.
  • num_lower – The number of diagonal bands to include below the central diagonal. If -1, include the entire lower triangle.
  • num_upper – The number of diagonal bands to include above the central diagonal. If -1, include the entire upper triangle.
  • exclude – If true, invert the selection of elements to mask. Elements in the band are set to zero.

• Returns:

  A symbolic tensor value with the configured selection masked out to 0 values, and the remaining values copied from the input tensor.

layer_norm()

max.graph.ops.layer_norm(input: TensorValue, gamma: Value | Value | ndarray, beta: Value | Value | ndarray, epsilon: float) → TensorValue

Performs layer normalization.

• Parameters:

  • input – The input tensor to normalize.
  • gamma – The gamma parameter of the normalization.
  • beta – The beta parameter of the normalization.
  • epsilon – The epsilon parameter of the normalization.

• Returns:

  A graph tensor value with the normalization applied.

matmul()

max.graph.ops.matmul(lhs: Value | Value | ndarray, rhs: Value | Value | ndarray) → TensorValue

Computes the matrix multiplication of two tensor graph values.

Performs general matrix multiplication with broadcasting.

If the lhs is 1D, it will be reshaped to 1xD. If the rhs is 1D, it will be reshaped to Dx1. In both cases, the additional 1 dimensions will be removed from the output shape.

For the multiplication, the innermost (rightmost) 2 dimensions are treated as a matrix. The lhs matrix will have the shape MxK. The rhs matrix will have the shape KxN. The output will have the shape MxN The K dimensions must be equivalent in both matrices.

The remaining outer dimensions will be broadcasted.

• Parameters:

  • lhs – The left-hand-side of the matmul.
  • rhs – The right-hand-side of the matmul.
  • location – An optional location for a more specific error message.

• Returns:

  A tensor graph value representing he result of broadcasting the two matricies together and then performing a matrix multiply along the innermost two dimension of each tensor.

Quantized

dequantize()

max.graph.ops.dequantize(encoding: QuantizationEncoding, quantized: TensorValue) → TensorValue

Dequantizes a quantized tensor to floating point.

NOTE: Currently this supports Q4_0, Q4_K, and Q6_K encodings only.

• Parameters:

  • encoding – The quantization encoding to use.
  • quantized – The quantized tensor to dequantize.

• Returns:

  The dequantized result (a floating point tensor).

qmatmul()

max.graph.ops.qmatmul(encoding: QuantizationEncoding, lhs: TensorValue, rhs: TensorValue) → TensorValue

Performs matrix multiplication between floating point and quantized tensors.

This quantizes the lhs floating point value to match the encoding of the rhs quantized value, performs matmul, and then dequantizes the result. Beware that, compared to a regular matmul op, this one expects the rhs value to be transposed. For example, if the lhs shape is [32, 64], and the quantized rhs shape is also [32, 64], then the output shape is [32, 32].

That is, this function returns the result from:

dequantize(quantize(lhs) @ transpose(rhs))

The last two dimensions in lhs are treated as matrices and multiplied by rhs (which must be a 2D tensor). Any remaining dimensions in lhs are broadcast dimensions.

NOTE: Currently this supports Q4_0, Q4_K, and Q6_K encodings only.

• Parameters:

  • encoding – The quantization encoding to use.
  • lhs – The non-quantized, left-hand-side of the matmul.
  • rhs – The transposed and quantized right-hand-side of the matmul. Must be rank 2 (a 2D tensor/matrix) and in a supported quantization encoding.

• Returns:

  The dequantized result (a floating point tensor).

Reduction

mean()

max.graph.ops.mean(x: Value | Value | ndarray, axis=-1) → TensorValue

Reduces a symbolic tensor using a mean operation.

• Parameters:

  • v – The input tensor for the operation.
  • axis – The axis along which to compute the reduction. If negative, indexes from the last dimension. For example, a value of -1 will compute the reduction along the last dimension.

• Returns:

  A symbolic tensor representing the result of the mean operation. The tensor will have the same rank as the input tensor, and the same shape except along the axis dimension which will have size 1.

Slicing

concat()

max.graph.ops.concat(vals: Iterable[Value | Value | ndarray], axis: int = 0, new_dim: int | str | Dim | integer | None = None) → TensorValue

Concatenates a list of symbolic tensors along an axis.

• Parameters:

  • vals – A list of symbolic tensor values. Each tensor must have the same dtype and rank, and must have the same dimension size for each dimension other than axis.
  • axis – The axis to concatenate along. If negative, indexes relative to the end of the tensor shape. For instance, concat(vs, -1) will concat along the last dimension.
  • new_dim – The expected output dimension of the concat axis. If provided, this will add a runtime assertion to the graph. If the resulting dimension can’t be known statically, new_dim _must_ be provided.

• Returns:

  A new symbolic tensor representing the concatenation result. It will have the same rank as each input tensor, and its dimensions will be the same as each input tensor’s for each dimension other than axis, which will have size equal to the sum of all tensor’s size for that dimension.

gather()

max.graph.ops.gather(input: Value | Value | ndarray, indices: Value | Value | ndarray, axis: int = -1) → TensorValue

Selects elements out of an input tensor by index.

• Parameters:

  • input – The input symbolic tensor to select elements from.
  • indices – A symbolic tensor of index values to use for selection.
  • axis – The dimension which indices indexes from input. If negative, indexes relative to the end of the input tensor. For instance, gather(input, indices, axis=-1) will index against the last dimension of input.

• Returns:

  A new symbolic tensor representing the result of the gather operation.

select()

max.graph.ops.select(cond: Value | Value | ndarray, x: Value | Value | ndarray, y: Value | Value | ndarray) → TensorValue

Returns condition ? x : y (element-wise), where cond, x and y are input tensors.

• Parameters:

  • condition – The condition tensor to use for selecting elementwise values.
  • x – If the condition is true at a position, the value from the same position in this tensor will be selected.
  • y – If the condition is false at a position, the value from the same position in this tensor will be selected.

• Returns:

  A new symbolic tensor holding either values from either x or y, based on the elements in condition.

stack()

max.graph.ops.stack(vals: Iterable[Value | Value | ndarray], axis: int = 0) → TensorValue

Stacks a list of tensors along a new axis.

• Parameters:

  • values – A list of symbolic tensor values. Each tensor must have the same dtype and rank, and must have the same dimension size for each dimension.
  • axis – The axis to concatenate along. If negative, indexes relative to the end of the tensor shape plus 1. For instance, stack(vs, -1) will create and stack along a new axis as the last dimension, aad stack(vs, -2) will create and stack along a new dimension which is inserted immediately before the last dimension.

• Returns:

  A new symbolic tensor representing the result of the stack. It will have rank n+1 where n is the rank of each input tensor. Its size on each dimension other than axis will be the same as each input tensors’, with the new axis inserted. Along the new dimension it will have size len(values).

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!
If you'd like to share more information, please report an issue on GitHub

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other