For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Python class

BufferValue

`BufferValue`

class max.graph.BufferValue(value)

source

Bases: Value[BufferType]

Represents a mutable semantic tensor within a Graph.

Unlike a TensorValue, which is value semantic and produces a new value from every operation, a BufferValue refers to memory you can update in place. Read it into a tensor with buffer_load() and write a tensor back with buffer_store(). This is how a graph carries state that must persist across executions, such as a key-value cache.

The following example loads a buffer input, updates it, and stores the result back in place:

from max.dtype import DType
from max.graph import BufferType, DeviceRef, Graph, ops

buffer_type = BufferType(DType.float32, shape=[4], device=DeviceRef.CPU())

with Graph("buffer_demo", input_types=[buffer_type]) as graph:
    state = graph.inputs[0].buffer

    # Read the current contents into a value-semantic tensor.
    current = ops.buffer_load(state)

    # Write an updated tensor back into the same memory.
    ops.buffer_store(state, current + 1)
    graph.output(current)

    print(f"dtype: {state.dtype}")  # Output: dtype: DType.float32
    print(f"shape: {state.shape}")  # Output: shape: [Dim(4)]

Initializes a BufferValue from another value.

Parameters:: value (Value[Any] | _Value[mo.BufferType] | HasBufferValue) – The value to wrap, either an MLIR value of buffer type or another BufferValue.

`device`

property device: DeviceRef

source

Returns the device of the BufferValue.

`dtype`

property dtype: DType

source

Returns the tensor data type.

`from_mlir()`

classmethod from_mlir(value)

source

Creates a BufferValue from an MLIR buffer value.

Parameters:: value (Value[BufferType]) – The MLIR buffer value to wrap.
Return type:: BufferValue

`print()`

print(label='debug_buffer')

source

Prints detailed information about the buffer.

Parameters:: label (str)
Return type:: None

`rank`

property rank: int

source

Returns the rank (number of dims) of the buffer.

`shape`

property shape: Shape

source

Returns the shape of the BufferValue.

`type`

property type: BufferType

source

Returns the type of the BufferValue as a BufferType.

BufferValue​

device​

dtype​

from_mlir()​

print()​

rank​

shape​

type​