Python class
DevicePinnedBuffer
DevicePinnedBuffer
class max.driver.DevicePinnedBuffer(self, dtype: max.dtype.DType, shape: collections.abc.Sequence[int], device: max.driver.Device)
class max.driver.DevicePinnedBuffer(self, dtype: max.dtype.DType, shape: collections.abc.Sequence[int], stream: max.driver.DeviceStream)
Bases: Buffer
Creates a pinned host memory allocation tied to the given device.
Device-pinned memory allocations can provide faster DMA speeds and allow properly asynchronous copies between the device and the host.
Since device-pinned buffers can be used for asynchronous copies they don’t perform automatic synchronizations in operations like to_numpy, so synchronization should be handled manually to ensure GPU tasks writing to the buffer are complete before reading it on the host.
Since this class provides device-pinned memory it doesn’t work on CPU, for regular host memory use the Buffer class.
from max.driver import DevicePinnedBuffer, Accelerator
from max.dtype import DType
import numpy as np
device = Accelerator()
buffer = DevicePinnedBuffer(
dtype=DType.float32, shape=[1024], device=device
)
# Fill with data and transfer to GPU
np_data = buffer.to_numpy()
np_data[:] = np.arange(1024, dtype=np.float32)
gpu_buffer = buffer.to(device)-
Parameters:
-
- dtype (DType) – Data type of buffer elements.
- shape (Sequence[int]) – Tuple of positive, non-zero integers denoting the buffer shape.
- device (Device) – GPU/Accelerator device to associate buffer with. Must not be CPU.
- stream (DeviceStream, optional) – Stream to associate the buffer with.
-
Raises:
-
ValueError – If is a CPU device.
zeros
zeros = <nanobind.nb_func object>
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!