NPUData¶
A model input or output tensor that can reside on the host (CPU) or NPU.
An NPUData backs a single model input/output tensor: it owns its host buffer, holds NPU memory that belongs to the Accelerator while the data is on the NPU, and lets the data be moved between CPU and NPU. Instances are created through Model.acquire_input_npu_data / Model.acquire_output_npu_data, not constructed directly.
The data is accessible as a numpy array (via indexing, e.g. nd[...]) only while it resides on the CPU. Bring it back with cpu before reading results that were produced on the NPU.
An NPU-resident NPUData may be passed to any model that expects a tensor of the same shape and dtype on the same accelerator, so several models can share one NPU-resident tensor without copying it back to the host in between.
An NPUData is not owned by the Model it came from and may outlive it. Its NPU memory belongs to the Accelerator it resides on: if that accelerator is destroyed while the data is still on the NPU, the memory is reclaimed and the NPUData becomes invalid (cpu, launch, copy_from and indexing raise, and dev_no returns -1). A CPU-resident NPUData is unaffected.
Note
This is an advanced API rather than a typical usage.
Warning
This API is in beta: it may still contain bugs, and its behavior may change in a future release.
Warning
NPUData objects are not fully thread-safe; the client must serialize access to an NPUData from different threads. In particular, launch and cpu change where the data resides and must not run concurrently with each other.
Attributes¶
| Name | Description |
|---|---|
shape |
Returns the shape of the data. |
dtype |
Returns the numpy dtype of the data's elements. |
dev_no |
Device number of the accelerator the data resides on. |
hardware_name |
Hardware name (device type) of the accelerator the data resides on. |
shape¶
Returns the shape of the data.
dtype¶
Returns the numpy dtype of the data's elements.
dev_no¶
Device number of the accelerator the data resides on.
Returns
The device number when the data is on the NPU, or -1 when it is on the host (CPU) or has been invalidated (see the class-level note on accelerator destruction).
hardware_name¶
Hardware name (device type) of the accelerator the data resides on.
Device numbers are assigned per hardware type (e.g. an "aries-rb" and a "regulus-ra" can both be device #0), so dev_no alone does not identify a physical device; pair it with this name to distinguish devices.
Returns
The hardware name (e.g. "aries-rb") when the data is on the NPU, or an empty string when it is on the host (CPU) or has been invalidated (see the class-level note on accelerator destruction).
Methods¶
| Name | Description |
|---|---|
__init__ |
Internal constructor. |
launch |
Uploads the data to the NPU memory of the given accelerator. |
cpu |
Brings the data back to host (CPU) memory. |
copy_from |
Copies the data from another NPUData. |
__getitem__ |
Reads element(s) from the host buffer via numpy indexing. |
__setitem__ |
Writes value(s) into the host buffer via numpy indexing. |
__len__ |
Returns the length of the buffer's first dimension. |
__init__¶
Internal constructor.
Parameters
_npu_data— The underlying C++ NPUData object. UseModel.acquire_input_npu_data/Model.acquire_output_npu_datato create instances.
launch¶
Uploads the data to the NPU memory of the given accelerator.
Warning
Not thread-safe: see the class-level note.
Parameters
acc— The accelerator on which to place the data.
cpu¶
Brings the data back to host (CPU) memory.
Warning
Not thread-safe: see the class-level note.
copy_from¶
Copies the data from another NPUData.
Both tensors must have the same shape and dtype; the source may reside on either side (a CPU-to-NPU or NPU-to-CPU copy is staged through the host).
Parameters
src— The source NPUData to copy from.
__getitem__¶
Reads element(s) from the host buffer via numpy indexing.
__setitem__¶
Writes value(s) into the host buffer via numpy indexing.
__len__¶
Returns the length of the buffer's first dimension.