qbruntime.pinned_memory.PinnedMemory Class Reference#
|
SDK qb Runtime Library v1.3
MCS001-EN
|
An NPU-accessible pinned (physically contiguous) memory buffer. More...
Public Member Functions | |
| __init__ (self, Union[_cQbRuntime.PinnedMemoryFloat, _cQbRuntime.PinnedMemoryUint8] _pinned_memory) | |
| Internal constructor. | |
| __getitem__ (self, key) | |
| Reads element(s) from the pinned buffer via indexing. | |
| __setitem__ (self, key, value) | |
| Writes value(s) into the pinned buffer via indexing. | |
| int | __len__ (self) |
| Returns the length of the buffer's first dimension. | |
| _Shape | shape (self) |
| Returns the shape of the buffer. | |
| np.dtype | dtype (self) |
| Returns the numpy dtype of the buffer's elements. | |
| int | size (self) |
| Returns the size of the buffer in bytes. | |
| bool | is_available (self) |
| Returns whether the buffer is allocated and still usable. | |
Protected Attributes | |
| _pinned_memory = _pinned_memory | |
| _array = np.asarray(_pinned_memory) | |
Detailed Description
An NPU-accessible pinned (physically contiguous) memory buffer.
A PinnedMemory wraps a buffer that is pinned in device memory so the NPU can read from / write to it directly, enabling zero-copy inference via Model.infer_pinned_memory. Instances are created through Accelerator.allocate_pinned_memory, not constructed directly.
- Note
- This is an experimental API and is only supported on Regulus hardware.
Definition at line 20 of file pinned_memory.py.
Constructor & Destructor Documentation
◆ __init__()
| qbruntime.pinned_memory.PinnedMemory.__init__ | ( | self, | |
| Union[ _cQbRuntime.PinnedMemoryFloat, _cQbRuntime.PinnedMemoryUint8 ] | _pinned_memory ) |
Internal constructor.
- Parameters
-
_pinned_memory The underlying C++ pinned-memory object. Use Accelerator.allocate_pinned_memory to create instances.
Definition at line 31 of file pinned_memory.py.
Member Function Documentation
◆ __getitem__()
| qbruntime.pinned_memory.PinnedMemory.__getitem__ | ( | self, | |
| key ) |
Reads element(s) from the pinned buffer via indexing.
Supports the same indexing as numpy (integers, slices, tuples, etc.), e.g. pm[0], pm[1, 2], pm[:10]. The data is read directly from the pinned buffer without an intermediate copy.
- Parameters
-
key The index or slice.
- Returns
- The selected element or a view into the pinned buffer.
Definition at line 48 of file pinned_memory.py.
◆ __setitem__()
| qbruntime.pinned_memory.PinnedMemory.__setitem__ | ( | self, | |
| key, | |||
| value ) |
Writes value(s) into the pinned buffer via indexing.
Writes directly into the pinned buffer, e.g. pm[0] = 1.0, pm[:10] = src. This is the recommended way to fill an input buffer before Model.infer_pinned_memory.
- Parameters
-
key The index or slice. value The value(s) to write.
Definition at line 60 of file pinned_memory.py.
◆ __len__()
| int qbruntime.pinned_memory.PinnedMemory.__len__ | ( | self | ) |
Returns the length of the buffer's first dimension.
Definition at line 72 of file pinned_memory.py.
◆ shape()
| _Shape qbruntime.pinned_memory.PinnedMemory.shape | ( | self | ) |
Returns the shape of the buffer.
Definition at line 77 of file pinned_memory.py.
◆ dtype()
| np.dtype qbruntime.pinned_memory.PinnedMemory.dtype | ( | self | ) |
Returns the numpy dtype of the buffer's elements.
Definition at line 82 of file pinned_memory.py.
◆ size()
| int qbruntime.pinned_memory.PinnedMemory.size | ( | self | ) |
Returns the size of the buffer in bytes.
Definition at line 87 of file pinned_memory.py.
◆ is_available()
| bool qbruntime.pinned_memory.PinnedMemory.is_available | ( | self | ) |
Returns whether the buffer is allocated and still usable.
Definition at line 92 of file pinned_memory.py.
Member Data Documentation
◆ _pinned_memory
|
protected |
Definition at line 42 of file pinned_memory.py.
◆ _array
|
protected |
Definition at line 46 of file pinned_memory.py.
The documentation for this class was generated from the following file:
Generated by
1.14.0