GPU Acceleration¶

PtyLab.py uses CuPy as a drop-in GPU backend for NumPy. The same reconstruction code runs on both CPU and GPU — no changes to your script needed beyond toggling a single parameter.

Installation with GPU support¶

GPU acceleration requires installing PtyLab.py with a CUDA extra. See Installation for the full instructions.

pip install "ptylab[gpu] @ git+https://github.com/PtyLab/PtyLab.py.git"

Verifying GPU availability¶

From the command line (within your environment):

ptylab check gpu

Toggle GPU switch¶

By default, GPU is automatically detected and used for your reconstructions, however this can be also manually controlled via the following parameter

params.gpuSwitch = True

Memory optimization¶

For large datasets that do not fit entirely in GPU memory:

params.saveMemory = True

This reduces peak GPU memory usage at a slight performance cost by keeping some arrays on the CPU and transferring them as needed.

Single precision¶

Converting to single precision (float32) halves memory usage and can significantly speed up GPU reconstruction:

engine.convert2single()

Call this after creating the engine but before reconstruct().

Manual GPU/CPU transfer¶

In advanced workflows you can control data placement explicitly:

engine._move_data_to_gpu()    # move all arrays to GPU
engine._move_data_to_cpu()    # move all arrays back to CPU

Notes¶

GPU support is optional — PtyLab.py runs identically on CPU without CuPy installed.
The params.gpuSwitch property validates that CuPy is available before enabling GPU mode. It will raise an informative error if CuPy is not installed.
Results are automatically transferred back to CPU before saving with reconstruction.saveResults().