Solvers and Profiling#
The benchmark provides an iterative plug-and-play solver for reconstruction and an unrolled plug-and-play solver for supervised training. Both use the same distributed execution context and profiling interface, making it possible to compare numerical quality and hardware behavior across single-GPU, multi-GPU, and multi-node runs.
Plug-and-Play Reconstruction
Alternate a data-fidelity gradient step and a learned denoising prior, with independent distribution of physics and denoising.
Unrolled PnP Training
Train an unrolled PGD model one supervised optimization step at a time, with distributed physics and patch-based model execution.
Profilers
Choose lightweight wall-clock metrics, operator-level PyTorch traces, or Nsight Systems timelines.