caesar

CAESAR (Conditional AutoEncoder with Super-resolution for Augmented Reduction) is a unified framework for spatio-temporal scientific data reduction, ported to C++ with LibTorch for deployment in high-performance computing and scientific workflows. The baseline model, CAESAR-V, is built on a variational autoencoder with scale hyperpriors and super-resolution modules. It encodes data into a compact latent space and uses learned priors to build an information-rich representation, aiming at high compression ratios while preserving scientific fidelity. The project cites a paper on a foundation model for lossy compression of spatiotemporal scientific data and a related work on generative latent diffusion for data reduction. Two model generations are described: CAESAR v1 (caesar_v.pt), the original foundation model, and CAESAR v2 (model_bs64_ep100k.pt), a newer optimized foundation model that is the default. Checkpoints are fetched through a model registry script, must be registered in the UFL catalog, and must match a registered SHA-256 hash before compilation. Model files are resolved from the CAESAR_MODEL_DIR environment variable, then an exported_model directory relative to the executable, then an installed system path. Build instructions cover Linux, macOS and Windows. Core dependencies include LibTorch 2.8 or newer, CMake 3.10, Zstandard 1.5 (required), Python 3.10 and PyYAML 6.0. The build uses CMake with the PyTorch CMake prefix path and optional tests. GPU acceleration is available on NVIDIA hardware via CUDA and nvCOMP, and on Apple Silicon via PyTorch's Metal Performance Shaders backend, with the MPS path noted as verified only on a nightly torch build. CPU execution is the default. Documentation includes developer notes on the codebase and a page on model installation and identity covering offline downloads, registration, devices, caching and an ADIOS contract. The repository is licensed Apache-2.0 and accepts questions, bug reports and contributions through GitHub issues.