v0.9.0

Highlights

• NERSC Perlmutter A100 training support — VAE, LDM, and broadband training all work on Perlmutter
• Perlmutter benchmark results added to docs/BENCHMARKS.md
• New notebooks/ directory with three analysis/presentation notebooks
• Broadband --data-path + --wandb compatibility fixes

Added

• NERSC Perlmutter A100 training support — enables VAE, LDM, and broadband training on Perlmutter
  • scripts/nersc_train_vae.py — standalone VAE training script with jax.lax.dynamic_slice workaround for a CUDA graph capture bug on Perlmutter A100s
  • scripts/prepare_training_data.py — reproducible data prep for training (quality filters + metadata enrichment: ECLIPSE_FRAC, SOLFLUX, galactic/ecliptic coords, broadband aliases)
  • jobs/ — SLURM job scripts for VAE, LDM (dark/moon/twilight), and broadband training on NERSC
  • docs/NERSC_SETUP.md — training section documenting required XLA flags (--xla_gpu_autotune_level=0) and when the NERSC VAE script is needed
• Perlmutter benchmark results in docs/BENCHMARKS.md — CPU and A100 GPU timings at default and 10K sample sizes
• Analysis & research notebooks in a new notebooks/ directory (separate from examples/, which hosts user-facing tutorials):
  • notebooks/investigate_outlier_spectra.ipynb — investigation of anomalous LDM-generated spectra
  • notebooks/specsim_polar_comparison.ipynb — desisky vs specsim polar comparison
  • notebooks/desi_data_pres.ipynb — DESI data systems presentation companion

Fixed

• Broadband --data-path + --wandb compatibility — gather_full_data previously assumed SkyBrightnessDataset internals; now works with TensorDataset loaded from user data paths
• Broadband _save_checkpoint — added fallback for datasets without .metadata (e.g. TensorDataset from --data-path)
• Duplicate quality filter messages when loading the DESI wavelength grid

Install

​bash pip install --upgrade desisky ​

Full changelog

See CHANGELOG.md for the complete entry.