flux2distill/__init__.py

"""flux2distill — compress FLUX.2 [klein] distilled 4B into a ~2B 4-step student.

See plan.md for the full design. Modules:
  config       — dataclasses for model / surgery / data / train / eval
  surrogate    — low-rank + nonlinearity residual surrogate, with SVD/least-squares init
  surgery      — block selection (SVD-energy) + replacement -> student   (model-specific)
  calibration  — surrogate warm-start (SVD init + short offline fit)
  data         — jasperai/monet streaming + VAE latent caching
  losses       — velocity matching, masked feature matching, flow-matching
  optim.muon   — Muon optimizer (2D hidden weights)
  model_utils  — load teacher/student, block naming, param counting        (model-specific)
"""

__version__ = "0.1.0"