0.1.0 Release

[0.1.0] — 2026-03-08

First release under the Hamon name. Hamon is a spiritual successor to
Extropic AI's thrml (v0.1.3), diverging
as an independent library focused on GPU-accelerated thermal sampling for discrete
energy-based models.

Added

• Non-reversible parallel tempering (nrpt, nrpt_adaptive)
  • Vectorized swap pass exploiting temperature-linearity: 1 energy eval per chain
    instead of 4 per adjacent pair
  • Single-pass DEO (deterministic even-odd) swap scheduling
  • Adaptive schedule optimization (Algorithm 4, Syed et al. 2021) to equalize
    rejection rates and minimize the global communication barrier Λ
  • Iterative chain count discovery (discover_chain_count)
• Round trip tracking (round_trips module)
  • Index process monitoring carried through lax.scan with minimal overhead
  • Communication barrier estimation: local λ(β) and global Λ
  • Predicted vs observed round trip rate diagnostics
  • Chain count recommendation from Λ estimates
• Dynamic block construction (dynamic_blocks module)
  • Influence-aware partitioning: aggregate influence A(w) identifies heavy vertices
  • Per-temperature block sizing based on correlation length heuristics
  • Correlation-based re-blocking from empirical samples (Venugopal & Gogate 2013)
  • Influence-weighted Hamming distance for mixing diagnostics
• Boundary energy deltas (boundary_energy module)
  • Edge classification (incident, boundary, interior, external) per block partition
  • Rectangular block construction with 4-coloring for 2D grids
• vmap parallel tempering — all chains run in a single kernel via jax.vmap,
  replacing the original Python for-loop that unrolled N copies into XLA
• Scan carry threading — global state carried through lax.scan with targeted
  scatter updates; no redundant block_state_to_global per iteration
• BlockSpec fast path — energy() accepts pre-built BlockSpec / BlockGibbsSpec
  directly, skipping reconstruction on every call
• Precomputed scatter indices on BlockSamplingProgram (_block_sd_inds,
  _block_positions, _block_output_sds)
• Comprehensive test suite for all new modules

Fixed

• Deterministic global state layout — replaced set with dict.fromkeys for
  global_sd_order in BlockSpec.__init__; state ordering is now reproducible
• MomentAccumulatorObserver dtype — pinned at construction to avoid silent
  float64 promotion on GPU
• Non-array pytree leaves under vmap — _stack_pbi_across_chains preserves
  Python ints for slice indexing inside vmapped function bodies

Changed

• Renamed from thrml-boost to hamon; this project no longer tracks upstream
  thrml changes
• Package directory: thrml_boost/ → hamon/
• Version reset to 0.1.0 to reflect new project identity

Attribution

Core block sampling, factor, PGM, and observer infrastructure derived from
thrml (v0.1.3) by Extropic AI,
licensed under Apache 2.0. See NOTICE for details.