0.3.0 Release

[0.3.0] — 2026-06-10

Added

• Jit-once NRPT round loop — the Gibbs + DEO swap scan now lives in a
  module-level eqx.filter_jit function, so the compilation cache persists
  across nrpt calls. Templates already at β = 1 are reused without
  rebasing, and nrpt_adaptive rebases once before the phase loop, so all
  tuning phases plus production trace and compile exactly once (β arrays
  and states are traced data). Verified by a trace-count regression test;
  measured 1.80s → 0.54s (~3.3×) per nrpt_adaptive call on an 8-chain
  48×48 Ising benchmark (CPU).

• Temperature-linear NRPT mode — nrpt now accepts a single template
  (ebm, program) pair (plus an explicit betas array) instead of per-chain
  sequences. One base program is built at β = 1 and every interaction array
  is scaled by the chain's β inside the vmapped Gibbs kernel — valid for any
  model whose interactions are linear in β (the DiscreteEBMFactor family),
  consistent with the E_β = β·E_base assumption the swap math already makes.
  This avoids constructing one program per chain and storing per-chain copies
  of every interaction tensor (n_chains× less interaction memory).
  nrpt_adaptive and discover_chain_count use this mode automatically on
  their template (ebm=/program=) routes, eliminating all per-phase
  EBM/program rebuilds during schedule tuning (~22% faster adaptive tuning
  on CPU for an 8-chain 48×48 Ising benchmark). Results are bit-identical
  to the per-chain-programs path. Explicit factory routes are unchanged.

Breaking

• Minimum supported Python is now 3.11 (was 3.10). The JAX (≥ 0.9) and
  jaxtyping (≥ 0.3) releases hamon is developed against both require
  Python ≥ 3.11, so the previous 3.10 floor could only resolve to stale
  dependency versions. Python 3.14 is now supported and tested in CI.
• Padded interaction entries are pre-zeroed at program construction —
  BlockSamplingProgram masks the sliced interaction tensors with the active
  flags once, and the built-in spin/categorical conditionals no longer
  multiply by the active mask on every Gibbs step. Custom samplers invoked
  through a BlockSamplingProgram may now rely on inactive entries being
  zero; samplers called directly with hand-built (unmasked) interactions must
  keep applying the active flags themselves.

Changed

• jaxtyping floor raised to 0.3.10 (was 0.2.23) — picks up the fix for
  PyTree[A | B] isinstance checks silently passing (0.3.9) and cloudpickle
  round-trips of variadic annotations like Shaped[Array, "..."] (0.3.10)
• optax floor (testing extra) raised to 0.2.8 (was 0.2.4) — older optax
  imports the jax_pmap_shmap_merge config option removed in JAX 0.10.0 and
  fails at import
• CI test and example matrices now cover Python 3.11–3.14 (was 3.10–3.13)
• Gibbs scan carries one copy of the sampling state, not two — the
  _run_blocks scan previously threaded both the per-block state list and
  the concatenated global state through the carry, although samplers only
  read the global state. The carry now holds just the sampler states and
  the global state; per-block states are extracted once after the scan.
  from_global_state gained the same contiguous-slice fast path as the
  write-back side, so the extraction lowers to static slices. Results are
  bit-identical; measured ~6% faster NRPT rounds on CPU (the duplicate
  carry was multiplied by vmap across chains) and ~3% faster plain Gibbs.
• Block write-back uses contiguous slice updates instead of scatters —
  free blocks always occupy contiguous ranges of the global state (a
  BlockSpec layout invariant), so the per-block write-back in the Gibbs
  scan and in scatter_block_to_global now lowers to
  lax.dynamic_update_slice with a precomputed static offset instead of a
  gather-index scatter, which XLA fuses far better (the isolated op is ~30×
  faster on CPU; scatters are disproportionately expensive on GPU).
  Non-contiguous node sets keep the scatter fallback. Results are
  bit-identical.
• track_round_trips=False now skips the index-process update inside the
  swap pass instead of only omitting the summary.
• Conditional samplers accumulate in the weights' dtype — the spin and
  categorical Gibbs conditionals previously seeded their parameter
  accumulators with float32 zeros regardless of the model dtype.
• Documentation refreshed for the new internals: architecture.md now
  describes the concatenated (not padded) global state layout and the actual
  index-process representation; stale hinton_init and CategoricalNode
  docstrings corrected.

Fixed

• sample_blocks no longer mutates the caller's state and sampler-state
  lists.
• NRPT observers received post-swap states paired with pre-swap energies —
  in the default (non-cached) energy mode, the base-energy vector was not
  permuted after accepted swaps before being handed to the observer, so
  base_energies[c] described the state that used to occupy chain c.
  Energies are now permuted alongside the states in both energy modes.
• β₀ = 0 produced NaN base energies in NRPT, silently rejecting every swap —
  nrpt recovered base energies by dividing the hottest chain's energy by β₀,
  which is 0/0 when the ladder is anchored at the reference distribution
  (beta_range=(0.0, ...), the discover_chain_count default and the range
  ising_sample uses). Swap acceptance became NaN and every swap was rejected
  with no error, degrading parallel tempering into independent chains and
  inflating Λ estimates. Base energies are now computed from an exact β = 1
  copy of the EBM via with_beta(), falling back to the coldest chain's β for
  EBM classes without with_beta() (raising a clear error if that β is 0).