polystep 0.9.0__tar.gz → 0.10.1__tar.gz

polystep-0.10.1/CHANGELOG.md

@@ -0,0 +1,625 @@
+# Changelog
+
+## 0.10.1 - 2026-08-01
+
+### Fixed
+
+- `adaptive_probes` reused the previous batch's cost matrix unless the caller passed
+  `objective_token`, which is now required for reuse. Pass `objective_token=0` for a
+  stationary full-batch objective.
+- `block_group_size < 1` produced no blocks and wrote zeros over every parameter.
+- A scheduled `epsilon` reaching 0 froze the run; a negative one made the optimizer
+  ascend. Both are rejected when the schedule resolves.
+- The multi-fidelity screen gave `top_k_mean` mass on vertices never scored at full
+  fidelity: `keep_v` capped at `k` where the solver needs at least `k`.
+- A closure returning a scalar broadcast over every candidate, so the step was zero.
+- `HybridSubspace(rank=0)` gave a spec with no coordinates and froze those layers.
+- Blockwise averaged probes before sanitizing, so it disagreed with the monolithic
+  driver whenever one probe was non-finite.
+- `load_state_dict` aliased the checkpoint's tensors, so running the optimizer mutated
+  the loaded dict. `SparseRandomProjection` is rebuilt rather than stored by reference,
+  and a control key missing from an older format no longer overwrites the live value.
+- `absorb` dropped buffers from `base_params`.
+- Newton refinement scored against the post-momentum position, double-counting momentum.
+- `LinearSubspace` left `num_coords` uncapped, giving a rank-deficient projection and,
+  at high rank, an allocation larger than the layer.
+- The adaptive-radius controller read the probe mean, which grows with the radius it
+  scales. It now uses the min-over-vertices estimator the trust region uses.
+- `PolyStepES` overwrote a caller-supplied solver's `epsilon`, and a `-inf` fitness held
+  `best_solution` permanently.
+- `train()` rejects a model the optimizer was not built on, rejects `trust_region`
+  (which it silently pinned at 1.0), and restores the caller's train mode.
+- CMA covariance adaptation raises on a sparse projection instead of updating state that
+  reaches nothing. `SparseRandomProjection` warns when a basis reaches under half the
+  parameters.
+
+Crashes: blockwise buffer overrun at `chunk_size < V*K`; mixed-dtype models on the fused,
+factored and delta paths; `resync_from_model` after a dtype or device change; zero-numel
+parameters under `per_layer`; a seeded CPU generator with a CUDA point; a fully masked
+attention row.
+
+### Performance
+
+- The subspace site path built a dense coordinate matrix with `particle_dim` nonzeros per
+  row. It is block-diagonal by particle group, so it is now a per-group `bmm`: 53-96x on
+  that matmul, 2.4x on a ConvNet step at `subspace_dim=3630`.
+- `SiteVmapEvaluator` runs under blockwise, which previously materialized every candidate
+  on conv, normalization, attention and custom-forward models: 40x on a ConvNet.
+- Blockwise caches its index buffers instead of rebuilding them per block per step.
+- `train()` scores its tracked loss on the same subsample the candidates used.
+- Counting the quadratic model's centre evaluations against the forward budget drops the
+  measured `orthoplex` margin in `docs/performance.md` from 2.0x to 1.7x.
+
+## 0.10.0 - 2026-07-30
+
+New default polytope, four site-aware evaluation paths, and a pass over fast-path
+correctness. Retune `step_radius` wherever `polytope_type` was left at its default.
+Measured numbers are in `docs/performance.md`.
+
+### Breaking
+
+- `polytope_type` defaults to `'simplex'` (was `'orthoplex'`): `k+1` evaluations per step
+  instead of `2k`. Features reading the antithetic ordering (`use_quadratic_model`,
+  `newton_refinement`, `trust_region`, `multifidelity_screen`) warn on another polytope.
+- Rotations use a sign-corrected QR at or below `geometry._QR_MAX_BATCH` and Householder
+  reflections above. Both Haar on `SO(d)`, but the draw per seed differs from 0.9.0 and
+  across the threshold, so seeded runs do not reproduce step for step.
+- `compile_forward` defaults to the in-place path; pass `False` to opt out.
+- `adaptive_num_probe` defaults on only when `num_probe > 1`.
+- Momentum is zeroed on a basis change instead of carried into the new basis.
+- `projection_mode` removed: `'structured'` truncated its block-diagonal from the right,
+  leaving the trailing output rows with a zero projection.
+- `use_csa` raises `TypeError`; use `use_adaptive_radius`. Gone with it:
+  `polystep.cma.update_step_size_csa`, `SolverState.sigma`/`.use_csa`,
+  `CMAAdaptiveSubspace.d_sigma`/`expected_norm`. A 0.9.0 checkpoint still loads.
+- Top-level exports removed. Still importable from their modules: `PolyStep`
+  (`polystep.solver`), `VmapSafeLSTMCell` (`polystep.layers`). Deleted outright:
+  `Solver`, `SoftmaxResult`, `SinkhornResult`, `get_device`, `POLYTOPE_NUM_VERTICES_MAP`,
+  `get_sampled_polytope_vertices`, `get_probe_points`, `compute_nn_cost_matrix`,
+  `update_adaptive_radius`, `StyblinskiTang`, `Levy`, `Griewank`, `Beale`, `Branin`.
+- `PolyStep` drops `layout`, `block_strategy` and `block_group_size`, with
+  `polystep.blockwise.compute_block_cost_matrix` and `blocks_to_layout_flat_batch`.
+  `PolyStep.compile` defaults to `False`.
+- `update_evolution_path_sigma` takes `C_diag=None` for an already-whitened displacement;
+  `update_covariance_diagonal` takes `trace` and normalizes before clamping.
+- Rejected at construction instead of failing later: `epsilon`/`ent_epsilon <= 0`,
+  `probe_radius <= 0`, `step_radius < 0`, `cost_batch_size=0`,
+  `AdaptiveSubspace(subspace_dim > full_dim)`, unknown `block_strategy`/`polytope_type`.
+- `numpy` is no longer a core dependency; it moved to the `examples`, `dev` and
+  `experiments` extras.
+- `step_radius` is not comparable across subspace classes: `LinearSubspace` amplifies by
+  `sqrt(num_params / num_coords)` where `HybridSubspace` and `FactoredSubspace` are unit
+  gain.
+
+### Added
+
+- `SubspaceDeltaEvaluator` and `SiteVmapEvaluator` score candidates without building a
+  candidate weight. The second assumes nothing about the module set, so it covers conv,
+  normalization, attention and custom `forward`.
+- `polystep_elementwise` and `polystep_weight_transform` put a custom layer on the
+  batched paths, probed at build time so a false declaration falls back rather than
+  returning a wrong loss.
+- `candidate_autocast`: BF16 candidate arithmetic with parameters at their own dtype.
+- `use_quadratic_model` and `trust_region` run at `num_probe=1`, reading curvature from
+  `L(+s) + L(-s) - 2 L(0)` with one shared `f(X)` per particle.
+- `solver='kl_softmax'`, interpolating softmax (`kl_softmax_lam=0`) and Sinkhorn (`inf`).
+- `AdaptiveSubspace.rotation_interval`, `NNCostEvaluator(per_sample=True)` and
+  `get_diagnostics()` (`ess`, `rho`, `evals` per iteration).
+- `examples/10_cnn_mnist.py` (LeNet-5) and `examples/11_transformer_selective_copy.py`,
+  both forward-pass only, with `--compare` against the materializing path.
+
+### Fixed
+
+- A model computing in two dtypes raised in subspace mode: every per-layer projection was
+  built at the layout's `dominant_dtype`, and `SiteVmapEvaluator` cast inputs to the
+  perturbed layer's dtype rather than the model's. `HybridSubspace`, `LinearSubspace` and
+  `LowRankSubspace` now train such a model at each parameter's own dtype; full space and
+  `AdaptiveSubspace` raise at construction naming the fix.
+- Fast-path selection accepted models it scored wrong: the bmm plan deduplicated shared
+  modules, loss subclasses and forward hooks were skipped, tied-weight detection merged a
+  square parameter with its transpose, aliasing was keyed on the first element's address,
+  frozen parameters raised `KeyError` or dropped a bias, `ParamLayout` merged differently
+  shaped views of one buffer, and `register_evaluator` and `reset_vmap()` left stale
+  plans. All decline or rebuild now.
+- Numerics: the Sinkhorn and KL-softmax plans overflowed to `inf` after a dual
+  divergence, giving NaN parameters; `sanitize_cost` overflowed its own penalty near
+  dtype max; the in-place path's FP32 loss buffer collapsed distinct FP64 candidates;
+  site-aware chunks ignored `candidate_autocast`; the CMA clamp ran before renormalizing.
+- Basis changes leaked state: rotations left velocity, the Newton direction and the
+  transport EMA in the replaced basis; `HybridSubspace` kept displacement history across
+  one and rotated at step 0; a rank transition never rebuilt the fused projection, so
+  `_fused_P` kept the old rank's width; `RankSchedule` applied stage 0 late and omitted
+  `_applied_rank` from `state_dict`; `_write_params` wrote a stale buffer snapshot.
+- Step accounting: block-wise averaged loss, ESS and rho over blocks rather than
+  candidates; the trust-region ratio compared across a rotation and a minibatch and
+  double-shrank; `adaptive_probes` could latch permanently and reused rows across a
+  covariance change; `adaptive_num_probe` left a stale Newton direction;
+  `multifidelity_screen` billed stage one at the wrong fidelity.
+- Solvers: `MinCostGreedySolver` stepped on an all-infeasible row; `TopKMeanSolver` gave
+  mass to masked vertices; `KLSoftmaxSolver` built its plan one dual update behind and
+  never converged below the working dtype's smallest normal; `SinkhornSolver` reset
+  non-finite duals silently; the low-level `PolyStep` froze at one particle and left
+  diagnostics empty.
+- Config that silently did nothing: `LinearEpsilon` accepted a negative `decay`,
+  `max_subspace_dim` was not a bound, `RankSchedule` rebuilt the subspace every step,
+  `ProgressiveEpsilon` never advanced outside the monolithic step, explicit CMA rates
+  were discarded, `svd_ratio=0` still forced an SVD, and `TernaryLinear`'s default
+  threshold returned a constant.
+- `examples/05` annealed epsilon to 0.1, which concentrates the plan toward argmax so the
+  effective step grows to the full `step_radius` as `step_radius` anneals; the last epoch
+  diverged on every seed. `examples/10` built its schedules positionally, so its radii
+  were constant.
+- Crashes and unclear failures now raise, warn or decline: empty cost matrices, tiny
+  solver temperatures, an evaluator left alive after an exception in `train()`, NaN
+  fitness in `PolyStepES.tell`, `torch.set_default_device`, vector-only models, fp16 CPU
+  and CUDA `particle_dim > 2` rotation paths, a singular biased-rotation frame, and an
+  unprojected weight site.
+
+### Performance
+
+- `SubspaceDeltaEvaluator` reads the projection as a strided view instead of gathering
+  and transposing a column block per chunk, and batches over columns rather than `d_out`.
+  Losses bit-identical.
+- `HybridSubspace.init_projections` memoizes the step-0 basis, so an absorb re-anchors
+  the origin without redrawing it. The QR was 30% of the step under the default
+  `absorb_mode='stagnation'`.
+- Haar rotations pick QR or Householder reflections by batch size, each on the side of
+  the crossover where it wins.
+- Chunks break at parameter and coordinate boundaries, so the site-aware paths stop
+  falling back mid-sweep, and block-wise mode reaches the delta and factored evaluators.
+- The non-differentiable model zoo takes the batched paths; rate-coded SNNs compute the
+  static input current once instead of per timestep, bit-identical.
+- `SparseDeltaEvaluator` takes candidates grouped by particle, so each gather runs once
+  per group. `mixed_precision=True` no longer disables the subspace-delta path.
+- The block-diagonal fuse is capped at 32 MB and skips a single dense block, past which
+  the padding read costs more than the launches saved.
+- Every example pins a thread count below `nproc`, where the OpenMP spin-wait takes over.
+  Ten pin 1; `07` pins 8, its own objective being wide enough to pay for the pool.
+- Smaller: cached per-chunk casts and projections, one `sanitize_cost` per step,
+  candidates written straight into layout order, `torch._foreach_copy_` weight restore,
+  bias folded into `baddbmm`, batched sparse chunk projection, and a Haar guard that no
+  longer forces `dim-1` host syncs on CUDA.
+
+## 0.9.0 - 2026-07-27
+
+### Added
+
+- `SparseDeltaEvaluator`: a full-space candidate replaces `particle_dim` scalars, so its
+  output differs from the base in at most that many columns. Propagates the delta through
+  elementwise layers and goes dense one `Linear` past the perturbed one, instead of a
+  dense `bmm` over `N` weight sets. Automatic for `nn.Sequential` of `Linear` and
+  elementwise layers, with a per-chunk fallback.
+- `FactoredSubspace`: perturbs each 2D parameter by `dW = A @ B` with `B` fixed, so
+  `x (W + A B)^T = x W^T + (x B^T) A^T` and no candidate weight is built. Not the
+  default: the fixed `B` confines perturbations to `rank` input directions.
+- `step(..., objective_token=...)` marks which objective a step measures, so cost rows
+  cannot be reused across minibatches. `release_evaluator()`, `resync_from_model()`,
+  `MinCostGreedySolver` and `TopKMeanSolver` are public.
+- `MANIFEST.in`, so the sdist ships a runnable test suite. CI runs it from the tarball.
+
+### Fixed
+
+- `api.train()` never called `register_evaluator`, the only assignment of
+  `_cost_evaluator`, `_fused_inputs`, `_factored_evaluator` and
+  `_sparse_delta_evaluator`, so every fast evaluation path was unreachable through the
+  training API.
+- Cost-row reuse under `adaptive_probes` was per particle, but a candidate is the whole
+  configuration with one row replaced, so a stagnant particle's row was still measured
+  against every other particle's position. Reuse is now all or nothing.
+- `multifidelity_screen` averaged direction contrast across particles, which carry
+  independent rotations, so the kept set was chosen at random. Ranked per particle now.
+- Tied weights raised on both evaluation paths. `batch_unflatten` emits the canonical key
+  only; `unflatten` still carries aliases for `load_state_dict`.
+- `restore_best` keyed on the OT cost when no callback was registered, restoring the step
+  with the cheapest probe cloud rather than the best weights. It also snapshots to CPU.
+- Chunk size ignored the weight set a non-fused subspace builds per candidate and the
+  vmap activations. A 120K-param step peaked at 1661 MB in full space and 10565 MB in
+  subspace mode; now 946 MB and 1461 MB.
+- `AdaptiveSubspace` and per-layer rotation replaced the projection while the coordinates
+  were non-zero, moving the weights with nothing evaluated behind them.
+- `CMAAdaptiveSubspace` never rotated or absorbed, left its hyperparameters at `0.0` so
+  the covariance update was inert, normalized the evolution paths by `sigma` instead of
+  `step_radius` so `C_diag` decayed to its floor, carried state through basis rotations
+  untransformed, and accumulated at bf16 where `1 + c_1*x` rounds back to `1`.
+- Anderson acceleration subtracted `dX` instead of `dX + dR`, so `anderson_depth > 0`
+  converged slower than no acceleration.
+- `num_particles = 1` froze the optimizer: balanced OT with one row forces a uniform
+  plan. The default now selects `SoftmaxSolver` and warns.
+- `scale_cost` divided before recentering, so adding a constant to every cost changed the
+  plan. `SinkhornSolver` also reported `ent_reg_cost` in the wrong frame; the plan and
+  duals were unaffected. The greedy and top-k solvers had the same reporting bug.
+- Softmax paths recentred by the global minimum, so a row above it underflowed to `-inf`
+  and returned NaN; the fused kernel did not recenter at all.
+- `sanitize_cost` mapped `-inf` to the worst finite value instead of the best, and used
+  an absolute penalty that entered the `scale_cost='mean'` reduction.
+- `state_dict` captured only `SolverState`, so resume was not exact for `amortize_steps`,
+  `adaptive_probes`, `trust_region` or `HybridSubspace`. Format is 3; format 2 loads with
+  a warning.
+- A stagnation absorb never cleared `stagnation_count`, redrawing the basis every step on
+  a plateau. `resync_from_model()` left momentum and rotation state on the old anchor.
+- Probe losses accumulated in FP32 unconditionally, so a float64 objective whose costs
+  differ below FP32 resolution produced a constant cost matrix.
+- Block-wise biased rotation raised on bf16 under `mixed_precision=True`. A zero descent
+  direction produced a singular rotation collapsing `+e0` and `-e0`.
+- Smaller: the in-place evaluator dropped unknown parameter keys silently; the
+  trust-region ratio scored a Newton step that was never applied; displacement history
+  mixed bases; randomized PCA drew from the global RNG; rank transitions dropped
+  `HybridSubspace` settings; `data_dependent_init` had the wrong sign; a float64 model
+  with a subspace raised; a frozen model gave an opaque error; `pyproject.toml` declared
+  `torch>=2.4` where 0.2.0 announced `torch>=2.8`.
+
+### Changed
+
+- `adaptive_probes` and `adaptive_num_probe` default on under
+  `block_strategy='monolithic'`.
+- `use_csa` is refused with a warning: the OT step is a deterministic descent direction,
+  so CSA reads agreement as the normal state and `sigma` grows without bound.
+- `multifidelity_screen` cuts the forward budget instead of only reweighting, and runs
+  only
+  when `screen_fidelity/num_probe + screen_keep_ratio < 1`. Needs a cheap closure from
+`screen_closure_from()`; `api.train` supplies one.
+- Block-wise warns for options it ignores; `grouped` warns when combined with a subspace.
+  `LinearSubspace` amplifies rather than dilutes the perturbation, as its docstring said.
+- `_step_blockwise` claimed blockwise cuts OT cost from `O(N^2)` to `O(N^2/L)`. The
+  forward count is identical; `L` solves over `P/L` rows replace one over `P`.
+- Removed `get_sobol_rotation_matrices`, `HybridSubspace.reconstruct_base` and
+  `reconstruct_batch_delta`: no callers, no coverage.
+
+### Performance
+
+- `thin_qr` pins basis construction to one thread: `HybridSubspace` setup for the MNIST
+  example drops from 27.0 s to 97.8 ms. `apply_biased_rotation` uses modified
+  Gram-Schmidt instead of `torch.linalg.qr`, 7-33x faster at these sizes. Compiled
+  kernels take radius and epsilon as 0-d tensors, so a scheduled radius no longer
+  recompiles every step.
+- Example 06 on an RTX 5090: 1024 s to 845 s, shift-recovery +9.3 pp to +10.7 pp.
+
+### Examples
+
+- Examples pin CPU threads, check optional imports before use instead of failing after
+  the run, and register the evaluator in hand-rolled loops. `python-sat` moved into the
+  `examples` extra, which example 04 needs.
+
+## 0.8.0 - 2026-07-24
+
+### Added
+
+- `compile_forward` on `NNCostEvaluator` / `PolyStepOptimizer`: CUDA-graphs the
+  forward+loss on the in-place path and replays it per candidate. Rescues the
+  memory-forced path; does not beat compiled vmap. See `docs/performance.md`.
+- `TrainConfig(restore_best=True)` (default): restore the lowest-loss weights before
+  `train()` returns.
+- `PolyStepOptimizer.state_dict()` / `load_state_dict()` for checkpoint-resume (solver
+  state, `ProgressiveEpsilon` internals, RNG). Model weights save separately.
+
+### Changed
+
+- `compile_evaluator` / `compile_forward` propagate to a registered evaluator via
+  `register_evaluator`, so they work outside `api.train()`.
+- `train()` skips the extra per-step full-batch forward and its host syncs when no
+  callback consumes per-step metrics.
+
+### Fixed
+
+- Standalone `PolyStep` accepts any epsilon scheduler (`CosineEpsilon`,
+  `ProgressiveEpsilon`), not only `LinearEpsilon` (was `TypeError`).
+- `mixed_precision=True` builds the `ParamLayout` after the BF16 cast, so full-space
+  candidate params are BF16 (were FP32).
+- `evaluate_subspace_inplace` casts float inputs to the param dtype, fixing a mismatch
+  under `mixed_precision` + `HybridSubspace` + FP32 inputs.
+- `cost_batch_size` subsampling keeps the seeded RNG and compares the full device on a
+  mismatch (was non-deterministic; fixes a `cuda:0`/`cuda:1` `randperm` crash).
+- `PolyStep.run()` accepts a 1-D initial point (normalized to one particle).
+- `ProgressiveEpsilon` no longer inflates epsilon every step under a fixed-iteration
+  Sinkhorn solver (`threshold<=0`).
+- Greedy solvers: `MinCostGreedySolver` sanitizes non-finite costs; `TopKMeanSolver`
+  rejects `k=0`.
+- `PolyStep.step()` uses the shared `sanitize_cost` (was an inline copy that synced and
+  could overflow the penalty to `inf`); `state.costs` records the raw mean cost.
+- `HybridSubspace` docstring corrected (QR-orthonormal columns; `step_radius` retune).
+- `NNCostEvaluator` compile docstrings match the shipped modes; the vmap-fallback filter
+  matches functorch markers instead of any "batched" error.
+
+## 0.7.0 - 2026-07-20
+
+### Added
+
+- On PyPI now: `pip install polystep` (or `uv add polystep`).
+
+### Changed
+
+- Wider install range: `torch >= 2.4` (was 2.10) and `numpy >= 1.24` (was 2.0).
+- Blockwise and subspace-blockwise steps reuse one buffer for the per-chunk configs
+  instead of reallocating it every chunk, so a step allocates less.
+
+### Fixed
+
+- Solvers stay accurate when the cost matrix has a large constant offset (`|C|` far above
+  `epsilon`). The cost is recentered before the log-domain math, which leaves the
+  transport plan unchanged, so `SinkhornSolver.matrix`, `KLSoftmaxSolver(lam=0)`, and the
+  softmax solvers no longer lose precision or return NaN in that regime.
+- `SinkhornSolver` computes `ent_reg_cost` from the actual plan mass, so it is correct
+  before convergence too, and it warns when the two marginals do not sum to the same
+  total.
+- A 1D biased rotation returns the identity instead of flipping the axis it just aligned
+  (`SO(1)` cannot represent a reflection).
+
+## 0.6.1 - 2026-07-15
+
+### Added
+
+- `examples/09_hard_decision_tree.py`: a hard oblique decision tree with strict argmax
+  routing and no relaxation, trained on an XOR checkerboard. PolyStep optimizes the hard
+  tree directly while OpenAI-ES and SPSA stall on its piecewise-constant loss; a
+  soft-tree Adam baseline is scored after hardening.
+
+### Changed
+
+- The HybridSubspace fused block-diagonal projection is rebuilt only when the projection
+  rotates. With the default `rotation_interval=0` the projection is static, so the
+  previous per-step `block_diag` rebuild was pure overhead.
+
+### Fixed
+
+- `mixed_precision=True` runs end to end. The barycentric and fused-softmax projections
+  cast the FP32 OT weights to the BF16 geometry dtype, `HybridSubspace` runs its
+  projection QR in FP32 (no BF16 CPU `geqrf`), and the cost evaluator matches float
+  inputs to the parameter dtype. Previously the first step raised a dtype mismatch.
+- The barycentric projection normalizes by the realized transport row sum instead of the
+  target marginal, so an unconverged Sinkhorn plan gives a translation-invariant step.
+  The softmax path is unchanged.
+- Full-space monolithic steps bound the default evaluation chunk to a fixed memory
+  budget, avoiding the `O(n_params^2)` config buffer that ran out of memory at the
+  default `chunk_size=None`.
+- `apply_biased_rotation` realigns the first search axis with the requested bias (QR
+  fixes the column only up to sign), so the biased search points toward descent rather
+  than ascent, and runs its QR and determinant in FP32 for BF16 inputs.
+
+## 0.6.0 - 2026-07-12
+
+### Added
+
+- `examples/08_direct_loss_minimization.py`: direct F1 minimization on an imbalanced
+  checkerboard where PolyStepES beats both a biased gradient (Adam+STE) and OpenAI-ES.
+- `experiments/runners/variant_sweep.py`: a reproducible sweep of the optimizer variants
+  (solver, representation, block strategy, adaptation flags, schedules, geometry) across
+  nine tasks, ranked by forward-eval budget with a state-mutation self-check that flags
+  any variant whose state never changed.
+
+### Changed
+
+- CMA covariance updates use the covariance-metric offspring step `y = sqrt(C_diag) * z`,
+  so the evolution paths and rank-mu match the covariance-scaled sampling.
+- AdaptiveSubspace runs the per-step basis QR on the GPU in fp32 for CUDA targets,
+  several times faster than the previous CPU round-trip for large models.
+
+### Fixed
+
+- `SinkhornSolver` rejects `max_iterations < 1` and validates user-supplied marginals
+(finite, nonnegative, positive mass) instead of clamping to an infeasible plan.
+- `PolyStepES` validates `dim`, `num_particles`, `epsilon`, and `step_radius` at
+  construction.
+- CMA covariance and step-size adaptation are disabled with a warning under
+  non-monolithic block strategies, where they previously never updated.
+- `trust_region` now records its predicted improvement independent of `biased_rotation`,
+  so the step-radius multiplier adapts whenever `use_quadratic_model` and
+  `num_probe >= 2` hold instead of staying frozen at 1.0. `trust_region` also
+  auto-enables the quadratic model.
+- Heaviside stall detection uses the correct generation index (`generation + 1`).
+
+### Removed
+
+- Dead `compute_ot_weights` export; the rank-mu path uses the transport masses directly.
+
+## 0.5.0 - 2026-07-07
+
+### Added
+
+- `PolyStepES` and `minimize`: an ask/tell interface so PolyStep drops into
+  evolution-strategy harnesses (evosax, NeuroEvoBench) as a gradient-free optimizer.
+- `examples/07_binary_net_no_ste.py`: STE-free binary-net training, PolyStepES vs
+  OpenAI-ES.
+- `experiments/bench_ask_tell.py`: ask/tell comparison against a Gaussian ES on the
+  synthetic suite.
+
+### Changed
+
+- CMA covariance rank-mu now estimates per-coordinate variance from the
+  transport-weighted polytope vertices, and drops the per-step scatter allocation.
+- CSA calibrates the step-size against the running norm of the evolution path instead of
+  sqrt(n), so sigma no longer collapses on OT-scale steps.
+- Anderson acceleration uses a Tikhonov ridge solve and only accepts steps that do not
+  regress the dual objective.
+- Newton refinement keeps its step only when the quadratic model says it beats the plain
+  OT step.
+- Amortized momentum coasts along the OT step, not the OT-plus-momentum step.
+
+### Fixed
+
+- Multi-fidelity screening no longer feeds its own dampened cost back in, which had
+  permanently starved low-contrast directions.
+- Sinkhorn fixed-iteration mode reports convergence on a finite result, so
+  ProgressiveEpsilon stops inflating epsilon.
+- `adaptive_omega` no longer erases the divergence back-off within the same check.
+- Guarded the barycentric projection against a zero source marginal.
+
+## 0.4.0 - 2026-06-21
+
+### Added
+
+- `solvers/_shared.py`: one set of shared solver helpers (FP32 promotion, device-side
+  non-finite cost sanitization, device/dtype-aligned marginals and warm-start duals) used
+  by the Sinkhorn, softmax, and tempered-softmax solvers.
+- `experiments/scripts/bench_eps_rescale.py` and `tests/test_input_validation.py`.
+
+### Changed
+
+- Fused-softmax now honors every `scale_cost` mode (`'mean'/'max_cost'/float/None`),
+  matching `SoftmaxSolver`; the compiled kernel no longer scales internally.
+- `TemperedSoftmaxSolver` gains FP32 promotion, non-finite handling, and an
+  autocast-disabled softmax (previously absent).
+- Warm-start dual re-centering uses the valid gauge `f -> f+c, g -> g-c` instead of
+  independent mean subtraction, which perturbed the iterate under `omega != 1`.
+- `BatchedLinearEvaluator` applies the real activation modules (exact
+  `LeakyReLU.negative_slope`, `GELU.approximate`) rather than hardcoded functional
+  defaults; non-default `Flatten` falls back to vmap.
+- Low-level `PolyStep.step()` forwards `init_eps` like the optimizer path.
+
+### Removed
+
+- Low-rank Sinkhorn (`SinkhornSolver.{rank,gamma,auto_rank_threshold}`,
+  `_solve_low_rank`, `SinkhornResult.{_Q,_R,_g_lr}`, the `solve(seed=...)` arg, and the
+  `rank` params on `PolyStepOptimizer` / `PolyStep`).
+
+### Fixed
+
+- Re-validate `epsilon > 0` inside `SinkhornSolver.solve()` (schedules mutate
+  it); reject `check_every < 1`, `num_probe < 1`, and `scale_cost` of 0 / non-finite.
+- Mezzadri rotation sign correction treats `sign(0)` as `+1` so an underflowed QR
+  diagonal can't null a column; dropped a redundant `Q.clone()`.
+
+## 0.3.0 - 2026-05-27
+
+Cleanup. No public API changes.
+
+### Changed
+
+- `solvers/sinkhorn.py`: dropped two GPU->CPU syncs per Sinkhorn solve. `cost_scale` (the
+  warm-start `clamp_` bound) stays on-device as a 0-d
+  tensor, and `ent_reg_cost = <f,a> + <g,b>` resolves both inner products
+in one host transfer via `torch.stack([...]).sum().item()`. Same pattern applied to the
+low-rank convergence loop and its `err_a/err_b` marginal transfers.
+- `_step_blockwise.py`: per-block fused-softmax `ent_cost_tensor.item()` calls now defer
+  to a single `torch.stack([...]).sum().item()` after the per-block loop, mirroring
+  `block_disp_terms` and `block_model_loss_terms`. Saves `O(num_blocks)` syncs per step
+  in the fused-softmax path.
+- `cma.py`: `update_step_size_csa` now accepts an optional pre-computed `p_sigma_norm`;
+  both call sites (`_step_common.update_cma_state`, `_step_monolithic.step`) forward the
+  norm they already paid for in `compute_heaviside_sigma`, removing one redundant
+  `torch.norm(...).item()` per CMA generation.
+
+### Fixed
+
+- `CMAAdaptiveSubspace.rotate()` now forwards `transport_matrix`, `X_vertices`, and
+  `X_current` to the wrapped `AdaptiveSubspace`, so the `'ot_bias'` rotation mode fires
+  when CMA is enabled (it previously fell through to random rotation).
+- `docs/api_overview.md`: the `PolyStepOptimizer.step` snippet used a zero-argument
+  lambda; replaced with the real `closure(batched_params) -> losses` signature via
+  `NNCostEvaluator`. The `SparseRandomProjection` example used `input_dim` / `output_dim`
+  keyword args; the actual constructor takes `full_dim` / `subspace_dim`.
+- Tightened two test tolerances that were 100-1000x looser than the solver's convergence
+  threshold (`tests/test_numerical_stress.py` row-sum checks;
+  `tests/test_ablation_solvers.py` tempered-softmax-vs-greedy match).
+
+## 0.2.2 - 2026-05-21
+
+Codebase cleanup, test modernization, and documentation correctness pass.
+
+### Breaking
+
+- Vertex generator API (`get_orthoplex_vertices`, `get_simplex_vertices`,
+  `get_cube_vertices`) signature changed from `(dim, origin, radius)` to
+  `(dim, *, radius=1.0)`. The `origin` parameter has been removed - all generators now
+  produce unit-radius templates centered at the origin; translation is handled by the
+  solver. This affects direct callers only; `PolyStep` and `PolyStepOptimizer` are
+  unaffected.
+
+### Fixed
+
+- `compute_ot_weights` in `cma.py`: replaced linear softmax normalization with
+  entropy-based OT weight computation matching the theoretical derivation.
+- `@torch.inference_mode()` migration across `transform.py` (removed unused
+  `import warnings`).
+- README and API overview: `max_iteration` -> `max_iterations` (matching the actual
+  `PolyStep` dataclass field).
+- API overview: `AdaptiveSubspace` and `LinearSubspace` code snippets updated from stale
+  direct constructors to the `from_layout()` factory.
+- API overview: `SolverState(X=X_init)` -> `solver.init_state(X_init)`.
+
+## 0.2.1 - 2026-05-14
+
+### Added
+
+- `examples/06_loihi_snn_polystep.py`: end-to-end skeleton for a Loihi 2-style two-stage
+  workflow. Stage 1 pretrains a hard-LIF MNIST SNN with PolyStep using
+  `PSTORCH_CONFIGS["snn"]`. Stage 2 adapts only the writable subset a real Loihi 2 chip
+  exposes at runtime (`fc2`, per-population `vth`, and `beta`; about 1.3% of model
+  parameters) under an `N(0, 1)` Gaussian input shift. Stage 2 uses TENT-style
+  safeguards: mixed-batch (half clean / half shifted), rank-8 probing on the writable
+  subspace, and two probes per step. Both stages use best-test early stopping (patience
+  4) since zeroth-order test curves are noisier. Shifted-test evaluations share a fixed
+  seeded noise mask across pre / post / baseline so the reported recovery is a paired
+  comparison. On default settings the example reaches ~83% best clean accuracy and a +13
+  pp paired shift-recovery over the frozen-readout baseline with near-zero clean-accuracy
+  degradation in about 17 minutes on a single GPU. Reported numbers shift by a few
+  percentage points run-to-run on CUDA because of non-deterministic cuBLAS reductions,
+  but the qualitative recovery is robust. The host loop is backend-agnostic:
+  `LoihiSpikeEvaluator` is the single swap point against a Lava `netx` deployment path.
+
+## 0.2.0 - 2026-05-14
+
+Dependency floor lift, native fast-path adoption, and performance pass for PyTorch 2.12.
+
+### Changed
+
+- Minimum Python is now 3.11 (was 3.10); minimum PyTorch is now 2.8 (was 2.4); NumPy
+  floor is now `>=2.0`.
+- Ruff `target-version` bumped to `py311`. Classifiers updated to advertise 3.11 / 3.12 /
+  3.13 / 3.14.
+
+### Performance
+
+- `SinkhornSolver` convergence loop: batched the per-check scalar measurements (err_a,
+  err_b, dual norm, Lyapunov) into a single device-to-host transfer, and switched the
+  Anderson-acceleration Lyapunov accept gate to a device-side `torch.where`. Eliminates
+  4-7 GPU-CPU syncs per `check_every` interval. Microbench (200 iter, n=m=512, fp32 on
+  RTX 5090, full Anderson + adaptive omega): **490 ms -> 48 ms (~10x)**.
+- Block-wise step (`_step_blockwise.py`): per-block displacement and model-loss scalars
+  are accumulated as device tensors and reduced to the host once per step instead of
+  twice per block.
+- `NNCostEvaluator.evaluate` and `BatchedLinearEvaluator.evaluate` now use
+  `@torch.inference_mode()` instead of `@torch.no_grad()` (matches the rest of
+  `cost_nn.py`).
+
+## 0.1.0 - 2026-05-03 (Initial Release)
+
+First public release alongside the arXiv preprint *Training Non-Differentiable Networks
+via Optimal Transport*.
+
+### Core
+
+- `PolyStepOptimizer` for gradient-free training of any `nn.Module`
+- Softmax solver (default; fast path) and log-domain Sinkhorn solver with warm-started
+  dual potentials
+- Polytope sampling (orthoplex, simplex, cube) with random rotations
+- Vectorised NN evaluation via `torch.func` + optional `torch.vmap`
+- High-level `train()` API with callbacks and early stopping
+
+### Subspace compression
+
+- `LinearSubspace` - random projection
+- `AdaptiveSubspace` - rotating orthogonal projection
+- `HybridSubspace` - per-layer projections (recommended)
+- `CMAAdaptiveSubspace` - CMA-ES covariance adaptation
+- `SparseRandomProjection` - for models with 100M+ parameters
+
+### Scalability
+
+- Block-wise per-layer OT decomposition
+- Chunked cost evaluation for memory control
+- `torch.compile` support on GPU hot paths
+- Mixed precision (BF16 model + FP32 solver)
+
+### Layers
+
+- `VmapSafeMultiHeadAttention` - attention compatible with `torch.vmap`
+- `VmapSafeLSTM` - LSTM compatible with `torch.vmap`
+
+### Examples and experiments
+
+- 5 runnable demos under `examples/` (quickstart, SNN, RL, MAX-SAT, MNIST)
+- Paper-reproduction harness under `experiments/runners/` covering SNN, INT8, argmax,
+  staircase, hard MoE, MAX-SAT (100K-1M vars), MNIST, ETTh1 timeseries, RL, and GPT-2
+  head-only fine-tuning
+- Ablations: OT vs softmax solver, epsilon / radius / particles / subspace grid, MAX-SAT
+  scaling

{polystep-0.9.0 → polystep-0.10.1}/CITATION.cff

@@ -6,8 +6,8 @@ authors:
   - family-names: Le
     given-names: An T.
 license: Apache-2.0
-version: 0.9.0
-date-released: "2026-07-27"
+version: 0.10.1
+date-released: "2026-08-01"
 repository-code: "https://github.com/anindex/polystep"
 keywords:
   - gradient-free optimization