diff --git a/CHANGELOG.md b/CHANGELOG.md index f6b03f769..dc0565e9b 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,5 +1,60 @@ # Changelog +## 0.11.1 — 2026-11-20 + +- **Fix autograd-hook VRAM leak in live-endpoint critic captures (PR #585).** + Break the closure reference cycle across all six critic hook sites (`_unified_v11`, + `_unified_v4`, `_unified`, `_t2_critic`, `_nash_v3`, `_minimal`) where `_live_captures` + entries held strong references to autograd-attached `unscaled` logit tensors via + `register_hook` lambdas, causing linear GPU memory growth during long training runs + (e.g., `920 -> 3,898 MiB` over 800 CIFAR-10 steps). Adds CPU & CUDA weakref / flat-VRAM + regression tests (`tests/test_unified_critic_vram_leak.py`), flattening peak VRAM to + `953–968 MiB` across 2,000 steps. +- **Universal 1st-order `fwAD` Laws 2–4 and function-space equilibrated optimizer for `unified_v11` (PR #587).** + Fixes `unified_v11`'s 4D high-resolution image failure while preserving `0` + `torch.autograd.grad`, `0` `svd`/`eigh`/`pinv`, exact `179` `Recipe` fields, and untouched + frozen `_unified/` & `_unified_v4/` baselines: + - **Law 2 (`particlegan/_unified_v11/penalty.py`):** Replaces the `dim <= 1024` and + `dim > 1024` finite-difference image secant branches (BF16 ULP underflow at + `img_delta=0.02`, ReLU `O(1/delta)` kink impulses, mixed real+fake BN stat shift, and + 6× over-strong logit gauge) with one universal 1st-order `fwAD.dual_level()` per-sample + Born-Infeld critic penalty (`live_endpoints_reused=True`, `eff_floor=0.0005/dim`, + `eff_gauge=min(gamma_gauge, 0.005)`, temporary BN `eval()` + BF16 dual-tangent cast + hook `_bn_dual_bf16_hook`). + - **Law 3 (`particlegan/_unified_v11/loss.py`, `critic.py`):** Keeps `ema_margin`, + `ema_data_scale`, and `ema_push_res` live across all tensor ranks without caching 4D + images in `_t1c_capture`, and retains `fibre_rel_blend=0.50` scalar logit-space + relativistic coupling (`softplus(f - r)` / `softplus(r - f)`) when `loss == "non_saturating"` + and pixel-space `t1c` capture is skipped (respecting explicit user `loss` overrides). + - **Law 4 (`particlegan/_unified_v11/critic.py`) & Universal Equilibrated Optimizer (`optimizer.py`):** + Removes syntactic `L_count` / `deep_gate` and `ndim != 2` / `dim > 32` generator probe + gates; replaces them with an element-budgeted, BN-safe generator secant probe driving + checkpointed envelope-tracked `_ema_func_gain` + generator-only `stiff_gate`, degree-0 + `peak_rms` batched Tikhonov smoothing, exact `1×1` Conv2d `sqrt(out_ch / in_ch)` scaling, + and lifted `min(r, c) <= 512` per-singular-value Tikhonov polar. + - **Checkpoint compatibility note:** `UnifiedV11EquilibratedOptimizer` now checkpoints + `_ema_func_gain` and per-parameter `peak_rms` with strict validation and explicitly + rejects older `0.11.0` optimizer checkpoints rather than silently reconstructing missing + history. + - **Verified evidence (seed `0`, zero recipe overrides):** On real CIFAR-10 + (`benchmarks.cifar_toy`, `B=128`), `unified_v11` drops FID from `2730.72 -> 458.95` + (`N=200`, `s800`, `TV=0.109`, best `436.99` @ `s2000`, `8.22–8.49 it/s`, `953 MiB`) and + `2618.93 -> 728.10` @ `s800` -> `432.43` @ `s2000` (`N=16384`, `10/10` cov, + `8.17–8.56 it/s`, `967 MiB`, `1.67×–2.18×` faster than `empf`). Official 7,000-step + 100-Gaussians gate passes at `100/100` modes, `hq = 99.025%` (`+0.50%` over `0.11.0`), + `mode_tv = 0.0436`. Forge `discriminator_stability` Tiers 1–2 reaches `25/27` required + PASS (`11/11` T1, `14/16` T2 — `+1` net required gain over `0.11.0`'s `24/27`, gaining + `five_word_joint_hold` and `staggered100`, with `39/46` total graded PASS and 7 + non-passes transparently disclosed: `model_ddgan_gaussian_chain`, + `model_ae_ddgan_gaussian_chain`, `vector_unequal_width`, `interp_sparse_arc`, + `interp_img_patch_shift`, `interp_saddle_barycentric`, `t3_iso_scale_mass_8x`). + Historical `0.11.0` Forge receipts remain untouched. +- **Reproducible CIFAR-10 Transformer-G / ResNet-18-feature-D benchmark (PR #586).** + Add `benchmarks/cifar_toy/` and `tests/test_cifar_toy_benchmark.py` (registered in + `experiments/forge/history.py`) with SHA-256-verified frozen ImageNet ResNet-18 critic and + independent ResNet-34 Feature-FID/KID/coverage evaluator, pure public API, and matched + cross-trainer provenance digests. + ## 0.11.0 — 2026-11-20 - **Default Unified 4-Law research selection (`DEFAULT_RECIPE = "unified_v11"`).** diff --git a/pyproject.toml b/pyproject.toml index 3643b73c7..d315a7abf 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta" [project] name = "particlegan" -version = "0.11.0" +version = "0.11.1" description = "Composable PyTorch primitives for learnable particle priors, GANs, and DDGAN" readme = "README.md" license = "MIT" @@ -17,7 +17,7 @@ dependencies = [ [project.urls] Repository = "https://github.com/255BITS/ParticleGAN" -Documentation = "https://github.com/255BITS/ParticleGAN/blob/v0.11.0/docs/api.md" +Documentation = "https://github.com/255BITS/ParticleGAN/blob/v0.11.1/docs/api.md" [project.optional-dependencies] mog = ["scipy>=1.10,<2"]