v2026.4.0

Noether Release Notes

This release moves Noether to calendar-based versioning (#189) and ships a major round of new recipes, a higher-level training API, broader AB-UPT capabilities, and a measurable quality improvement on the DrivAerML baseline.

Highlights

Improved AB-UPT baseline on DrivAerML

RoPE scaling (#184) combined with the new Muon optimizer (#158) lifts AB-UPT performance on DrivAerML over the previous baseline. See the full comparison in the W&B report.

New recipes

• AB-UPT showcase with scalable query-based inference (#153) — an end-to-end DrivAerML walkthrough covering data fetching, training, and evaluation. Designed to be the recommended starting point for AB-UPT users.
• Heat transfer recipe (#146) — Neural surrogate for the SIMSHIFT Heatsink benchmark (819 simulations) predicting velocity, temperature, and pressure on 3D volume fields conditioned on simulation parameters. Supports both AB-UPT and Transolver++.
• Emmi-Wing HF subset recipe (#159) — small, fast-to-iterate aerodynamics recipe backed by the public EmmiAI/Emmi-Wing HuggingFace dataset, with train/val/eval splits.
• Aero CFD recipe (#143) — the previous tutorial promoted to a full recipe.

High-level training API

Training presets (#128) — DomainPreset encapsulates dataset stats, normalizers, pipeline defaults, and forward properties. Per-dataset training scripts (examples/train_*.py) now expose only experiment-level knobs (model, hidden dim, epochs, device, accelerator, paths).

AB-UPT capabilities

• Untied physics blocks (#177) — self_untied, cross_untied, perceiver_untied variants with per-branch weights via UntiedLinear (batched or grouped matmul). Removes the need to configure decoder-block counts per domain and enables cross-attention between branches with untied weights.
• Arbitrary domains (#139) (breaking) — replaces the hardcoded surface or volume split with a ModelDataSpecs dict, so AB-UPT can be applied to any number of named domains (gas, liquid, solid, multi-phase, …).
• Optional geometry encoder (#150) and Muon optimizer support (#158) (breaking) — Muon is paired with a secondary optimizer for non-2D tensors. QKV is split into separate Q, K, V matrices (breaks old checkpoints).

Workflow and infrastructure

• submitit-based multirun job submission (#164) (breaking) — replaces the custom sbatch --wrap path with submitit.AutoExecutor, giving Hydra multirun sweeps as SLURM array jobs with login-node validation. SlurmConfig field names now match AutoExecutor.update_parameters (name, slurm_partition, mem_gb, timeout_min, …).
• Development CLI (#142) — noether-development runs individual modules (datasets, models, trainers) in isolation with verbose logging for fast iteration without the full pipeline.
• EMA evaluation (#183, #192) — EmaCallback now accepts nested eval_callbacks that run on swapped-in EMA weights and emit metrics under an ema=<factor>/ prefix, removing the need for a separate eval run.
• Profiling callback (#161) — wraps the PyTorch profiler as a standard callback.
• Reproducibility (#179, #180) — DataLoader workers are seeded and use_deterministic_algorithms is enabled. Non-deterministic dataset ordering fixed.
• Dataset hooks (#154) — pre_getitem and post_getitem hooks on the base Dataset for lightweight per-sample customization.

Breaking changes

• #138 — shared attention renamed to self. Auto-converted with a deprecation warning during the transition.
• #139 — fixed surface or volume domains generalized to arbitrary domains. Replace AeroDataSpecs with the ModelDataSpecs dict.
• #142 — development CLI introduces new entry points (additive; review CLI docs).
• #158 — Muon support splits QKV into separate Q, K, V matrices. Old AB-UPT checkpoints will not load.
• #164 — submitit replaces the custom sbatch wrapper. Rename SlurmConfig fields. Remove --output and --error. Sweeps may not vary slurm: fields.
• #189 — versioning switched to CalVer. Manual release process. Release-please config removed.

Detailed changes

Features

• Rename shared → self attention by @Ndles in #138
• Expose training presets and high-level API by @Ndles in #128
• Convert tutorial to aero_cfd recipe by @HennerM in #143
• Generalize fixed surface/volume to arbitrary domains by @richardk53 in #139
• Make geometry encoder optional in AB-UPT by @HennerM in #150
• pre_getitem and post_getitem hooks on Dataset by @HennerM in #154
• Emmi-Wing HF subset recipe by @Ndles in #159
• Profiling callback by @HennerM in #161
• Heat transfer recipe by @HennerM in #146
• AB-UPT showcase with scalable query-based inference by @Ndles in #153
• Module-wise development CLI by @MauritsBleeker in #142
• submitit replaces custom sbatch wrap, adds multirun by @HennerM in #164
• Muon optimizer support by @kinggongzilla in #158
• Seed DataLoader workers and enable deterministic algorithms by @HennerM in #179
• Evaluate on EMA weights by @HennerM in #183
• Route DomainPreset.build_optimizer through optimizer discriminator by @Ndles in #187
• Basic timing and precision option in showcase preset by @Ndles in #188
• Reuse callback configs in EMA eval by @HennerM in #192
• Untied physics blocks by @HennerM in #177

Fixes

• Distributed training hanging with submit_job by @HennerM in #141
• calculate_statistics compatibility with any dataset subclass by @HennerM in #147
• Attempt HF download without HF_TOKEN by @HennerM in #152
• Support configs without custom top-level schema in submit script by @HennerM in #151
• move_item_to_device on MPS by @Ndles in #157
• Log total parameter count for non-composite models by @HennerM in #155
• Support normalization and denormalization on CUDA by @HennerM in #156
• StopIteration error when training on fractional epoch by @HennerM in #162
• Transolver uses Transolver attention, not regular attention by @MauritsBleeker in #166
• Missed import rename by @HennerM in #168
• Aero callback uses correct denormalization key by @MauritsBleeker in #169
• aero_cfd ShapeNet recipe by @HennerM in #173
• Default to unlimited job time by @HennerM in #176
• Non-deterministic dataset order by @HennerM in #180
• Avoid error when loading model with noether-eval due to stored computed fields by @kinggongzilla in #170
• Catch OOM in slurm submission by @HennerM in #190
• Flush tracker on eval and exception by @HennerM in #182
• resume_from_checkpoint for eval callbacks by @HennerM in #193

Docs

• Add "Understanding the Data Pipeline" page by @tcapelle in #140
• Minor documentation fixes by @MauritsBleeker in #160
• SEO enhancements by @HennerM in #175
• Increase visibility of recipes in docs and README by @Ndles in #181

Chore

• Fix leftover tutorial kind strings by @kinggongzilla in #144
• Use _get_int_env instead of os.environ.get by @ghtyrant in #145
• Trainer config validation and transformer input validation by @HennerM in #148
• Add description to periodic callback progress bar by @HennerM in #149
• CLAUDE.md and concise AGENTS.md by @HennerM in #163
• Upgrade torch to 2.11 by @kinggongzilla in #165
• Type guard to avoid mypy error in OptimizerWrapper by @kinggongzilla in #171
• Instructions on SIGTERM and CheckpointCallback by @Ndles in #172
• Tests for trainer eval, failing callbacks, grad scaler loading by @HennerM in #167
• Add ABUPT forward/backward benchmarks by @HennerM in #174
• Change RoPE params for DrivAerML by @HennerM in #184

CI

• Publish target release to PyPI by @ghtyrant in #136
• Fix mypy pre-commit hook by @HennerM in #191
• Convert versioning to calendar-based system by @hfurkanvural in #189

Tests

• Fix dataloader test on macOS by @HennerM in #186