v0.1.18

v0.1.18 (2026-05-02)

• 🐛 Bug Fixes (DDP + mosaic data)
  • Default sampler_type='auto' now picks the DDP sampler when a
    process group is initialized. Previously 'auto' silently fell
    back to MultiBatchSampler (a rank-agnostic sampler), so DDP
    runs computed each batch on every rank in parallel — correct
    but with no throughput gain over single-GPU. Users who already
    passed sampler_type='ddp' explicitly are unaffected.
  • MyDistributedSampler now derives its shuffle order from a
    seeded random.Random instance (cross-rank-consistent for the
    dataset visit order, rank-specific for the within-dataset
    shuffle), and properly initialises the base
    DistributedSampler. Previously it used the global Python
    random module, so each DDP rank sampled a different sub-batch
    at the same step. With non-uniform per-sub-batch modality
    combinations (mosaic data), this produced different encoder
    graphs per rank and caused NCCL all-reduce to hang under
    find_unused_parameters=False (Lightning default), eventually
    triggering a watchdog timeout.
  • Heads-up — DDP reproducibility: the DDP sampling order has
    changed as a side-effect of the fix. Existing seeded DDP runs
    will produce different numerics; checkpoints from prior
    versions still load and continue training, but the post-fix
    sampling sequence is not bit-equivalent to the pre-fix one.
    Single-GPU users (using MultiBatchSampler) are unaffected.
• 🐛 Bug Fixes (API hardening)
  • MIDAS.configure_optimizers no longer raises AttributeError
    when entered through the simpler configure_data path
    (load_optimizer_state was only set by
    configure_data_from_dir / configure_data_from_mdata /
    load_checkpoint).
  • MIDAS.configure_data default batch_names now use f-string
    formatting (f'batch_{i}') instead of the literal string
    'batch_%d' repeated len(datalist) times.
  • Bad ATAC configuration in configure_data now raises
    ValueError instead of calling exit() (which killed the
    Jupyter kernel without a traceback).
  • download_file now accepts both str and pathlib.Path for
    dest_path. The signature was annotated str but the body
    called .name.
  • Encoder.forward no longer mutates the caller's batch dict.
    The mask multiply is now out-of-place; the previous in-place
    data[m] *= mask corrupted upstream tensors for any modality
    without a trsf_before_enc_* transform. Mathematically
    equivalent (the mask is a 0/1 modality-presence indicator, and
    calc_recon_loss already multiplies the loss by the same
    mask), but makes the encoder safe to re-call on the same
    batch (e.g. predict's mod_latent / translate paths).
  • VAE.forward no longer wraps the PoE call in a bare
    try/except that swallowed real errors with a malformed
    logging.debug call.
• ✅ Tests
  • Added tests/test_invariants.py pinning down the bugs above
    plus the DDP sampler determinism fix (cross-rank disjoint
    indices, set_epoch actually changes ordering).
• 📚 Documentation
  • Each basics demo now exposes a single # === GPU configuration ===
    block (GPUS + STRATEGY) at the top so switching from
    single-GPU to multi-GPU only requires editing two values.
  • Removed the redundant standalone advanced/multi_gpu.rst
    tutorial — its contents now live inline in the basics demos
    where the failure modes would actually be encountered.
  • README: removed the duplicated MuData section (the from_mdata
    path is one link away in the docs), corrected the Quick
    Example comment about input format, and fixed the License
    badge link.
• ⚙️ Packaging
  • Version is now single-sourced from pyproject.toml;
    scmidas.__version__ and the Sphinx release both read it via
    importlib.metadata.version("scmidas") instead of duplicating
    the literal in three files.
  • Relax the torch pin from >=2.5,<2.6 to >=2.5,<3 (and the
    matching torchvision / torchaudio companions). The previous
    <2.6 cap was a workaround for a suspected Lightning-DDP
    incompatibility; torch 2.8 has now been verified end-to-end in
    the mosaic DDP path (1000-epoch run with UMAP and numerics
    consistent with the single-GPU baseline), so users on torch 2.6
    / 2.7 / 2.8 no longer have to manually override the pin.