Skip to content

v0.2.2

Latest

Choose a tag to compare

@phduytien phduytien released this 19 Nov 21:56
· 2 commits to main since this release

Compatibility

  • RFProteína backends. Select the execution backend with
    inference_backend: eager, compile, or bioir. compile remains
    the default, so existing compiled runs keep their behavior.

Breaking Changes

  • RFProteína is a numerical engine. BioIR runs the encoder, the
    velocity network, the integrator, and the decoder. The host application
    owns generation, partial diffusion, search, scoring, sequence
    constraints, motif assignment, and structure output. The same host code
    runs on every backend.
  • Public API. sample, sample_partial, redesign, inverse_fold,
    and search are removed. Advance a prepared state with
    partial_simulation, which can resume from the previous clean
    prediction. For custom feature preparation, integration, or readouts, use
    RFProteinaTrunk.forward_features.
  • Explicit sampler settings. from_reference_state_dicts requires
    simulation_config. The built-in defaults serve tests and toy examples
    only. They do not match the released sampler settings, so a rollout that
    relies on them follows a different trajectory.
  • Batch lengths. Sampling uses the actual length of each batch. Compiler
    buckets stay inside the native trunk, so seeds that came from the older
    bucket-padded path can produce different outputs.

Performance

  • RFProteína sampling. On an H200 with 60 steps and 180 trunk passes,
    BioIR is faster than the compiled backend in every configuration we
    measured. Single monomers run 2.3× to 4.7× faster, and batches of 8 and
    32 run 1.4× to 1.9× faster. Peak memory stays within 1.7% of the
    previous BioIR path.

See the full release notes.