Skip to content

Release 0.8.384

Need Help?

The SuperCowPowers team is happy to give any assistance needed when setting up AWS and Workbench. So please contact us at workbench@supercowpowers.com or on chat us up on Discord

This release upgrades the 3D descriptor energy model, expands plugin / ml_pipeline script support, adds model promotion primitives, and lands a batch of endpoint, performance, and robustness improvements. The headline is the move to quantum-grade conformer energies for the 3D feature endpoint.

GFN2-xTB Conformer Energy Ranking

The smiles-to-3d-v1 endpoint now ranks conformers with GFN2-xTB semi-empirical quantum energies (via the tblite library) instead of MMFF94s force-field energies. MMFF94s rankings are known to be poor for flexible and polar molecules — we measured near-zero rank correlation against GFN2-xTB on common drug-like compounds — which biased the Boltzmann-weighted averages toward the wrong conformers. Geometry is still MMFF94s-optimized; only the energies that set the Boltzmann weights now come from xTB, which improves all 74 ensemble-averaged 3D features at once.

  • Transparent fallback. If tblite is unavailable or a molecule fails to converge, the pipeline falls back to MMFF94s/UFF. The new desc3d_energy_method diagnostic column records which model actually produced the weights (GFN2-xTB, MMFF94s, or UFF) so a fallback is never silent.
  • Full endpoint only. xTB adds ~0.1–0.5 s per conformer; it runs on the async full endpoint where the per-molecule budget accommodates it.

Full write-up: 3D Molecular Descriptors.

Plugin & ML-Pipeline Scripts

  • workbench: and plugin: script resolution for ml_pipeline scripts — pipelines can reference scripts shipped with Workbench or supplied by a plugin, with path resolution validated before AWS submission.
  • schemas support on scripts, plus an expanded plugin example to demonstrate the pattern.

Model Promotion

  • Model.copy() copies a model to a new name (with a corrected return type), the building block for promotion workflows.
  • Example pipelines now include model-promotion script/logic, moving promotion toward a first-class pipeline step.

Endpoints & Pipelines

  • Endpoint outputs in pipeline DAGs — pipeline_meta is now generic enough to model endpoint outputs as downstream nodes; MetaEndpoint auto-imports.
  • Async progress heartbeat for long-running async inference, so liveness is visible during multi-minute jobs.
  • Realtime endpoint scaling added, with notes toward a future scale-out policy.
  • Cross-pipeline dependency groups — depends_on relationships can now span pipelines, and select() takes a downstream= arg to pull in downstream nodes.

Performance

  • Device-responsive inference & SHAP — more work runs on GPU when available.
  • Cached featurized MolGraphs for chemprop, plus training/inference timing reports to diagnose slow runs.

Robustness

  • @aws_throttle decorators on describe-style calls and a more robust SageMaker config (additional retries, larger connection pool) to ride out API throttling.
  • ModelError with a timeout message on the back-off path, and ServiceUnavailable added to caught exceptions.
  • Orphaned-artifact tooling — a script to find orphaned artifacts and prune pipelines with missing sources.

Questions?

The SuperCowPowers team is happy to answer any questions you may have about AWS and Workbench. Please contact us at workbench@supercowpowers.com or on chat us up on Discord