Skip to content

Release 0.8.396

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

The headline of this release is the move to Chemprop 2.2.4, which also brings a handful of new model-tuning knobs to the Chemprop template. The rest of the release is a concise batch of endpoint, pipeline, and tooling improvements.

Chemprop 2.2.4

Both the PyTorch/Chem training and inference images are rebuilt on Chemprop 2.2.4 (training + inference parity), and the model template now exposes new tuning options that come with it:

  • Learning-rate schedule — warmup_epochs, init_lr, max_lr, and final_lr are now settable hyperparameters. Defaults match Chemprop's, so existing models are unchanged.
  • Per-layer FFN dimensions — ffn_hidden_dim now accepts a list (e.g. [1024, 256, 64]) for a tapered prediction head, in addition to a single fixed width.
  • CheMeleon foundation-model fine-tuning — initialize the message-passing layers from the CheMeleon foundation model via from_foundation, with freeze_mpnn_epochs to stabilize the FFN head before unfreezing the encoder.

Endpoints

  • New 2D (salt-keeping) + 3D feature endpoint, plus clearer endpoint naming.
  • Dynamic batch sizing for meta endpoints, and async child endpoints now inherit max_instances.

Pipelines & Plugins

  • The ml_pipeline launcher can now run a single script by its full filename.
  • Refined plugin resolution.

Tooling

  • Secret-leak detection — a pre-commit config plus a CI workflow (PR-scoped, local binary) to catch accidental credential commits.
  • Improved chunk-progress display for long-running operations.

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