Skip to content

Future Work

This is the public, openly-scoped roadmap. We keep deferred work recorded so outside contributors can pick up well-defined pieces. If you want to take one on, open an issue to claim it.

Items here are intentionally not in the current release. They're scoped enough to start, but each is a meaningful project on its own.

Off-Modal Dockerfiles (per eligible model)

Generate a Dockerfile + requirements.txt per model so models can run outside Modal. Feasibility varies by model:

  • Eligible (realistic): models with no GPU-at-build step, weights available from a public source (HuggingFace Hub or a direct URL), and a standard public base image. The model-source layer maps to a plain COPY; the weight-download step maps to a RUN with build secrets for credentials.
  • Hard / out of scope: models that compile CUDA/flash-attn extensions during the image build (these need build-time GPU access, which standard docker build doesn't provide), and any model whose base image is a private/closed registry.

The split is the work: define per-model eligibility, generate Dockerfiles for the eligible majority, and document the tail.

Benchmarks

Wire standard biological benchmarks (e.g. ProteinGym for variant-effect prediction) into the catalog so each model's comparison.yaml can cite reproducible numbers rather than paper-reported ones.

Self-improving model-implementation skill

The "add a model" skill should learn from each accepted contribution — capturing recurring review feedback into the template and the guide so the next contribution starts further along.

Faster builds

Adopt BuildKit-style caching and shared base layers to cut cold-build times for the heavier (conda/GPU) models.

Self-healing weight bake

Every model's setup_model() / @modal.enter trusts the build-time weight bake — nothing verifies at runtime that the baked image actually contains the weights. An incomplete or stale build-cache layer can therefore crash-loop silently (e.g. a cached image missing a required config file). Proposed: a commons-level runtime required_files check at container start that falls back to download_model_assets() when the bake is incomplete, turning a silent crash-loop into a slower-but-successful cold start. Deferred: it's a commons runtime-path change touching every model's startup, and fresh deploys bake correctly.

Have something else you think belongs in the catalog? Open an issue — see CONTRIBUTING.md.