BASALT v1.2.2 improves installation reliability, model reproducibility, and documentation while retaining the Python 3.12 workflow introduced in v1.2.1.
Highlights
- Resolves the base environment in a single Conda transaction using the validated CPU PyTorch stack.
- Pins the BASALT model revision and adds resumable downloads, connection timeouts, and fallback sources.
- Improves mainland-China installation with Micromamba and selectable Conda mirrors.
- Isolates VAMB and other conflicting optional dependencies from the BASALT base environment.
- Expands documentation for LorBin integration, multi-assembly study design, troubleshooting, reproducibility, and BASALT-Air.
Fixes
- Removed the duplicate uv/pip installation that could download an additional PyTorch and large CUDA dependencies.
- Added
--hf-timeoutfor blocked Hugging Face connections. - Pinned the default model revision to
bc98b102522d1c80dd8c2594df4ab3155438320e. - Prevented VAMB 5.0.4 from replacing the validated BASALT PyTorch runtime.
- Standardized model discovery through one absolute
BASALT_WEIGHTpath. - Removed obsolete Conda channels and separated legacy CheckM from the maintained CheckM2 workflow.
Validation
BASALT v1.2.2 was installed and tested on Ubuntu 22.04 x86-64 with Micromamba. The BASALT launcher, Python environment, CheckM2 database, required executables, 81 model checkpoints, five ensemble descriptors, and CPU model loading all passed validation. The documentation also passed strict Sphinx and external-link checks.