0.2.1
Released 2026-10-30.
Compatibility
- Portable wheels. Each Linux wheel supports CPython 3.12, 3.13, and 3.14 on its CPU architecture, either x86_64 or aarch64. The host needs NVIDIA driver 580 or newer and glibc 2.34 or newer. An NGC image and a local CUDA Toolkit are no longer required to install or run the wheel. Some dependencies need build tools; Triton also needs a C compiler and Python headers when it first runs a model.
Breaking Changes
- CUDA graph configuration.
optimize()accepts graph region roles or paths. Former module keys such astoken_transformerandstructure_pairformernow raiseValueError; use the enclosing region’s role or path instead.
Performance
- Protenix sampling. An optional frame-reuse mode skips some denoiser evaluations during a 200-step rollout. It is off by default and is an approximation, so enabling it can change predictions.
- Attention. Protenix atom attention and pair-biased attention use faster paths for supported inputs.
Fixed Issues
- Graph replay. Each CUDA graph owns its cuBLAS workspace, preventing invalid memory access when another graph is evicted.
- RFProteína checkpoints. Released checkpoints with LoRA adapters can
be loaded directly through
from_reference_state_dicts. - Template chain IDs. Templates with blank PDB chain IDs now raise a descriptive error instead of failing with an attribute error.
- Dependencies. Runtime dependency minimum versions include fixes for known security issues.