TMol 0.1.62 includes the latest merged scoring and repacking changes and repairs release publication and pip installation.
- CUDA whole-pose LJ/LK and electrostatic energies accumulate block-pair totals in double precision before producing the usual FP32 scores. This prevents accumulation-order error from exceeding the existing scoring accuracy bounds when energy terms cancel.
- Annotated release tags are verified without rewriting local tag refs, and every fresh build is pinned to the verified release commit.
- Repacking reuses topology data and annealer buffers, releases consumed pages promptly, and rejects distant atom pairs before expensive exclusions. The changes preserve scoring formulas, RNG consumption and packing results, without new GPU-model-specific tuning.
- Missing Dunbrack rotamers use the nearest populated neighbor. Unused protonation-alternative hooks are removed. Hydroxyl torsion weights match Rosetta.
- PyPI provides 12 CPU wheels for PyTorch 2.14 and Python 3.11–3.14 across Linux x86_64, Linux ARM64 and Apple Silicon. GitHub provides these CPU variants and 34 CUDA variants, plus source and the ARM64 dependency-metadata repair script.
- Versioned wheel pages select an explicit CUDA/PyTorch pairing. Installation commands require prebuilt wheels for every dependency, so unavailable combinations fail instead of compiling locally.
- Every release candidate must pass fresh installation, dependency checks, native loading, protein scoring and finite gradients before publication; CUDA scoring is also checked on a GPU.
Precompiled CPU installation
python -m pip install "tmol==0.1.62" --only-binary=:all:On Linux, install CPU-only PyTorch first to avoid its CUDA runtime downloads:
python -m pip install "torch==2.14.*" --only-binary=:all: --index-url https://download.pytorch.org/whl/cpu
python -m pip install "tmol==0.1.62" --only-binary=:all:Precompiled GPU installation
For Linux with a CUDA 13.2-compatible NVIDIA driver, in a fresh environment:
python -m pip install "torch==2.14.*" --only-binary=:all: --index-url https://download.pytorch.org/whl/cu132
python -m pip install "tmol==0.1.62+cu132torch2.14" --only-binary=:all: \
--find-links https://uw-ipd.github.io/tmol/wheels/v0.1.62/cu132torch2.14/
python -m pip checkUse the matching variant page for other CUDA/PyTorch combinations. Pip selects Python and platform tags; it does not detect the GPU or choose CUDA. No local CUDA toolkit or C++ compiler is needed for these wheels.
On Linux ARM64 with CUDA 13, run the fingerprint-checked cuSPARSELt metadata repair after installing and before the dependency check:
curl -fLO https://github.com/uw-ipd/tmol/releases/download/v0.1.62/repair_cuda13_arm64_metadata.py
python repair_cuda13_arm64_metadata.py
python -m pip checkThe helper corrects the platform metadata of NVIDIA cuSPARSELt 0.8.0/0.8.1 while preserving native bytes, versions and dependencies.
This is a fresh build from the 0.1.62 release commit. The earlier v0.1.61 and v0.1.60 tags are preserved and their candidates are not reused.