Enhanced sampling that never leaves the GPU.

GLUED is an OpenMM plugin. Collective variables, bias deposition and the chain-rule force scatter all run as kernels inside OpenMM's per-step pipeline, so positions stay on the device between integration steps. Nothing crosses PCIe and the CPU never waits.

Build from source View on GitHub Examples

OpenMM 8.x on the CUDA, OpenCL and Reference platforms. Linux, Windows and macOS. Academic license, source build only. Measured against openmm-plumed on the benchmarks page.

opes_quickstart.py

 1import openmm as mm
 2import glued
 3
 4# One force. Temperature and PBC are set once.
 5force = glued.Force(pbc=True, temperature=300.0)
 6
 7# Register a CV from a plain Python list.
 8phi   = force.add_dihedral([4, 6, 8, 14])
 9
10# Bias it. OPES needs only sigma and a bias factor.
11force.add_opes(phi, sigma=0.35, gamma=10.0, pace=500)
12
13# Standard OpenMM from here.
14system.addForce(force)
15simulation.step(10_000_000)

One GPU context for CV evaluation, bias deposition and force scatter.

GLUED subclasses OpenMM's ForceImpl and adds its own NVRTC-compiled kernels to the per-step compute pipeline. There is no host-side buffer to download into and upload back. The benchmarks show what that round-trip costs a PLUMED bridge on the same hardware.

Positionsnever leave posq[]

Forcesatomic-add into the OpenMM force buffer

Host sync per stepnone

Host trafficCOLVAR logging only, if you enable it

Collective variables

Distance, angle, dihedral, centre of mass, radius of gyration, coordination, RMSD, path, contact map, PCA, and a TorchScript model of your own.

Browse the CV catalog

Bias methods

Well-tempered and parallel-bias MetaD, OPES and OPES Expanded, walls, harmonic and moving restraints, ABMD, extended Lagrangian, EDS, MaxEnt.

Browse the bias catalog

Multi-GPU shapes

One system across cards, N replicas for REUS or T-REMD, or multi-walker MetaD and OPES with one grid shared across GPUs.

See the multi-walker recipe

Build it once. Run on every platform OpenMM does.

The Reference, CUDA and OpenCL platforms build from one CMake invocation on Linux, Windows, WSL2 and macOS. There is no pip or conda package. You clone and build.

Installation guide View on GitHub Benchmarks