Loading...
Searching...
No Matches
Version 2.10

Unreleased MPI buffer lifetime fix

  • Complete outstanding asynchronous domain-decomposition sends before resizing their buffers or destroying the PLUMED instance. This prevents a possible use after free during domain redistribution or shutdown.

Version 2.10 (Aug 25, 2025)

Version 2.10 is a major upgrade of PLUMED, with a lot of new features and an extensive rewrite of the core of the code. Starting with version 2.10, we will stop supporting older compilers. For this reason, we will likely extend the support of version 2.9 longer than usual. Before switching to version 2.10, users are invited to carefully read the following text.

  • Changes relevant for users:
    • PLUMED 2.10 requires a C++17 compatible compiler.
    • The passing of data between actions has been made more flexible. Now, in addition to passing scalars between actions, you can also pass vectors, matrices and functions computed on grids between actions. The ways in which you can use these new functionalities are detailed at this link. The old input syntax for most actions should work the same as it did in previous versions of the code. However if you investigate the log you will see that these old functionalities are reproduced by using shortcut actions that generate more complicated plumed inputs.
    • A number of functionalities in the code have been carefully optimized. An incomplete list includes grids (with an impact on METAD, sum_hills, and other grid-based methods) and overhead associated to the simulation of small systems, where the overall performance should be significantly better than in previous versions.
    • The additional flexibility comes at the price of some decrease in performance. In most cases this decrease in performance should be negligible or largely mitigated by the other optimizations that we implemented. We are very interested in feedback on PLUMED performances in real applications. If you have a real application where you can see a measurable slowdown when upgrading from version 2.9 to version 2.10, please open an issue on GitHub and report it. Ideally, you could provide a simple input file for the benchmark which highlights the performance regression.
    • It is now possible to use an embedded Python interpreter to implement collective variables and functions in Python. This is largely based on a previous implementation by Toni Giorgino that is described in this paper and has been integrated in PLUMED by Daniele Rapetti. This feature is provided as a separate plugin that can be linked against the proper Python interpreter and loaded at runtime with LOAD. Documentation about how to install and use this feature can be found in directory plugins/pycv.
    • A prototype of the COORDINATION collective variable with limited functionalities is now included in a CUDA implementation which is orders of magnitude faster when using large groups. This implementation has been contributed by Daniele Rapetti and is provided as a separate plugin that can be linked against the proper CUDA libraries and loaded at runtime with LOAD. Documentation about how to install and use this feature can be found in directory plugins/cudaCoord.
    • New CLTool plotswitch for tabulating any switching function and its derivative.
    • A new @ndx selector can be used to read GROMACS index files (see Specifying Atoms) without the need to explicitly declare a GROUP.
    • The LOAD action has been improved (see this pull request).
      • It can be used in any position in the input file.
      • It is now working properly when multiple Plumed objects are present (e.g., in multithread applications or when creating multiple Plumed() objects in Python).
  • New patches for MD codes:
    • A patch for GROMACS 2025 has been added, which is based on the native GROMACS interface to PLUMED and adds support for multiple walkers.
  • Changes leading to differences with previous versions:
    • GHOST now automatically makes molecules whole. In case you do not want it to do it, use the NOPBC flag.
    • In ANGLES you can no longer use the GROUP, GROUPA/GROUPB/GROUPC keywords. Explicitly list all the distances you want to compute using ATOMS1,ATOMS2,ATOMS3...
    • Action NBONDS no longer exists. The same effect can be achieved through a more transparent implementation that you can read about here: https://www.plumed-tutorials.org/lessons/23/001/data/Steinhardt.html
    • Action CENTER_OF_MULTICOLVAR no longer exists. You now simply use CENTER with the PHASES option and a vector as input for the weights.
    • Dimensionality reduction methods and landmark selection actions have a new syntax. A good introduction that explains how to use these actions can be found in this tutorial
    • if you use DISTANCES with the GROUP or GROUPA/GROUPB syntax each distance is no longer allocated to a position in space. You thus cannot use the GROUP or GROUPA/GROUPB variants of action with methods such as DUMPMULTICOLVAR, MULTICOLVARDENS or SMAC. To be clear if you use \DISTANCES with the ATOMS1/ATOMS2/ATOMS3/... syntax you can still use it with DUMPMULTICOLVAR, MULTICOLVARDENS or SMAC. We believe this second use case was the variant of the old version that was widely used.
  • Places where we strongly recommend using the new syntax:
  • Modules that were built from code taken from the old crystllization and multicolvar modules.
    • symfunc contains code for computing symmetry functions much of which was taken from the old crystallization module
    • volumes contains code for determing whether atoms are within a particular part of the box or not. This is largely built from code that was formerly in the multicolvar module
    • clusters contains code for determining if atoms are clustered together or not. This is build from code that was formerly in the crystallization module
    • gridtools contains code for doing kernel density estimation and manipulating functions on grids
    • contour contains code for computing Willard-Chandler dividing surfaces.
    • refdist contains code for calculating distances between configuations.
    • fourier contains code for computing fourier transforms of functions on grids
  • New contributed modules:
    • A new crystdistrib module by Jake McKibben, Gareth Tribello and Erik Santiso for computing order parameters that can be used to study the formation of molecular crystals.
    • A new metatomic module by Guillaume Fraux and Gareth Tribello integrating with https://docs.metatensor.org/metatomic/ to enable the use of arbitrary machine learning models as collective variables.
  • Changes relevant for developers:
    • Removed some shortcuts for ActionRegister.h (from the modules bias colvar function sasa and vatom) and CLToolRegister.h (from the module cltools), now the two headers can only be included with #include "core/ActionRegister.h" or #include "core/CLToolRegister.h".
    • plumed mklib can now process more than one cpp file. Files are combined in a single library.
    • When loading PLUMED at runtime using PLUMED_KERNEL, plumed is loaded with RTLD_LOCAL by default. This facilitates running multiple PLUMED versions, e.g. for benchmark.
    • On Linux, plumed-runtime executable can now find a working libplumedKernel.so by looking in the dlopen path (including LD_LIBRARY_PATH and stored RPATH). On MacOS, plumed-runtime executable can now find a working libplumedKernel.dylib by using a relative rpath. These changes should make this executable as functional as the normal plumed executable but with the advantage of loading symbols in a local namespace (see above).
    • A new benchmark tool has been added to time execution of sample input files and facilitate the comparison across multiple PLUMED versions.
    • You can now pass any scalars, vectors, matrices and functions (and forces on these objects) in and out of PLUMED using the functionality described here: https://www.plumed-tutorials.org/lessons/23/001/data/MDInterfaceI.html
    • You can also pass vectors that are scattered across the domains by using the functionality detailed here: https://www.plumed-tutorials.org/lessons/23/001/data/MDInterfaceII.html
    • You can use ActionShortcut to create complex inputs from a simpler initial input. The nest will then allow users to explore these more complex inputs.
    • You can use show_graph to create diagrams showing how values and forces are passed between the various actions in your input files.
    • Complete refactor of SwitchingFunction.cpp and SwitchingFunction.h, now adding new switching function is more straightforward and all the "book-keeping" can be done within a single class
    • Python (cython) wrappers now use nogil. This should facilitate integration with Python. See https://github.com/plumed/plumed2/pull/1129#issuecomment-2410867829 (thanks to Guillaume Fraux).

Version 2.10.1 (Jul 1, 2026)

  • All fixes as in 2.9.5
  • Fixed a pointer in GROMACS patch
  • Fixed issue in the PLANE command
  • Added check that the neighbor list does not fill RAM. Can be removed with export PLUMED_IGNORE_NL_MEMORY_ERROR=1
  • Fixed a bug with CONVERT_TO_FES and MINTOZERO
  • Fixed issue with roundoff of timestep when using GROMACS in single precision

Version 2.10.2 (Oct 19, 2026)

  • Fixed the GROMACS 2024 patch so that the potential energy is computed on every step where PLUMED needs it (ENERGY CV, OPES multithermal, well-tempered ensemble). See issue 1205.
  • Fixed conversion of negative (and zero) float scalars from mixed-precision MD codes. ENERGY printed as NaN with GROMACS in mixed precision because log10 was applied to the raw value. The rounding that recovers a clean timestep is now requested with the new PUT flag ROUND_TO_BASE_TEN, and used for the timestep only, since its grid is coarser than the float itself and was discarding up to 4 bits of the energy.
  • Enabled ENERGY (and energy biasing) with the native GROMACS 2025 ForceProvider interface: request energy/virial in the MD loop and apply energy biases after F_EPOT is accumulated.
  • Fixed an out-of-bounds write when a force is applied on ENERGY and the atoms are scattered over the domains. The forces of the MD code were rescaled over the total number of atoms rather than over the atoms that are local to the rank, which corrupted the heap and made OPES multithermal and the well-tempered ensemble crash on more than one MPI rank.
  • Documentation: ENERGY does include the GROMACS long tail (DispCorr) corrections, contrary to what the manual stated. See issue 567.
  • Fixed the GROMACS 2025 and 2026 patches so that PLUMED is told before the force calculation that GROMACS writes a checkpoint, as with the GROMACS 2024 patch. Actions that write their state on checkpoints, such as OPES_METAD with STATE_WFILE and no STATE_WSTRIDE, wrote it one step after each GROMACS checkpoint, and never at the last step. With more than one MPI rank per simulation, only the main rank wrote it, and the run deadlocked at the first checkpoint. See #1430.
  • Fixed -replex together with -plumed in the GROMACS 2025 and 2026 patches: with GROMACS in mixed precision, mdrun stopped at the first step because the bias was read into a double.
  • Fixed the GROMACS 2025 patch so that -multidir simulations without -plumed do not have to run in lockstep, which made them deadlock when they had different numbers of steps. As in the GROMACS 2026 patch, the multi-simulation record is now used only when PLUMED is active.
  • Fixed the virial with a bias on ENERGY (e.g. OPES multithermal or the well-tempered ensemble) and domain decomposition: only the first rank rescaled its share of the virial of the MD code, so the pressure was wrong, by orders of magnitude in some cases. As in PLUMED 2.9, it is now rescaled on all ranks.
  • Replica exchange with more than one rank per replica (e.g. GROMACS -replex or -hrex) hung after an exchange attempt on which no action needed the atoms: the exchange of positions started for the exchange was never completed. With small systems, the following exchanges could instead receive the positions of an earlier step.
  • GROMACS 2025 and 2026 patches: PLUMED can now stop the simulation, e.g. with COMMITTOR. The request reaches all ranks and the simulations that share state, and mdrun runs one more step after the next step at which GROMACS communicates signals (every nstglobalcomm steps, and between simulations that share state every 200 steps or more; with the modular simulator, it stops at the next neighbor-search step after that), writing a checkpoint at that last step.
  • GROMACS 2025 and 2026 patches: when the PLUMED input needs the potential energy, the forces that PLUMED applied on virtual sites were lost. They are now spread on the constructing atoms, on the steps when PLUMED uses virtual sites.
  • GROMACS 2025 and 2026 patches: with multiple time stepping, the potential energy passed to PLUMED on the fast steps lacked the slow terms (e.g. the PME mesh, by 650-1250 kJ/mol in a water test). mdrun now stops with an error if PLUMED needs the energy on such a step.
  • GROMACS 2025 and 2026 patches: errors raised on purpose by the PLUMED module, e.g. for replica exchange with a bias, are no longer reported as internal errors.
  • GROMACS 2025 and 2026 patches: a request of PLUMED to stop (e.g. COMMITTOR) was ignored in reruns. The rerun now stops after the frame on which PLUMED asks, as with the GROMACS 2020-2024 patches.
  • GROMACS 2025 and 2026 patches: with shell (polarizable) particles, PLUMED also ran on each shell relaxation step, as step 1: e.g. PRINT printed extra lines. It now only calculates its forces again on those steps, without updating, in MD and in reruns. When the PLUMED input needs the energy, PLUMED updates with the energy before the shells are relaxed, and mdrun prints a warning.
  • GROMACS 2025 and 2026 patches: with multiple time stepping and a PLUMED input that needs the energy, the constraint virial lacked the forces of PLUMED on the slow steps: they are now also added to the force buffer that GROMACS uses for it.

Version 2.10.3 (in progress)

  • All fixes as in 2.9.7
  • ADAPTIVE_PATH: the TOLERANCE keyword was not passed to the path reparameterization, and giving it in the input gave an error. Its default value is now 1E-4, which is the value that was used. In addition, the displacement weights (beta column) of all the atoms were set equal to the weight of the first atom.
  • UPPER_WALLS and LOWER_WALLS with periodic vector arguments: as for scalar arguments (see 2.9.7), periodic boundary conditions are no longer applied to the difference between the argument and AT (see issue 658).
  • Forces applied on actions with many atoms (more than about 3300) could trigger multithreading in an external BLAS library (e.g. OpenBLAS or MKL), which ignored the number of threads set with PLUMED_NUM_THREADS or by the MD engine and could oversubscribe all the cores of the node (see issue 1121).
  • plumed mklib and LOAD failed to compile a source file given with a path containing a directory (e.g. plumed mklib ../code/MyCV.cpp or LOAD FILE=../code/MyCV.cpp): the object file was written to a directory that does not exist inside the temporary build directory.