Using the Python package#

After installing GROMACS, sourcing the “GMXRC” (see GROMACS docs), and installing the gmxapi Python package (see Full installation instructions), import the package in a Python script or interactive interpreter. This documentation assumes a convenient alias of gmx to refer to the gmxapi Python package.

import gmxapi as gmx

For full documentation of the Python-level interface and API, use the pydoc command line tool or the help() interactive Python function, or refer to the gmxapi Python module reference.

Any Python exception raised by gmxapi should be descended from (and catchable as) gmxapi.exceptions.Error. Additional status messages can be acquired through the Logging facility. Unfortunately, some errors occurring in the GROMACS library are not yet recoverable at the Python level, and much of the standard GROMACS terminal output is not yet accessible through Python. If you find a particularly problematic scenario, please file a GROMACS bug report.

During installation, the gmxapi Python package becomes tied to a specific GROMACS installation. If you would like to access multiple GROMACS installations from Python, build and install gmxapi in separate virtual environments.

Notes on parallelism and MPI#

The GROMACS library can be built for parallel computation using various strategies. If GROMACS was configured with -DGMX_MPI=ON, the same MPI library and compiler tool chain must be used for gmxapi and mpi4py. In any case, the Python package must be built with mpi4py installed. See MPI requirements.

Note

This section uses “mpiexec” generically to refer to the MPI program launcher. Depending on your MPI implementation and system details, your environment may use “mpirun”, or some other command instead.

gmxapi scripts manage batches of simulations (as “ensembles”) using MPI and mpi4py. To check whether your installed gmxapi package was built with MPI bindings, you can check for the mpi_bindings feature using gmxapi.version.has_feature(). The following command will produce an error if the feature is not available.

python -c 'import gmxapi; assert gmxapi.version.has_feature("mpi_bindings")'

Assuming you use mpiexec to launch MPI jobs in your environment, run a gmxapi script on two ranks with something like the following. Note that it can be helpful to provide mpiexec with the full path to the intended Python interpreter since new process environments are being created.

mpiexec -n 2 `which python` -m mpi4py myscript.py

The -m mpi4py ensures that the mpi4py package is available and allows for proper clean-up of resources. (See mpi4py.run for details.)

Mapping ranks to ensemble members#

gmxapi divides the root communicator into separate sub-communicators for each simulator in an ensemble simulation task. Consider a root communicator of size S being allocated to N simulators. Each rank R in the root communicator is assigned to ensemble member M(R) as follows.

When GROMACS is built with MPI library support, gmxapi allocates available MPI ranks to simulators in (approximately) equal size consecutive chunks.

\[M(R) = \text{trunc}(R * N / S)\]

For thread-MPI (or no-MPI) GROMACS builds, each simulator is assigned one process (with an attempt at even distribution). Based on the preceding formula, thread-MPI ensemble member assignment looks like the following.

\[\begin{split}M_T(R) = \begin{cases} M(R) &,\; M(R) \neq M(R-1) \\ \textrm{null} &,\; \textrm{otherwise} \end{cases}\end{split}\]

In other words, without an MPI library, only the root rank from M(R) is assigned.

Changed in version 0.4.0: In earlier releases, ranks were assigned to thread-MPI simulators contiguously, such that high-numbered ranks R>N were unused. MPI simulators were not supported for ensemble simulation tasks.

Caveats for MPI jobs#

Changed in version 0.3.0: By default, most commands outside gmxapi.simulation launch only on the root rank. (Results are synchronized to all ranks.) gmxapi.function_wrapper allows you to set allow_duplicate=True, if your script logic or data transfer overhead require tasks to be executed on all ranks (computation is duplicated).

If gmxapi.commandline_operation is used to wrap an MPI-enabled executable, the executable could behave unpredictably when the script is run in an MPI context. By default, commandline_operation subprocesses get a copy of the environment from the Python interpreter from which they are launched, and an executable may think it was launched directly by mpiexec, causing MPI errors when it tries to assert ownership of the MPI resources.

Changed in version 0.3.1: You can use the env key word argument to gmxapi.commandline_operation to replace the default map of environment variables. By pruning out the environment variables set by the MPI launcher, you can prevent the executable from automatically detecting an MPI context that it shouldn’t use. See also Issue 4421

gmxapi does not currently have an abstraction for subprocess launch methods. While such a feature is under investigation, allow_duplicate (function_wrapper()) and env (commandline_operation()) should allow users to wrap tools in custom launchers. Discussion welcome on the forum!

Running simple simulations#

Once the gmxapi package is installed, running simulations is easy with gmxapi.read_tpr().

import gmxapi as gmx
simulation_input = gmx.read_tpr(tpr_filename)
md = gmx.mdrun(simulation_input)

Note that this sets up the work you want to perform, but does not immediately trigger execution. You can explicitly trigger execution with:

md.run()

or you can let gmxapi automatically launch work in response to the data you request (by calling result() on a named output member).

The gmxapi.mdrun() operation produces a simulation trajectory output. You can use md.output.trajectory as input to other operations, or you can get the output directly by calling md.output.trajectory.result(). If the simulation has not been run yet when result() is called, the simulation will be run before the function returns.

Running ensemble simulations#

To run a batch of simulations, just pass an array of inputs.:

md = gmx.read_tpr([tpr_filename1, tpr_filename2, ...])
md.run()

Make sure to launch the script in an MPI environment with a sufficient number of ranks to allow one rank per simulation.

Input arguments and “ensemble” syntax#

When a list of input is provided to a command argument that expects some other type, gmxapi generates an ensemble operation. The command is applied to each element of input, and the result() will be a list. When an output member of an ensemble operation is provided as input to another command, the consuming command will also be an ensemble operation.

gmxapi uses MPI to manage ensemble members across available resources. It is important that the same gmxapi commands are called on all processes so that underlying collective MPI calls are made as expected. In other words, if you are using mpi4py in your script, be careful with conditional execution like the following.

if mpi4py.MPI.COMM_WORLD.Get_rank() == 0:
    # don't put any gmxapi commands here, including method calls
    # like `obj.result()`, unless you have an `else`
    # to make sure the same gmxapi command runs on every rank.
    ...

For commands that already integrate well with gmxapi’s MPI-based ensemble management (like mdrun()), available resources can be split up automatically, and applied to run the ensemble members concurrently. Other operations may require further development of Resource Management API features for the gmxapi framework to most effectively apply multi-core computing resources. See Issue 3718 and the wiki for more information.

See also Notes on parallelism and MPI.

Accessing command line tools#

In gmxapi 0.1, most GROMACS tools are not yet exposed as gmxapi Python operations. gmxapi.commandline_operation provides a way to convert a gmx (or other) command line tool into an operation that can be used in a gmxapi script.

In order to establish data dependencies, input and output files need to be indicated with the input_files and output_files parameters. input_files and output_files key word arguments are dictionaries consisting of files keyed by command line flags.

For example, you might create a gmx solvate operation as:

solvate = gmx.commandline_operation('gmx',
                                    arguments=['solvate', '-box', '5', '5', '5'],
                                    input_files={'-cs': structurefile},
                                    output_files={'-p': topfile,
                                                  '-o': structurefile,
                                                  }

To check the status or error output of a command line operation, refer to the returncode and stderr outputs. To access the results from the output file arguments, use the command line flags as keys in the file dictionary output.

Example:

structurefile = solvate.output.file['-o'].result()
if solvate.output.returncode.result() != 0:
    print(solvate.output.erroroutput.result())

Preparing simulations#

Continuing the previous example, the output of solvate may be used as the input for grompp:

grompp = gmx.commandline_operation('gmx', 'grompp',
                                   input_files={
                                       '-f': mdpfile,
                                       '-p': solvate.output.file['-p'],
                                       '-c': solvate.output.file['-o'],
                                       '-po': mdout_mdp,
                                   },
                                   output_files={'-o': tprfile})

Then, grompp.output.file['-o'] can be used as the input for gmxapi.read_tpr().

Simulation input can be modified with the gmxapi.modify_input() operation before being passed to gmxapi.mdrun(). For gmxapi 0.1, a subset of MDP parameters may be overridden using the dictionary passed with the parameters key word argument.

Example:

simulation_input = gmx.read_tpr(grompp.output.file['-o'])
modified_input = gmx.modify_input(input=simulation_input, parameters={'nsteps': 1000})
md = gmx.mdrun(input=modified_input)
md.run()

Using arbitrary Python functions#

Generally, a function in the gmxapi package returns an object that references a node in a work graph, representing an operation that will be run when the graph executes. The object has an output attribute providing access to data Futures that can be provided as inputs to other operations before computation has actually been performed.

You can also provide native Python data as input to operations, or you can operate on native results retrieved from a Future’s result() method. However, it is trivial to convert most Python functions into gmxapi compatible operations with gmxapi.function_wrapper(). All function inputs and outputs must have a name and type. Additionally, functions should be stateless and importable (e.g. via Python from some.module import myfunction) for future compatibility.

Simple functions can just use return() to publish their output, as long as they are defined with a return value type annotation. Functions with multiple outputs can accept an output key word argument and assign values to named attributes on the received argument.

Examples:

from gmxapi import function_wrapper

@function_wrapper(output={'data': float})
def add_float(a: float, b: float) -> float:
    return a + b


@function_wrapper(output={'data': bool})
def less_than(lhs: float, rhs: float, output=None):
    output.data = lhs < rhs

See also

For more on Python type hinting with function annotations, check out PEP 3107.

Subgraphs#

Basic gmxapi work consists of a flow of data from operation outputs to operation inputs, forming a directed acyclic graph (DAG). In many cases, it can be useful to repeat execution of a subgraph with updated inputs. You may want a data reference that is not tied to the immutable result of a single node in the work graph, but which instead refers to the most recent result of a repeated operation.

One or more operations can be staged in a gmxapi.operation.Subgraph, a sort of meta-operation factory that can store input binding behavior so that instances can be created without providing input arguments.

The subgraph variables serve as input, output, and mutable internal data references which can be updated by operations in the subgraph. Variables also allow state to be propagated between iterations when a subgraph is used in a while loop.

Use gmxapi.subgraph() to create a new empty subgraph. The variables argument declares data handles that define the state of the subgraph when it is run. To initialize input to the subgraph, give each variable a name and a value.

To populate a subgraph, enter a SubgraphContext by using a with() statement. Operations created in the with block will be captued by the SubgraphContext. Define the subgraph outputs by assigning operation outputs to subgraph variables within the with block.

After exiting the with block, the subgraph may be used to create operation instances or may be executed repeatedly in a while loop.

Note

The object returned by gmxapi.subgraph() is atypical of gmxapi operations, and has some special behaviors. When used as a Python context manager, it enters a “builder” state that changes the behavior of its attribute variables and of operaton instantiation. After exiting the with() block, the subgraph variables are no longer assignable, and operation references obtained within the block are no longer valid.

Looping#

An operation can be executed an arbitrary number of times with a gmxapi.while_loop() by providing a factory function as the operation argument. When the loop operation is run, the operation is instantiated and run repeatedly until condition evaluates True.

gmxapi.while_loop() does not provide a direct way to provide operation arguments. Use a subgraph to define the data flow for iterative operations.

When a condition is a subgraph variable, the variable is evaluated in the running subgraph instance at the beginning of an iteration.

Example:

subgraph = gmx.subgraph(variables={'float_with_default': 1.0, 'bool_data': True})
with subgraph:
    # Define the update for float_with_default to come from an add_float operation.
    subgraph.float_with_default = add_float(subgraph.float_with_default, 1.).output.data
    subgraph.bool_data = less_than(lhs=subgraph.float_with_default, rhs=6.).output.data
operation_instance = subgraph()
operation_instance.run()
assert operation_instance.values['float_with_default'] == 2.

loop = gmx.while_loop(operation=subgraph, condition=subgraph.bool_data)
handle = loop()
assert handle.output.float_with_default.result() == 6

Logging#

gmxapi uses the Python logging module to provide hierarchical logging, organized by submodule. You can access the logger at gmxapi.logger or, after importing gmxapi, through the Python logging framework:

import gmxapi as gmx
import logging

# Get the root gmxapi logger.
gmx_logger = logging.getLogger('gmxapi')
# Set a low default logging level
gmx_logger.setLevel(logging.WARNING)
# Make some tools very verbose
#  by descending the hierarchy
gmx_logger.getChild('commandline').setLevel(logging.DEBUG)
#  or by direct reference
logging.getLogger('gmxapi.mdrun').setLevel(logging.DEBUG)

You may prefer to adjust the log format or manipulate the log handlers. For example, tag the log output with MPI rank:

try:
    from mpi4py import MPI
    rank_number = MPI.COMM_WORLD.Get_rank()
except ImportError:
    rank_number = 0
    rank_tag = ''
    MPI = None
else:
    rank_tag = 'rank{}:'.format(rank_number)

formatter = logging.Formatter(rank_tag + '%(name)s:%(levelname)s: %(message)s')

# For additional console logging, create and attach a stream handler.
ch = logging.StreamHandler()
ch.setFormatter(formatter)
logging.getLogger().addHandler(ch)

For more information, refer to the Python logging documentation.

More#

Refer to the gmxapi Python module reference for complete and granular documentation.

For more information on writing or using pluggable simulation extension code, refer to https://gitlab.com/gromacs/gromacs/-/issues/3133. (For gmxapi 0.0.7 and GROMACS 2019, see https://github.com/kassonlab/sample_restraint)