Profiling your models

Do you feel like your model is too slow? Do you want to make it faster? Instead of guessing which part of the code is responsible for any slowdown, you should profile your code to learn how much time is spent in each function and where to focus any optimization efforts.

In this tutorial you’ll learn how to profile your model using PyTorch profiler, how to read the output of the profiler, and how to add your own labels for new functions/steps in your model forward function.

from typing import Dict, List, Optional

import matplotlib.pyplot as plt
import numpy as np
import torch

from metatensor.torch import Labels, TensorBlock, TensorMap
from metatensor.torch.atomistic import (
from metatensor.torch.atomistic.ase_calculator import MetatensorCalculator

When profiling your code, it is important to run the model on a representative system to ensure you are actually exercising the behavior of your model at the right scale. Here we’ll use a relatively large system with many atoms.

primitive ="C", crystalstructure="diamond", a=3.567)
atoms =, 10 * np.eye(3))
print(f"We have {len(atoms)} atoms in our system")
We have 2000 atoms in our system

We will use the same HarmonicModel as in the previous tutorial as our machine learning potential.

Click to see the definition of HarmonicModel
class HarmonicModel(torch.nn.Module):
    def __init__(self, force_constant: float, equilibrium_positions: torch.Tensor):
        """Create an ``HarmonicModel``.

        :param force_constant: force constant, in ``energy unit / (length unit)^2``
        :param equilibrium_positions: torch tensor with shape ``n x 3``, containing the
            equilibrium positions of all atoms
        assert force_constant > 0
        self.force_constant = force_constant
        self.equilibrium_positions = equilibrium_positions

    def forward(
        systems: List[System],
        outputs: Dict[str, ModelOutput],
        selected_atoms: Optional[Labels],
    ) -> Dict[str, TensorMap]:
        # if the model user did not request an energy calculation, we have nothing to do
        if "energy" not in outputs:
            return {}

        # we don't want to worry about selected_atoms yet
        if selected_atoms is not None:
            raise NotImplementedError("selected_atoms is not implemented")

        if outputs["energy"].per_atom:
            raise NotImplementedError("per atom energy is not implemented")

        # compute the energy for each system by adding together the energy for each atom
        energy = torch.zeros((len(systems), 1), dtype=systems[0].positions.dtype)
        for i, system in enumerate(systems):
            assert len(system) == self.equilibrium_positions.shape[0]
            r0 = self.equilibrium_positions
            energy[i] += torch.sum(self.force_constant * (system.positions - r0) ** 2)

        # add metadata to the output
        block = TensorBlock(
            samples=Labels("system", torch.arange(len(systems)).reshape(-1, 1)),
            properties=Labels("energy", torch.tensor([[0]])),
        return {
            "energy": TensorMap(keys=Labels("_", torch.tensor([[0]])), blocks=[block])

model = HarmonicModel(

capabilities = ModelCapabilities(
        "energy": ModelOutput(quantity="energy", unit="eV", per_atom=False),

metadata = ModelMetadata()
wrapper = MetatensorAtomisticModel(model.eval(), metadata, capabilities)

/home/runner/work/metatensor/metatensor/python/examples/atomistic/ DeprecationWarning: `export()` is deprecated, use `save()` instead

If you are trying to profile your own model, you can start here and create a MetatensorCalculator with your own model.

atoms.calc = MetatensorCalculator("")

Before trying to profile the code, it is a good idea to run it a couple of times to allow torch to warmup internally.


Profiling energy calculation

Now we can run code using torch.profiler.profile() to collect statistic on how long each function takes to run. We randomize the positions to force ASE to recompute the energy of the system

atoms.positions += np.random.rand(*atoms.positions.shape)
with torch.profiler.profile() as energy_profiler:

print(energy_profiler.key_averages().table(sort_by="self_cpu_time_total", row_limit=10))
--------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
                                              Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls
--------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
                                    Model::forward        44.08%     473.000us        58.06%     623.000us     623.000us             1
                     ASECalculator::prepare_inputs        21.90%     235.000us        25.72%     276.000us     276.000us             1
                    ASECalculator::convert_outputs         8.57%      92.000us         9.79%     105.000us      52.500us             2
                  ASECalculator::compute_neighbors         2.70%      29.000us         2.70%      29.000us      29.000us             1
                                    aten::_to_copy         2.61%      28.000us         5.31%      57.000us       6.333us             9
                                         aten::sum         2.14%      23.000us         2.14%      23.000us      23.000us             1
                                       aten::copy_         1.96%      21.000us         1.96%      21.000us       2.333us             9
                                      aten::arange         1.77%      19.000us         3.26%      35.000us      17.500us             2
                                         aten::sub         1.40%      15.000us         2.42%      26.000us      26.000us             1
                                         aten::mul         1.21%      13.000us         1.21%      13.000us      13.000us             1
--------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
Self CPU time total: 1.073ms

There are a couple of interesting things to see here. First the total runtime of the code is shown in the bottom; and then the most costly functions are visible on top, one line per function. For each function, Self CPU refers to the time spent in this function excluding any called functions; and CPU total refers to the time spent in this function, including called functions.

For more options to record operations and display the output, please refer to the official documentation for PyTorch profiler.

Here, Model::forward indicates the time taken by your model’s forward(). Anything starting with aten:: comes from operations on torch tensors, typically with the same function name as the corresponding torch functions (e.g. aten::arange is torch.arange()). We can also see some internal functions from metatensor, with the name staring with MetatensorAtomisticModel:: for MetatensorAtomisticModel; and ASECalculator:: for ase_calculator.MetatensorCalculator.

If you want to see more details on the internal steps taken by your model, you can add torch.profiler.record_function() ( inside your model code to give names to different steps in the calculation. This is how we are internally adding names such as Model::forward or ASECalculator::prepare_inputs above.

Profiling forces calculation

Let’s now do the same, but computing the forces for this system. This mean we should now see some time spent in the backward() function, on top of everything else.

atoms.positions += np.random.rand(*atoms.positions.shape)
with torch.profiler.profile() as forces_profiler:

print(forces_profiler.key_averages().table(sort_by="self_cpu_time_total", row_limit=10))
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
                                                   Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg    # of Calls
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
torch::jit::(anonymous namespace)::DifferentiableGra...        35.87%     816.000us        42.99%     978.000us     978.000us             1
                                         Model::forward        20.35%     463.000us        28.57%     650.000us     650.000us             1
                          ASECalculator::prepare_inputs         8.70%     198.000us        10.24%     233.000us     233.000us             1
                         ASECalculator::convert_outputs         5.23%     119.000us         6.24%     142.000us      71.000us             2
                                              aten::mul         5.14%     117.000us         5.45%     124.000us      31.000us             4
                            ASECalculator::run_backward         4.97%     113.000us        51.87%       1.180ms       1.180ms             1
                                            aten::copy_         2.11%      48.000us         2.11%      48.000us       3.200us            15
                                              aten::sub         1.80%      41.000us         2.55%      58.000us      58.000us             1
                                          <backward op>         1.80%      41.000us         7.12%     162.000us     162.000us             1
                                         aten::_to_copy         1.54%      35.000us         3.21%      73.000us       6.083us            12
-------------------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------
Self CPU time total: 2.275ms

Let’s visualize this data in an other way:

events = forces_profiler.key_averages()
events = sorted(events, key=lambda u: u.self_cpu_time_total, reverse=True)
total_cpu_time = sum(map(lambda u: u.self_cpu_time_total, events))

bottom = 0.0
for event in events:
    self_time = event.self_cpu_time_total
    name = event.key
    if len(name) > 30:
        name = name[:12] + "[...]" + name[-12:]

    if self_time > 0.03 * total_cpu_time:, self_time, bottom=bottom, label=name)
        bottom += self_time
    else:, total_cpu_time - bottom, bottom=bottom, label="others")

plt.xlim(0, 1)
plt.ylabel("self time / µs")
3 profiling

Total running time of the script: (0 minutes 0.278 seconds)

Gallery generated by Sphinx-Gallery