The same benchmarks CodSpeed runs, measured locally¶

linopy benchmarks its model construction with pytest-benchmark and runs that suite on CodSpeed — which, since its 2026 memory instrument, tracks memory as well as time. pytest-benchmem reuses those exact tests: the benchmark(...) call is all it reads. No second suite, no new markers — point it at the benchmarks you already have and get a memray memory profile locally, on demand, without a CI round-trip or an account. It was extracted from linopy's suite to be exactly that: the terminal companion for the numbers CodSpeed keeps in CI.

Here is the suite's model — linopy's "basic problem", an N×N LP — construction plus a couple of LinearExpression operations, sized by n:

In [1]:

from numpy import arange

import linopy


def create_model(n: int) -> linopy.Model:
    m = linopy.Model()
    N, M = arange(n), arange(n)
    x = m.add_variables(coords=[N, M], name="x")
    y = m.add_variables(coords=[N, M], name="y")
    m.add_constraints(x - y >= N)
    m.add_constraints(x + y >= 0)
    m.add_objective(2 * x.sum() + y.sum())
    return m

from numpy import arange import linopy def create_model(n: int) -> linopy.Model: m = linopy.Model() N, M = arange(n), arange(n) x = m.add_variables(coords=[N, M], name="x") y = m.add_variables(coords=[N, M], name="y") m.add_constraints(x - y >= N) m.add_constraints(x + y >= 0) m.add_objective(2 * x.sum() + y.sum()) return m

The same test, in your terminal¶

An ordinary pytest-benchmark test — the one already in the suite — with --benchmark-memory adding the peak on the same call it already times:

@pytest.mark.parametrize("n", [100, 200, 400])
def test_create_model(benchmark, n):
    benchmark(create_model, n)

benchmark: 3 tests                                                                      
                                                                                        
  Name (time in ms)                   Mean   │   peak·min (MiB)   peak·mean   peak·max  
 ────────────────────────────────────────────────────────────────────────────────────── 
  test_create_model[100]     83.0831 (1.0)   │             2.16        2.16       2.16  
  test_create_model[200]    87.1668 (1.05)   │             8.64        8.64       8.64  
  test_create_model[400]   100.9533 (1.22)   │            34.52       34.52      34.52  
                                                                                        
memory (right of │): a separate, untimed pass, not the timed rounds  •  also available  
via --benchmark-memory-columns: allocated, allocations                                  
3 passed in 8.72s

Peak next to the time, straight from the call. The peak climbs with the square of n (~2 → ~35 MiB here) while the timer, mostly fixed overhead at this size, barely moves — the kind of thing you want to catch before it becomes a model that won't build.

The local half of the loop¶

CodSpeed keeps the history, the dashboards, the PR annotations. pytest-benchmem is the offline companion on the same tests, for the tight loop where a CI round-trip is too slow — run it, then diff two runs, flamegraph where it allocated, chart the scaling, or sweep across installed linopy versions, all in the terminal:

benchmem compare base.json head.json --columns peak          # what moved
benchmem flamegraph profiles/ --worst peak --open            # where it allocated
benchmem sweep linopy 0.7.0 0.8.0 --suite bench/             # across versions

One suite, measured both ways. The Quickstart starts from benchmarks you already have.