Current Behaviour

When trying to fit toys, I need to:

• Resample the data
• Recreate the likelihood with updated constraints

When I do that, the memory seems not to be released. Specifically, the likelihood recreation seems to be causing a memory leak:

import zfit

def fun():
    obs   = zfit.Space('x', limits=(0, 10))
    mu    = zfit.Parameter('mu', 5, 0, 10)
    sg    = zfit.Parameter('sg', 1, 0,  3)
    gauss = zfit.pdf.Gauss(obs=obs, mu=mu, sigma=sg)
    
    sampler   = gauss.create_sampler(n=1000)
    nll       = zfit.loss.UnbinnedNLL(model=gauss, data=sampler)
    minimizer = zfit.minimize.Minuit()
    
    for run_number in tqdm.trange(100, ascii=' -'):
        sampler.resample()
        nll_new = nll.create_new()
        minimizer.minimize(nll_new)

fun()

Expected Behaviour

nll_new stops existing after each iteration, that should trigger the garbage collector but that does not seem to happen.

Context (Environment)

• zfit version: 0.27.1
• Python version: 3.12.11
• Are you using conda, pipenv, etc? : micromamba
• Operating System: almalinux9
• Tensorflow version: 2.19.1

Possible Solution/Implementation

I might be able to re-use the likelihood by parametrizing the constraints, I am still checking. I tried to run the fit in a new subprocess, but it seems that the likelihood is not pickable.