BUG: `config.change_flags` is not thread-safe: concurrent graph construction can permanently clear `config.cxx`

Ouverte
#2,575 2 commentaires 0 réactions 0 personnes assignées Voir sur GitHub

Les mainteneurs répondent en général sous 1 jour

Personne n'a encore pris cette issue.

Évaluation

Difficulté
4/5
Temps estimé
3-5 jours
Accessibilité débutants
52/100
Type d'issue
Bug
Clarté
Clairement spécifiée
Activité
Active
Stack technique
python
Domaine
backend

Piste de recherche

Start by reading _ChangeFlagsDecorator in pytensor/configparser.py, then inspect the config.change_flags uses in pytensor/tensor/basic.py and the concurrent reproduction in the issue. Identify how overlapping or recursive scopes affect restoration, and add regression coverage for both cases. Done means concurrent graph construction leaves config.cxx and optdb__max_use_ratio at their original values.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Description

bug
Describe the issue:

config.change_flags saves the current global values on entry and writes them back on exit (pytensor/configparser.py, _ChangeFlagsDecorator). When two threads' scopes over the same flag interleave, the scope that exits last restores the temporary value that the other scope had set:

A enters: saves "g++", sets cxx=""
B enters: saves "",    sets cxx=""
A exits:  restores "g++"
B exits:  restores ""            <- process-wide, permanently

Every later scope then saves and restores the corrupted value.

Graph construction enters such a scope. Static-shape inference (pytensor/tensor/basic.py, with config.change_flags(optdb__max_use_ratio=10, cxx="")) runs whenever a shape is symbolic, for example in pt.zeros((n,)). PyMC's resolve_shapes (pymc/pytensorf.py) does the same. Building graphs on several threads at once therefore leaves cxx == "" and optdb__max_use_ratio == 10 for the rest of the process.

A shared decorator instance has the same problem without threads. For example, @config.change_flags(warn_float64="ignore") in pytensor/tensor/basic.py is a single object, and a recursive or concurrent call overwrites the values it saved.

Reproducible code example:
from concurrent.futures import ThreadPoolExecutor

import pytensor
import pytensor.tensor as pt

before = pytensor.config.cxx, pytensor.config.optdb__max_use_ratio


def build(_):
    n = pt.lscalar("n")
    for _ in range(500):
        pt.zeros((n + 1,))  # static-shape inference enters change_flags(cxx="")


with ThreadPoolExecutor(8) as executor:
    list(executor.map(build, range(8)))

print(before, (pytensor.config.cxx, pytensor.config.optdb__max_use_ratio))
# ('/usr/bin/g++', 8.0) ('', 10.0)

xs = pt.dvectors(*(f"x{i}" for i in range(33)))
f = pytensor.function(xs, pt.exp(sum(xs)), mode="CVM")
f(*[[0.0]] * 33)
Error message:
NotImplementedError: Elemwise.perform (Python mode) does not support more than 32 operands. Use the C backend or disable Elemwise fusion.
PyTensor version information:

PyTensor 3.3.1 (pip), Python 3.14, Linux x86_64, with g++ as cxx. The same code is on main (332ea8fb7).

Context for the issue:
  • Every C-linked function compiled afterwards in the process silently falls back to Python implementations. Already compiled functions are unaffected.
  • Fused elemwise graphs with more than 32 operands then raise the error above.
  • Python and C implementations differ at rounding level, so seeded results can change without an error, depending on thread timing earlier in the process. I measured differences of about 1e-14 in a downstream PyMC model.

I hit this in a library that fits independent models on worker threads with PyMC. A long-running process lost C compilation after one threaded fit, and an unrelated later computation failed with the error above.

Langage dominant
Python
Étoiles
654
Forks
258
Merge moyen
3 j 41 min
PR mergées (30 j)
42

Préparer son environnement

Par où commencer

  1. Lisez l'issue en entier, puis le guide de contribution du projet.
  2. Signalez en commentaire que vous la prenez — cela évite que deux personnes fassent le même travail.
  3. Forkez le dépôt et travaillez sur une branche.
  4. Ouvrez une pull request qui référence le numéro de l'issue.

Autres issues de pymc-devs/pytensor

Toutes les issues de pymc-devs/pytensor

Issues similaires

Plus d'issues Python

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.