Segfault in IndexHNSW.add() with Python 3.14 on macOS ARM (OpenMP threading)

[athoune]

Description

Segfault (SIGSEGV) in faiss.IndexHNSWFlat.add() when running with multiple OpenMP threads on Python 3.14.5 / macOS ARM (Apple Silicon). The crash disappears when setting OMP_NUM_THREADS=1.

Environment

• faiss-cpu: 1.14.2 (PyPI)
• Python: 3.14.5 (uv-managed, macOS aarch64)
• OS: macOS (Apple Silicon)
• NumPy: 2.3.1

Reproduction

import faiss
import numpy as np

dim = 128
n = 5000
M = 32
ef_construction = 400

rng = np.random.default_rng(42)
vectors = rng.standard_normal((n, dim)).astype(np.float32)

index = faiss.IndexHNSWFlat(dim, M, faiss.METRIC_L2)
index.hnsw.efConstruction = ef_construction
index.add(vectors)  # <-- SIGSEGV here

Works: OMP_NUM_THREADS=1 python repro.py
Crashes: python repro.py (uses all available cores)

Backtrace

Current thread 0x00000001f6a49e80 (most recent call first):
  File "faiss/swigfaiss.py", line 8152 in add
  File "faiss/class_wrappers.py", line 292 in replacement_add

  # C++ stack:
  libfaiss.dylib: faiss::search_neighbors_to_add
  libfaiss.dylib: faiss::HNSW::add_links_starting_from
  libfaiss.dylib: faiss::HNSW::add_with_locks
  libfaiss.dylib: faiss::IndexHNSW::add
  libomp.dylib: __kmpc_fork_call  (OpenMP parallel region)

The crash occurs in the OpenMP parallel region inside IndexHNSW::add(), specifically during neighbor search and link creation in the HNSW graph construction.

Additional context

• The issue is not specific to METRIC_L2 — it also crashes with METRIC_INNER_PRODUCT (COSINE metric via normalized vectors).
• With n=200 it seems to work; with n=5000 it crashes reliably. This suggests a race condition that manifests under higher thread contention.
• Setting OMP_NUM_THREADS=1 is an effective workaround.
• Tested with faiss-cpu 1.14.2 (latest on PyPI as of June 2026).

Request

Could this be a known issue with OpenMP thread safety in the HNSW add path? Is there a recommended configuration for Python 3.14 on macOS ARM?