dev: faiss-cpu>=1.12.0 + remove monkeypatch (#16) [skip ci]

Signed-off-by: Élie Goudout <[email protected]>

Author: eliegoudout

docs/src/user_guide/installation_compatibility.rst

@@ -18,21 +18,18 @@ If you wish to install from source or wheels manually, you can download `release
 
 OS Compatibility
 ----------------
-The library is available and functional on Ubuntu, Windows and MacOS. However, note the following regarding ``faiss``-related features.
-
-- They are **partially** supported on **Windows** and their use will trigger a warning: due to non-Python dependency conflicts, there might be uncontrolled and silent concurrency issues, potentially altering results. For more information, see this nice ressource about `native dependencies <https://pypackaging-native.github.io/key-issues/native-dependencies>`_. Note that we still run full CI tests on Windows.
-- They are **not officially supported** on **MacOS** and might generate random crashes. We should match the current Windows support status *relatively soon*.
+The library is available and functional on **Ubuntu**, **Windows** and *MacOS* (minus ``faiss``-related features ─ feel free to show your interest `here <https://github.com/ThalesGroup/scio/issues/2>`_).
 
 Framework Compatiblity
 ----------------------
 Although confidence scores are not a particularity of Neural Networks, ``scio`` is precisely developed to handle such models. More precisely, only *PyTorch* models (inheriting from ``torch.nn.Module``) are currently supported.
 
 GPU Compatibility
 -----------------
-Our library is **fully compatible** and tested with CUDA devices for native use of GPU acceleration! Note however that features using ``faiss`` currently use CPU-bound indexes, introducing a potential GPU > CPU > GPU bottleneck. This may be improved in future versions. Do not hesitate to express interest by opening a related issue.
+Our library is **fully compatible** and tested with CUDA devices for native use of GPU acceleration! Note however that features using ``faiss`` currently use CPU-bound indexes, introducing a potential GPU > CPU > GPU bottleneck. This may be improved in future versions ─ feel free to show your interest `here <https://github.com/ThalesGroup/scio/issues/18>`_.
 
 Supported Data Types
 --------------------
 The package is **fully compatible** and tested with ``torch.half`` (float16), ``torch.float`` (float32) and ``torch.double`` (float64) data types. However, we **discourage** the use of gradient-based algorithms with ``torch.half`` data, as they can easily generate irrelevant results due to potential ``nan`` values. Finally, note that ``faiss``-related operations temporarily convert to 32 bits data, regardless of the input type.
 
-There is currently no official support for ``torch.bfloat16`` even though many features may be compatible. Do not hesitate to express interest by opening a related issue.
+There is currently no official support for ``torch.bfloat16`` even though many features may be compatible. Feel free to show your interest by opening a related issue.

pyproject.toml

@@ -29,7 +29,7 @@ classifiers = [
 ]
 dependencies = [
     "dill>=0.3.8",
-    "faiss-cpu>=1.11",
+    "faiss-cpu>=1.12",
     "lazy-loader>=0.4",
     "matplotlib>=3.10",
     "numpy>=2.2.2",

scio/__init__.py

@@ -1,7 +1,5 @@
 """scio package."""
 
-import os
-import platform
 from importlib.metadata import PackageNotFoundError, version
 
 try:
@@ -13,48 +11,3 @@
 
 # Lazily load from adjacent `.pyi` file
 __getattr__, __dir__, __all__ = lazy.attach_stub(__name__, __file__)
-
-# Monkeypatch the use of ``faiss`` on non-Linux platforms
-if (
-    platform.system() != "Linux"
-    and os.environ.get("KMP_DUPLICATE_LIB_OK", None) != "TRUE"
-):  # pragma: no cover
-    import importlib
-    import sys
-    import types
-    from warnings import warn
-
-    null = object()
-
-    class DummyFaiss(types.ModuleType):
-        """Dummy faiss to intercept faiss loading."""
-
-        __faiss = null
-
-        def __getattribute__(self, attr: str) -> object:
-            """Intercept loading, warn & enable KMP_DUPLICATE_LIB_OK."""
-            __faiss = super().__getattribute__(f"_{type(self).__name__}__faiss")
-
-            # If necessary, warn and load faiss
-            if __faiss is null:
-                msg = (
-                    "On non-Linux platforms, the use of any `faiss`-dependant feature "
-                    "is partially supported by `scio`: you might experience slowdowns "
-                    "or even incorrect results! This is due to a non-Python dependency "
-                    "conflict regarding OpenMP (see https://pypackaging-native.github"
-                    ".io/key-issues/native-dependencies for interesting ressources). "
-                    "As a monkeypatch, we enable the `KMP_DUPLICATE_LIB_OK`."
-                )
-                warn(msg, stacklevel=2)
-                os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
-
-                if sys.modules.pop("faiss") is not self:
-                    msg = "Dummy faiss module should only be defined during scio init"
-                    raise RuntimeError(msg)
-
-                self.__faiss = __faiss = importlib.import_module("faiss")
-                sys.modules["faiss"] = __faiss
-
-            return __faiss.__getattribute__(attr)
-
-    sys.modules["faiss"] = DummyFaiss("faiss")