fix(core): lazy loading transformers, numpy, simsimd

libs/core/langchain_core/language_models/base.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import importlib.util
 import warnings
 from abc import ABC, abstractmethod
 from collections.abc import Callable, Mapping, Sequence
@@ -38,12 +39,7 @@
 if TYPE_CHECKING:
     from langchain_core.outputs import LLMResult
 
-try:
-    from transformers import GPT2TokenizerFast  # type: ignore[import-not-found]
-
-    _HAS_TRANSFORMERS = True
-except ImportError:
-    _HAS_TRANSFORMERS = False
+_HAS_TRANSFORMERS = importlib.util.find_spec("transformers") is not None
 
 
 class LangSmithParams(TypedDict, total=False):
@@ -93,6 +89,11 @@ def get_tokenizer() -> Any:
             "Please install it with `pip install transformers`."
         )
         raise ImportError(msg)
+
+    from transformers import (  # type: ignore[import-not-found]  # noqa: PLC0415
+        GPT2TokenizerFast,
+    )
+
     # create a GPT-2 tokenizer instance
     return GPT2TokenizerFast.from_pretrained("gpt2")
 

libs/core/langchain_core/vectorstores/in_memory.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import importlib.util
 import json
 import uuid
 from pathlib import Path
@@ -23,12 +24,7 @@
 
     from langchain_core.embeddings import Embeddings
 
-try:
-    import numpy as np
-
-    _HAS_NUMPY = True
-except ImportError:
-    _HAS_NUMPY = False
+_HAS_NUMPY = importlib.util.find_spec("numpy") is not None
 
 
 class InMemoryVectorStore(VectorStore):
@@ -439,6 +435,8 @@ def max_marginal_relevance_search_by_vector(
             )
             raise ImportError(msg)
 
+        import numpy as np  # noqa: PLC0415
+
         mmr_chosen_indices = maximal_marginal_relevance(
             np.array(embedding, dtype=np.float32),
             [vector for _, _, vector in prefetch_hits],

libs/core/langchain_core/vectorstores/utils.py

@@ -8,27 +8,21 @@
 
 from __future__ import annotations
 
+import importlib.util
 import logging
 import warnings
 from typing import TYPE_CHECKING, cast
 
-try:
-    import numpy as np
-
-    _HAS_NUMPY = True
-except ImportError:
-    _HAS_NUMPY = False
-
-try:
-    import simsimd as simd  # type: ignore[import-not-found]
+_HAS_NUMPY = importlib.util.find_spec("numpy") is not None
+_HAS_SIMSIMD = importlib.util.find_spec("simsimd") is not None
 
-    _HAS_SIMSIMD = True
-except ImportError:
-    _HAS_SIMSIMD = False
 
 if TYPE_CHECKING:
+    import numpy as np
+
     Matrix = list[list[float]] | list[np.ndarray] | np.ndarray
 
+
 logger = logging.getLogger(__name__)
 
 
@@ -54,6 +48,8 @@ def _cosine_similarity(x: Matrix, y: Matrix) -> np.ndarray:
         )
         raise ImportError(msg)
 
+    import numpy as np  # noqa: PLC0415
+
     if len(x) == 0 or len(y) == 0:
         return np.array([[]])
 
@@ -98,6 +94,8 @@ def _cosine_similarity(x: Matrix, y: Matrix) -> np.ndarray:
         similarity[np.isnan(similarity) | np.isinf(similarity)] = 0.0
         return cast("np.ndarray", similarity)
 
+    import simsimd as simd  # type: ignore[import-not-found]  # noqa: PLC0415
+
     x = np.array(x, dtype=np.float32)
     y = np.array(y, dtype=np.float32)
     return 1 - np.array(simd.cdist(x, y, metric="cosine"))
@@ -130,6 +128,8 @@ def maximal_marginal_relevance(
         )
         raise ImportError(msg)
 
+    import numpy as np  # noqa: PLC0415
+
     if min(k, len(embedding_list)) <= 0:
         return []
     if query_embedding.ndim == 1: