reverting removal of DTypeLike changes

DTypeLike was removed because pyright was choking on it against numpy 2.2.1 Now numpy at 2.2.6 seems to have changes needed for pyright to not have issues with it. additionally, DTypeLike is accurate in a way that np.dtype[Any] was not.

Author: robbiemu

gguf-py/gguf/gguf_reader.py

@@ -181,7 +181,7 @@ def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] =
         self.data_offset = offs
         self._build_tensors(offs, tensors_fields)
 
-    _DT = TypeVar('_DT', bound = np.dtype[Any])
+    _DT = TypeVar('_DT', bound = npt.DTypeLike)
 
     # Fetch a key/value metadata field by key.
     def get_field(self, key: str) -> Union[ReaderField, None]:
@@ -192,7 +192,7 @@ def get_tensor(self, idx: int) -> ReaderTensor:
         return self.tensors[idx]
 
     def _get(
-        self, offset: int, dtype: np.dtype[Any], count: int = 1, override_order: None | Literal['I', 'S', '<'] = None,
+        self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None,
     ) -> np.ndarray:
         count = int(count)
         itemsize = int(np.empty([], dtype = dtype).itemsize)
@@ -328,7 +328,7 @@ def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None:
             block_size, type_size = GGML_QUANT_SIZES[ggml_type]
             n_bytes = n_elems * type_size // block_size
             data_offs = int(start_offs + offset_tensor[0])
-            item_type: np.dtype[Any]
+            item_type: npt.DTypeLike
             if ggml_type == GGMLQuantizationType.F16:
                 item_count = n_elems
                 item_type = np.dtype(np.float16)

gguf-py/gguf/lazy.py

@@ -5,6 +5,7 @@
 from typing import Any, Callable
 
 import numpy as np
+from numpy.typing import DTypeLike
 
 
 logger = logging.getLogger(__name__)
@@ -106,7 +107,7 @@ def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
             return o
 
     @classmethod
-    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | np.dtype[Any] | tuple[np.dtype[Any], Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]:
+    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike | tuple[DTypeLike, Callable[[tuple[int, ...]], tuple[int, ...]]] = False) -> Callable[[Any], Any]:
         def wrapped_fn(*args, **kwargs):
             if kwargs is None:
                 kwargs = {}
@@ -203,7 +204,7 @@ class LazyNumpyTensor(LazyBase):
     shape: tuple[int, ...]  # Makes the type checker happy in quants.py
 
     @classmethod
-    def meta_with_dtype_and_shape(cls, dtype: np.dtype[Any], shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
+    def meta_with_dtype_and_shape(cls, dtype: DTypeLike, shape: tuple[int, ...]) -> np.ndarray[Any, Any]:
         # The initial idea was to use np.nan as the fill value,
         # but non-float types like np.int16 can't use that.
         # So zero it is.

gguf-py/gguf/quants.py

@@ -3,6 +3,8 @@
 from typing import Any, Callable, Sequence
 from math import log2, ceil
 
+from numpy.typing import DTypeLike
+
 from .constants import GGML_QUANT_SIZES, GGMLQuantizationType, QK_K
 from .lazy import LazyNumpyTensor
 
@@ -24,7 +26,7 @@ def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizati
 
 
 # This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
-def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: np.dtype[Any], oshape: tuple[int, ...]) -> np.ndarray:
+def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
     rows = arr.reshape((-1, arr.shape[-1]))
     osize = 1
     for dim in oshape: