8000 reverting unnecessary changes in gguf-py/gguf as of numpy 2.2.6 · robbiemu/llama.cpp@4802d01 · GitHub
[go: up one dir, main page]
Skip to content

Commit 4802d01

Browse files
robbiemurobbiemu
committed
reverting unnecessary changes in gguf-py/gguf as of numpy 2.2.6
1 parent 3400755 commit 4802d01

File tree

2 files changed

+29
-29
lines changed

2 files changed

+29
-29
lines changed

‎gguf-py/gguf/gguf_reader.py

Lines changed: 25 additions & 25 deletions
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,4 @@
1+
#
12
# GGUF file reading/modification support. For API usage information,
23
# please see the files scripts/ for some fairly simple examples.
34
#
@@ -14,7 +15,6 @@
1415

1516
from .quants import quant_shape_to_byte_shape
1617

17-
1818
if __name__ == "__main__":
1919
from pathlib import Path
2020

@@ -134,12 +134,12 @@ def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] =
134134
offs = 0
135135

136136
# Check for GGUF magic
137-
if self._get(offs, np.dtype(np.uint32), override_order = '<')[0] != GGUF_MAGIC:
137+
if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC:
138138
raise ValueError('GGUF magic invalid')
139139
offs += 4
140140

141141
# Check GGUF version
142-
temp_version = self._get(offs, np.dtype(np.uint32))
142+
temp_version = self._get(offs, np.uint32)
143143
if temp_version[0] & 65535 == 0:
144144
# If we get 0 here that means it's (probably) a GGUF file created for
145145
# the opposite byte order of the machine this script is running on.
@@ -162,7 +162,7 @@ def __init__(self, path: os.PathLike[str] | str, mode: Literal['r', 'r+', 'c'] =
162162
offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32]))
163163

164164
# Check tensor count and kv count
165-
temp_counts = self._get(offs, np.dtype(np.uint64), 2)
165+
temp_counts = self._get(offs, np.uint64, 2)
166166
offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
167167
offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
168168
tensor_count, kv_count = temp_counts
@@ -193,7 +193,7 @@ def get_tensor(self, idx: int) -> ReaderTensor:
193193

194194
def _get(
195195
self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I', 'S', '<'] = None,
196-
) -> np.ndarray:
196+
) -> npt.NDArray[Any]:
197197
count = int(count)
198198
itemsize = int(np.empty([], dtype = dtype).itemsize)
199199
end_offs = offset + itemsize * count
@@ -212,8 +212,8 @@ def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
212212
return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts)
213213

214214
def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]:
215-
slen = self._get(offset, np.dtype(np.uint64))
216-
return slen, self._get(offset + 8, np.dtype(np.uint8), slen[0].item())
215+
slen = self._get(offset, np.uint64)
216+
return slen, self._get(offset + 8, np.uint8, slen[0])
217217

218218
def _get_field_parts(
219219
self, orig_offs: int, raw_type: int,
@@ -230,19 +230,19 @@ def _get_field_parts(
230230
# Check if it's a simple scalar type.
231231
nptype = self.gguf_scalar_to_np.get(gtype)
232232
if nptype is not None:
233-
val = self._get(offs, np.dtype(nptype))
233+
val = self._get(offs, nptype)
234234
return int(val.nbytes), [val], [0], types
235235
# Handle arrays.
236236
if gtype == GGUFValueType.ARRAY:
237-
raw_itype = self._get(offs, np.dtype(np.uint32))
237+
raw_itype = self._get(offs, np.uint32)
238238
offs += int(raw_itype.nbytes)
239-
alen = self._get(offs, np.dtype(np.uint64))
239+
alen = self._get(offs, np.uint64)
240240
offs += int(alen.nbytes)
241241
aparts: list[npt.NDArray[Any]] = [raw_itype, alen]
242242
data_idxs: list[int] = []
243243
# FIXME: Handle multi-dimensional arrays properly instead of flattening
244244
for idx in range(alen[0]):
245-
curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0].item())
245+
curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0])
246246
if idx == 0:
247247
types += curr_types
248248
idxs_offs = len(aparts)
@@ -261,19 +261,19 @@ def _get_tensor_info_field(self, orig_offs: int) -> ReaderField:
261261
offs += int(name_len.nbytes + name_data.nbytes)
262262

263263
# Get Tensor Dimensions Count
264-
n_dims = self._get(offs, np.dtype(np.uint32))
264+
n_dims = self._get(offs, np.uint32)
265265
offs += int(n_dims.nbytes)
266266

267267
# Get Tensor Dimension Array
268-
dims = self._get(offs, np.dtype(np.uint64), n_dims[0].item())
268+
dims = self._get(offs, np.uint64, n_dims[0])
269269
offs += int(dims.nbytes)
270270

271271
# Get Tensor Encoding Scheme Type
272-
raw_dtype = self._get(offs, np.dtype(np.uint32))
272+
raw_dtype = self._get(offs, np.uint32)
273273
offs += int(raw_dtype.nbytes)
274274

275275
# Get Tensor Offset
276-
offset_tensor = self._get(offs, np.dtype(np.uint64))
276+
offset_tensor = self._get(offs, np.uint64)
277277
offs += int(offset_tensor.nbytes)
278278

279279
return ReaderField(
@@ -288,11 +288,11 @@ def _build_fields(self, offs: int, count: int) -> int:
288288
orig_offs = offs
289289
kv_klen, kv_kdata = self._get_str(offs)
290290
offs += int(kv_klen.nbytes + kv_kdata.nbytes)
291-
raw_kv_type = self._get(offs, np.dtype(np.uint32))
291+
raw_kv_type = self._get(offs, np.uint32)
292292
offs += int(raw_kv_type.nbytes)
293293
parts: list[npt.NDArray[Any]] = [kv_klen, kv_kdata, raw_kv_type]
294294
idxs_offs = len(parts)
295-
field_size, field_parts, field_idxs, field_types = self._get_field_parts(offs, raw_kv_type[0].item())
295+
field_size, field_parts, field_idxs, field_types = self._get_field_parts(offs, raw_kv_type[0])
296296
parts += field_parts
297297
self._push_field(ReaderField(
298298
orig_offs,
@@ -331,28 +331,28 @@ def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None:
331331
item_type: npt.DTypeLike
332332
if ggml_type == GGMLQuantizationType.F16:
333333
item_count = n_elems
334-
item_type = np.dtype(np.float16)
334+
item_type = np.float16
335335
elif ggml_type == GGMLQuantizationType.F32:
336336
item_count = n_elems
337-
item_type = np.dtype(np.float32)
337+
item_type = np.float32
338338
elif ggml_type == GGMLQuantizationType.F64:
339339
item_count = n_elems
340-
item_type = np.dtype(np.float64)
340+
item_type = np.float64
341341
elif ggml_type == GGMLQuantizationType.I8:
342342
item_count = n_elems
343-
item_type = np.dtype(np.int8)
343+
item_type = np.int8
344344
elif ggml_type == GGMLQuantizationType.I16:
345345
item_count = n_elems
346-
item_type = np.dtype(np.int16)
346+
item_type = np.int16
347347
elif ggml_type == GGMLQuantizationType.I32:
348348
item_count = n_elems
349-
item_type = np.dtype(np.int32)
349+
item_type = np.int32
350350
elif ggml_type == GGMLQuantizationType.I64:
351351
item_count = n_elems
352-
item_type = np.dtype(np.int64)
352+
item_type = np.int64
353353
else:
354354
item_count = n_bytes
355-
item_type = np.dtype(np.uint8)
355+
item_type = np.uint8
356356
np_dims = quant_shape_to_byte_shape(np_dims, ggml_type)
357357
tensors.append(ReaderTensor(
358358
name = tensor_name,

‎gguf-py/gguf/quants.py

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -93,11 +93, E281 11 @@ def __init_subclass__(cls, qtype: GGMLQuantizationType) -> None:
9393
cls.block_size, cls.type_size = GGML_QUANT_SIZES[qtype]
9494
cls.__quantize_lazy = LazyNumpyTensor._wrap_fn(
9595
cls.__quantize_array,
96-
meta_noop=(np.dtype(np.uint8), cls.__shape_to_bytes)
96+
meta_noop=(np.uint8, cls.__shape_to_bytes)
9797
)
9898
cls.__dequantize_lazy = LazyNumpyTensor._wrap_fn(
9999
cls.__dequantize_array,
100-
meta_noop=(np.dtype(np.float32), cls.__shape_from_bytes)
100+
meta_noop=(np.float32, cls.__shape_from_bytes)
101101
)
102102
assert qtype not in _type_traits
103103
_type_traits[qtype] = cls
@@ -165,12 +165,12 @@ def __shape_from_bytes(cls, shape: Sequence[int]):
165165

166166
@classmethod
167167
def __quantize_array(cls, array: np.ndarray) -> np.ndarray:
168-
return _apply_over_grouped_rows(cls.quantize_rows, arr=array, otype=np.dtype(np.uint8), oshape=cls.__shape_to_bytes(array.shape))
168+
return _apply_over_grouped_rows(cls.quantize_rows, arr=array, otype=np.uint8, oshape=cls.__shape_to_bytes(array.shape))
169169

170170
@classmethod
171171
def __dequantize_array(cls, array: np.ndarray) -> np.ndarray:
172172
cls.init_grid()
173-
return _apply_over_grouped_rows(cls.dequantize_rows, arr=array, otype=np.dtype(np.float32), oshape=cls.__shape_from_bytes(array.shape))
173+
return _apply_over_grouped_rows(cls.dequantize_rows, arr=array, otype=np.float32, oshape=cls.__shape_from_bytes(array.shape))
174174

175175
@classmethod
176176
def __quantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:

0 commit comments

Comments
 (0)
0