vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Make GGUFReader endian detection less arbitrary

Browse source
This commit is contained in:
KerfuffleV2 2023-11-09 16:55:58 -07:00
parent 52bdc7e946
commit a04f0487b0
1 changed files with 5 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

7
gguf-py/gguf/gguf_reader.py
View file

@ -62,7 +62,8 @@ class ReaderTensor(NamedTuple):
class GGUFReader: class GGUFReader:
byte_order: Literal['I' | 'S' | '<'] = 'I' # I - same as host, S - swapped
byte_order: Literal['I' | 'S'] = 'I'
alignment: int = GGUF_DEFAULT_ALIGNMENT alignment: int = GGUF_DEFAULT_ALIGNMENT
# Note: Internal helper, API may change. # Note: Internal helper, API may change.
@ -87,7 +88,9 @@ class GGUFReader:
raise ValueError('GGUF magic invalid') raise ValueError('GGUF magic invalid')
offs += 4 offs += 4
temp_version = self._get(offs, np.uint32) temp_version = self._get(offs, np.uint32)
if temp_version[0] > 2000: if temp_version[0] & 65535 == 0:
# If we get 0 here that means it's (probably) a GGUF file created for
# the opposite byte order of the machine this script is running on.
self.byte_order = 'S' self.byte_order = 'S'
temp_version = temp_version.newbyteorder(self.byte_order) temp_version = temp_version.newbyteorder(self.byte_order)
version = temp_version[0] version = temp_version[0]