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

Add big endian support

Browse source
This commit is contained in:
Galunid 2023-11-04 23:01:38 +01:00
parent 03c9683eb7
commit fd30850576
3 changed files with 11 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
convert-generic.py
View file

@ -28,12 +28,16 @@ else:
# output in the same directory as the model by default # output in the same directory as the model by default
fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
is_big_endian = False
if args.bigendian is not None:
is_big_endian = True
print(f"Loading model: {dir_model.name}") print(f"Loading model: {dir_model.name}")
hparams = model.Model.load_hparams(dir_model) hparams = model.Model.load_hparams(dir_model)
model_class = model.Model.from_model_architecture(hparams["architectures"][0]) model_class = model.Model.from_model_architecture(hparams["architectures"][0])
model_instance = model_class(dir_model, ftype, fname_out) model_instance = model_class(dir_model, ftype, fname_out, is_big_endian)
print("Set model parameters") print("Set model parameters")
model_instance.set_gguf_parameters() model_instance.set_gguf_parameters()

6
model.py
View file

@ -30,16 +30,18 @@ class SentencePieceTokenTypes(IntEnum):
class Model: class Model:
def __init__(self, dir_model: Path, ftype: int, fname_out: Path): def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
self.dir_model = dir_model self.dir_model = dir_model
self.ftype = ftype self.ftype = ftype
self.fname_out = fname_out self.fname_out = fname_out
self.is_big_endian = is_big_endian
self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
self.is_safetensors = self._is_model_safetensors() self.is_safetensors = self._is_model_safetensors()
self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin") self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
self.part_names = self._get_part_names() self.part_names = self._get_part_names()
self.hparams = Model.load_hparams(self.dir_model) self.hparams = Model.load_hparams(self.dir_model)
self.model_arch = self._get_model_architecture() self.model_arch = self._get_model_architecture()
self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch]) self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess)
def set_vocab(self): def set_vocab(self):
self._set_vocab_gpt2() self._set_vocab_gpt2()

2
util.py
View file

@ -20,4 +20,6 @@ def parse_args() -> argparse.Namespace:
"ftype", type=int, choices=[0, 1], default=1, nargs='?', "ftype", type=int, choices=[0, 1], default=1, nargs='?',
help="output format - use 0 for float32, 1 for float16", help="output format - use 0 for float32, 1 for float16",
) )
parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine")
return parser.parse_args() return parser.parse_args()