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

convert.py : add ftype when converting (does not work)

Browse source
This commit is contained in:
Georgi Gerganov 2023-08-22 16:19:15 +03:00
parent bee1f0e441
commit 32fc925ea8
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 23 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

23
convert.py
View file

@ -69,6 +69,9 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
'I32': DT_I32, 'I32': DT_I32,
} }
# TODO: match this with `llama_ftype`
# TODO: rename to LLAMAFileType
# TODO: move to `gguf.py`
class GGMLFileType(enum.Enum): class GGMLFileType(enum.Enum):
AllF32 = 0 AllF32 = 0
MostlyF16 = 1 # except 1d tensors MostlyF16 = 1 # except 1d tensors
@ -101,6 +104,8 @@ class Params:
n_head_kv: int n_head_kv: int
f_norm_eps: float f_norm_eps: float
ftype: Optional[GGMLFileType] = None
@staticmethod @staticmethod
def find_n_mult(n_ff: int, n_embd: int) -> int: def find_n_mult(n_ff: int, n_embd: int) -> int:
# hardcoded magic range # hardcoded magic range
@ -738,6 +743,9 @@ class OutputFile:
self.gguf.add_head_count_kv (params.n_head_kv) self.gguf.add_head_count_kv (params.n_head_kv)
self.gguf.add_layer_norm_rms_eps (params.f_norm_eps) self.gguf.add_layer_norm_rms_eps (params.f_norm_eps)
if params.ftype:
self.gguf.add_file_type(params.ftype)
def add_meta_vocab(self, vocab: Vocab) -> None: def add_meta_vocab(self, vocab: Vocab) -> None:
tokens = [] tokens = []
scores = [] scores = []
@ -1020,6 +1028,12 @@ def main(args_in: Optional[List[str]] = None) -> None:
" - LLaMA v2: --ctx 4096\n") " - LLaMA v2: --ctx 4096\n")
params.n_ctx = args.ctx params.n_ctx = args.ctx
if args.outtype:
params.ftype = {
"f32": GGMLFileType.AllF32,
"f16": GGMLFileType.MostlyF16,
}[args.outtype]
print(f"params = {params}") print(f"params = {params}")
vocab: Vocab vocab: Vocab
@ -1042,9 +1056,12 @@ def main(args_in: Optional[List[str]] = None) -> None:
model = model_plus.model model = model_plus.model
model = convert_model_names(model, params) model = convert_model_names(model, params)
output_type = pick_output_type(model, args.outtype) ftype = pick_output_type(model, args.outtype)
model = convert_to_output_type(model, output_type) model = convert_to_output_type(model, ftype)
outfile = args.outfile or default_outfile(model_plus.paths, output_type) outfile = args.outfile or default_outfile(model_plus.paths, ftype)
params.ftype = ftype
print(f"Writing {outfile}, format {ftype}")
OutputFile.write_all(outfile, params, model, vocab) OutputFile.write_all(outfile, params, model, vocab)
print(f"Wrote {outfile}") print(f"Wrote {outfile}")

2
gguf.py
View file

@ -597,7 +597,7 @@ class GGUFWriter:
self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo) self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo)
def add_file_type(self, ftype: int): def add_file_type(self, ftype: int):
self.add_string(KEY_GENERAL_FILE_TYPE, file_type) self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype)
def add_name(self, name: str): def add_name(self, name: str):
self.add_string(KEY_GENERAL_NAME, name) self.add_string(KEY_GENERAL_NAME, name)