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

convert: Handle special tokens and merges in vocab only mode

Browse source 
convert: Vocab only mode no longer requires loading model tensors
This commit is contained in:
KerfuffleV2 2023-08-27 13:17:18 -06:00
parent ea43267819
commit c7b0952eb7
1 changed files with 17 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

24
convert.py
View file

@ -299,8 +299,10 @@ class Params:
params = Params.loadHFTransformerJson(model_plus.model, hf_config_path) params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
elif orig_config_path.exists(): elif orig_config_path.exists():
params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path) params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
else: elif model_plus.format != 'none':
params = Params.guessed(model_plus.model) params = Params.guessed(model_plus.model)
else:
raise ValueError('Cannot guess params when model format is none')
params.path_model = model_plus.paths[0].parent params.path_model = model_plus.paths[0].parent
@ -597,7 +599,7 @@ LazyModel = Dict[str, LazyTensor]
class ModelPlus: class ModelPlus:
model: LazyModel model: LazyModel
paths: List[Path] # Where this was read from. paths: List[Path] # Where this was read from.
format: Literal['ggml', 'torch', 'safetensors'] format: Literal['ggml', 'torch', 'safetensors', 'none']
vocab: Optional[Vocab] # For GGML models (which have vocab built in), the vocab. vocab: Optional[Vocab] # For GGML models (which have vocab built in), the vocab.
@ -930,7 +932,7 @@ class OutputFile:
self.gguf.close() self.gguf.close()
@staticmethod @staticmethod
def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab) -> None: def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: SpecialVocab) -> None:
check_vocab_size(params, vocab) check_vocab_size(params, vocab)
of = OutputFile(fname_out) of = OutputFile(fname_out)
@ -938,6 +940,8 @@ class OutputFile:
# meta data # meta data
of.add_meta_arch(params) of.add_meta_arch(params)
of.add_meta_vocab(vocab) of.add_meta_vocab(vocab)
of.add_meta_special_vocab(svocab)
of.write_meta() of.write_meta()
of.close() of.close()
@ -1187,8 +1191,13 @@ def main(args_in: Optional[List[str]] = None) -> None:
if args.dump_single: if args.dump_single:
model_plus = lazy_load_file(args.model) model_plus = lazy_load_file(args.model)
do_dump_model(model_plus) do_dump_model(model_plus)
return
model_plus = load_some_model(args.model) if not args.vocab_only:
model_plus = load_some_model(args.model)
else:
# You can no longer use guessed parameters for your vocab only model. Does anyone actually care?
model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
if args.dump: if args.dump:
do_dump_model(model_plus) do_dump_model(model_plus)
@ -1214,11 +1223,12 @@ def main(args_in: Optional[List[str]] = None) -> None:
vocab: Vocab vocab: Vocab
if args.vocab_only: if args.vocab_only:
# FIXME: Handle special vocab here also.
vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
assert args.outfile, "need --outfile if using --vocab-only" assert args.outfile, "need --outfile if using --vocab-only"
# FIXME: Try to respect vocab_dir somehow?
vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
special_vocab = SpecialVocab(model_plus.paths[0].parent)
outfile = args.outfile outfile = args.outfile
OutputFile.write_vocab_only(outfile, params, vocab) OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
print(f"Wrote {outfile}") print(f"Wrote {outfile}")
return return