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

convert : various script cleanups/fixes + merges and special token handling (#2842)

Browse source 
* convert: Fix permute calls and method/func definitions

* Cleanups for gguf-py

* Minor types cleanups.

* Initial implementation of handling merges and special tokens

* convert: Handle special tokens and merges in vocab only mode

convert: Vocab only mode no longer requires loading model tensors

* gguf: Refactor tensor name mapping

* convert: Fix type hint for special_token_types in SpecialVocab

* Use common special vocab handling in various conversion scripts

* First pass at implementing suggested changes

* Second pass

* gguf: SpecialVocab: Fix issue with special token content not in a dict

gguf: SpecialVocab: Allow skipping handling of merges

* convert-falcon-hf-to-gguf: Support --vocab-only option, bail out if no tokenizer.json

* convert-gptneox-hf-to-gguf and convert: Only handle merges for BPE tokenizer

* gguf: SpecialVocab: Actually set load_merges in object

* Uniform args parsing and vocab only mode for convert examples

* convert.py: Set gpt2 as tokenizer model when using BPE

* Squish last type warning in gguf.py - yay!
This commit is contained in:
Kerfuffle 2023-08-30 02:25:50 -06:00 • committed by GitHub
parent ad9ddcff6e
commit dc07dc492e
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
10 changed files with 728 additions and 748 deletions
Download patch file Download diff file Expand all files Collapse all files

28
convert-llama-ggmlv3-to-gguf.py
View file

@ -75,7 +75,7 @@ class Tensor:
self.dims = ()
self.dtype = None
self.start_offset = 0
self.len_bytes = 0
self.len_bytes = np.int64(0)
def load(self, data, offset):
orig_offset = offset
@ -134,13 +134,14 @@ class GGMLV3Model:
return offset
class GGMLToGGUF:
def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None):
def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
hp = ggml_model.hyperparameters
self.model = ggml_model
self.data = data
self.cfg = cfg
self.params_override = params_override
self.vocab_override = vocab_override
self.special_vocab = special_vocab
if params_override is not None:
n_kv_head = params_override.n_head_kv
else:
@ -162,6 +163,8 @@ class GGMLToGGUF:
gguf_writer = gguf.GGUFWriter(self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
self.add_params(gguf_writer)
self.add_vocab(gguf_writer)
if self.special_vocab is not None:
self.special_vocab.add_to_gguf(gguf_writer)
self.add_tensors(gguf_writer)
print(" gguf: write header")
gguf_writer.write_header_to_file()
@ -259,20 +262,13 @@ class GGMLToGGUF:
gguf_writer.add_eos_token_id(2)
def add_tensors(self, gguf_writer):
nm = self.name_map
tensor_map = self.name_map
data = self.data
print(f'* Adding {len(self.model.tensors)} tensor(s)')
for tensor in self.model.tensors:
name = str(tensor.name, 'UTF-8')
if name.endswith('.weight'):
name = name[:-7]
suffix = '.weight'
elif name.endswith('.bias'):
name = name[:-5]
suffix = '.bias'
mapped_name = nm.get(name)
mapped_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
assert mapped_name is not None, f'Bad name {name}'
mapped_name += suffix
tempdims = list(tensor.dims[:])
if len(tempdims) > 1:
temp = tempdims[1]
@ -302,8 +298,10 @@ def handle_metadata(cfg, hp):
else:
raise ValueError('Unable to load metadata')
vocab = convert.load_vocab(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, cfg.vocabtype)
# FIXME: Respect cfg.vocab_dir?
svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
convert.check_vocab_size(params, vocab)
return (params, vocab)
return (params, vocab, svocab)
def handle_args():
parser = argparse.ArgumentParser(description = 'Convert GGMLv3 models to GGUF')
@ -330,14 +328,16 @@ def main():
print(f'* GGML model hyperparameters: {model.hyperparameters}')
vocab_override = None
params_override = None
special_vocab = None
if cfg.model_metadata_dir is not None:
(params_override, vocab_override) = handle_metadata(cfg, model.hyperparameters)
(params_override, vocab_override, special_vocab) = handle_metadata(cfg, model.hyperparameters)
print('!! Note: When overriding params the --gqa, --eps and --context-length options are ignored.')
print(f'* Overriding params: {params_override}')
print(f'* Overriding vocab: {vocab_override}')
print(f'* Special vocab: {special_vocab}')
else:
print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override)
converter = GGMLToGGUF(model, data, cfg, params_override = params_override, vocab_override = vocab_override, special_vocab = special_vocab)
converter.save()
print(f'* Successful completion. Output saved to: {cfg.output}')