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

Add test for MPT tokenization (#3728)

Browse source 
* Add test for MPT tokenization

* Revert code motion

* Remove unnecessary restriction in test case

* Clarify logic in conversion
This commit is contained in:
goerch 2023-10-22 21:21:42 +02:00 • committed by GitHub
parent 5a42a5f8e8
commit 9e70cc0322
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 21 additions and 12 deletions
Download patch file Download diff file Expand all files Collapse all files

15
convert-mpt-hf-to-gguf.py
View file

@ -128,15 +128,22 @@ vocab_size = hparams["vocab_size"]
# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
tokenizer = AutoTokenizer.from_pretrained(dir_model)
added_vocab = tokenizer.get_added_vocab()
reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
for i in range(vocab_size):
tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
scores.append(0.0) # dummy
toktypes.append(gguf.TokenType.NORMAL)
if i not in reverse_vocab:
tokens.append(f"[PAD{i}]")
toktypes.append(gguf.TokenType.USER_DEFINED)
elif reverse_vocab[i] in added_vocab:
# NOTE: wouldn't we like to distinguish CONTROL tokens here?
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.USER_DEFINED)
else:
tokens.append(reverse_vocab[i])
toktypes.append(gguf.TokenType.NORMAL)
gguf_writer.add_token_list(tokens)
gguf_writer.add_token_scores(scores)
gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))