From c6f15a752ae7733f7e5bcd84caae1a4bc6a3b8f4 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Tue, 7 May 2024 20:57:28 +0200 Subject: [PATCH] Read vocabulary for ArcticForCausalLM from sentencepiece model instead of HF tokenizer. Add/redefine tokens accordingly to added_tokens_decoder from tokenizer_config.json --- convert-hf-to-gguf.py | 82 ++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 81 insertions(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 17cdb0fc5..dd7b7c6fc 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -1522,7 +1522,87 @@ class ArcticModel(Model): model_arch = gguf.MODEL_ARCH.ARCTIC def set_vocab(self): - self._set_vocab_llama_hf() + # The reason for using a custom implementation here is that the + # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from + # tokenizer.model and used them as BOS and EOS instead of adding new tokens. + from sentencepiece import SentencePieceProcessor + + tokenizer_path = self.dir_model / 'tokenizer.model' + + if not tokenizer_path.is_file(): + print(f'Error: Missing {tokenizer_path}', file=sys.stderr) + sys.exit(1) + + # Read the whole vocabulary from the tokenizer.model file + tokenizer = SentencePieceProcessor(str(tokenizer_path)) + + vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) + + tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)] + scores: list[float] = [-10000.0] * vocab_size + toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size + + for token_id in range(tokenizer.vocab_size()): + + piece = tokenizer.id_to_piece(token_id) + text = piece.encode("utf-8") + score = tokenizer.get_score(token_id) + + toktype = SentencePieceTokenTypes.NORMAL + if tokenizer.is_unknown(token_id): + toktype = SentencePieceTokenTypes.UNKNOWN + elif tokenizer.is_control(token_id): + toktype = SentencePieceTokenTypes.CONTROL + elif tokenizer.is_unused(token_id): + toktype = SentencePieceTokenTypes.UNUSED + elif tokenizer.is_byte(token_id): + toktype = SentencePieceTokenTypes.BYTE + + tokens[token_id] = text + scores[token_id] = score + toktypes[token_id] = toktype + + # Use the added_tokens_decoder field from tokeniser_config.json as the source + # of information about added/redefined tokens and modify them accordingly. + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + if tokenizer_config_file.is_file(): + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + tokenizer_config_json = json.load(f) + + if "added_tokens_decoder" in tokenizer_config_json: + added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"] + for token_id, token_json in added_tokens_decoder.items(): + token_id = int(token_id) + if (token_id >= vocab_size): + print(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') + continue + + token_content = token_json["content"] + token_type = SentencePieceTokenTypes.USER_DEFINED + token_score = -10000.0 + + # Map unk_token to UNKNOWN, other special tokens to CONTROL + # Set the score to 0.0 as in the original tokenizer.model + if ("special" in token_json) and token_json["special"]: + if token_content == tokenizer_config_json["unk_token"]: + token_type = SentencePieceTokenTypes.UNKNOWN + else: + token_type = SentencePieceTokenTypes.CONTROL + token_score = 0.0 + + print(f"Setting token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})") + tokens[token_id] = token_content.encode("utf-8") + toktypes[token_id] = token_type + scores[token_id] = token_score + + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_tokenizer_pre("default") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): super().set_gguf_parameters()