Use common special vocab handling in various conversion scripts

convert-falcon-hf-to-gguf.py

@@ -13,6 +13,8 @@ from typing import Any, List
 from pathlib import Path
 from transformers import AutoTokenizer
 
+from convert import SpecialVocab
+
 def bytes_to_unicode():
     # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
     """
@@ -116,20 +118,13 @@ print("gguf: get tokenizer metadata")
 tokens: List[bytearray] = []
 scores: List[float] = []
 toktypes: List[int] = []
-merges: List[str] = []
-
 
 if Path(dir_model + "/tokenizer.json").is_file():
     # gpt2 tokenizer
     gguf_writer.add_tokenizer_model("gpt2")
 
-    print("gguf: get gpt2 tokenizer merges")
-
     with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
         tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
-
-    gguf_writer.add_token_merges(merges)
 
     print("gguf: get gpt2 tokenizer vocab")
 
@@ -166,24 +161,8 @@ if Path(dir_model + "/tokenizer.json").is_file():
     gguf_writer.add_token_scores(scores)
     gguf_writer.add_token_types(toktypes)
 
-print("gguf: get special token ids")
+special_vocab = SpecialVocab(Path(dir_model))
-# Look for special tokens in config.json
+special_vocab.add_to_gguf(gguf_writer)
-
-if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-    gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-    gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-    gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-    gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-    gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
 
 # TENSORS
 

convert-gptneox-hf-to-gguf.py

@@ -13,6 +13,8 @@ from typing import Any, List
 from pathlib import Path
 from transformers import AutoTokenizer
 
+from convert import SpecialVocab
+
 # ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
 
 
@@ -112,20 +114,13 @@ gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
 print("gguf: get tokenizer metadata")
 
 tokens: List[bytearray] = []
-merges: List[str] = []
-
 
 if Path(dir_model + "/tokenizer.json").is_file():
     # gpt2 tokenizer
     gguf_writer.add_tokenizer_model("gpt2")
 
-    print("gguf: get gpt2 tokenizer merges")
-
     with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
         tokenizer_json = json.load(f)
-    merges = tokenizer_json["model"]["merges"]
-
-    gguf_writer.add_token_merges(merges)
 
     print("gguf: get gpt2 tokenizer vocab")
 
@@ -158,39 +153,8 @@ if Path(dir_model + "/tokenizer.json").is_file():
 
     gguf_writer.add_token_list(tokens)
 
-    if "added_tokens" in tokenizer_json and Path(dir_model + "/tokenizer_config.json").is_file():
+special_vocab = SpecialVocab(Path(dir_model))
-        print("gguf: get special token ids")
+special_vocab.add_to_gguf(gguf_writer)
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        # find special token ids
-
-        if "bos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config:
-            for key in tokenizer_json["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-
 
 # TENSORS
 

convert-llama-7b-pth-to-gguf.py

@@ -15,6 +15,8 @@ from typing import Any, List, TypeAlias
 from pathlib import Path
 from sentencepiece import SentencePieceProcessor
 
+from convert import SpecialVocab
+
 #NDArray = np.ndarray[Any, Any]
 # compatible with python < 3.9
 NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
@@ -180,62 +182,8 @@ if Path(dir_model + "/tokenizer.model").is_file():
     gguf_writer.add_token_scores(scores)
     gguf_writer.add_token_types(toktypes)
 
-
+special_vocab = SpecialVocab(Path(dir_model))
-print("gguf: get special token ids")
+special_vocab.add_to_gguf(gguf_writer)
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
 
 # TENSORS
 

convert-llama-ggmlv3-to-gguf.py

@@ -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()
@@ -295,8 +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 = convert.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')
@@ -323,14 +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}')
 

convert-llama-hf-to-gguf.py

@@ -13,6 +13,8 @@ from typing import Any, List, Optional, TypeAlias
 from pathlib import Path
 from sentencepiece import SentencePieceProcessor
 
+from convert import SpecialVocab
+
 #NDArray = np.ndarray[Any, Any]
 # compatible with python < 3.9
 NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
@@ -189,62 +191,8 @@ if Path(dir_model + "/tokenizer.model").is_file():
     gguf_writer.add_token_scores(scores)
     gguf_writer.add_token_types(toktypes)
 
-
+special_vocab = SpecialVocab(Path(dir_model))
-print("gguf: get special token ids")
+special_vocab.add_to_gguf(gguf_writer)
-
-if Path(dir_model + "/tokenizer.json").is_file():
-    # Look for special tokens in tokenizer.json if it exists
-
-    with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
-        tokenizer = json.load(f)
-
-    if "added_tokens" in tokenizer and Path(dir_model + "/tokenizer_config.json").is_file():
-
-        with open(dir_model + "/tokenizer_config.json", "r", encoding="utf-8") as f:
-            tokenizer_config = json.load(f)
-
-        if "bos_token" in tokenizer_config and tokenizer_config["bos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["bos_token"]["content"]:
-                    gguf_writer.add_bos_token_id(key["id"])
-
-        if "eos_token" in tokenizer_config and tokenizer_config["eos_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["eos_token"]["content"]:
-                    gguf_writer.add_eos_token_id(key["id"])
-
-        if "unk_token" in tokenizer_config and tokenizer_config["unk_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["unk_token"]["content"]:
-                    gguf_writer.add_unk_token_id(key["id"])
-
-        if "sep_token" in tokenizer_config and tokenizer_config["sep_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["sep_token"]["content"]:
-                    gguf_writer.add_sep_token_id(key["id"])
-
-        if "pad_token" in tokenizer_config and tokenizer_config["pad_token"] != None:
-            for key in tokenizer["added_tokens"]:
-                if key["content"] == tokenizer_config["pad_token"]["content"]:
-                    gguf_writer.add_pad_token_id(key["id"])
-else:
-    # If no tokenizer.json: Look for special tokens in config.json
-
-    if "bos_token_id" in hparams and hparams["bos_token_id"] != None:
-        gguf_writer.add_bos_token_id(hparams["bos_token_id"])
-
-    if "eos_token_id" in hparams and hparams["eos_token_id"] != None:
-        gguf_writer.add_eos_token_id(hparams["eos_token_id"])
-
-    if "unk_token_id" in hparams and hparams["unk_token_id"] != None:
-        gguf_writer.add_unk_token_id(hparams["unk_token_id"])
-
-    if "sep_token_id" in hparams and hparams["sep_token_id"] != None:
-        gguf_writer.add_sep_token_id(hparams["sep_token_id"])
-
-    if "pad_token_id" in hparams and hparams["pad_token_id"] != None:
-        gguf_writer.add_pad_token_id(hparams["pad_token_id"])
-
 
 # TENSORS