convert : add "tokenizer.ggml.pre" GGUF KV (wip)

convert-hf-to-gguf.py

@@ -346,7 +346,7 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH:
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
     # used for GPT-2 BPE and WordPiece vocabs
-    def get_basic_vocab(self) -> tuple[list[str], list[int]]:
+    def get_vocab_base(self) -> tuple[list[str], list[int], str]:
         tokens: list[str] = []
         toktypes: list[int] = []
 
@@ -355,6 +355,8 @@ def get_basic_vocab(self) -> tuple[list[str], list[int]]:
         vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))
         assert max(tokenizer.vocab.values()) < vocab_size
 
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
         added_vocab = tokenizer.get_added_vocab()
 
@@ -372,11 +374,41 @@ def get_basic_vocab(self) -> tuple[list[str], list[int]]:
                 tokens.append(reverse_vocab[i])
                 toktypes.append(gguf.TokenType.NORMAL)
 
-        return tokens, toktypes
+        return tokens, toktypes, tokpre
+
+    def get_vocab_base_pre(self, tokenizer) -> str:
+        # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
+        # is specific for the BPE pre-tokenizer used by the model
+        # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
+        # use in llama.cpp to implement the same pre-tokenizer
+
+        chktxt = "\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български what's ''''''```````\"\"\"\"......!!!!!!??????"
+
+        chktok = tokenizer.encode(chktxt)
+        chkhsh = hash(tuple(chktok))
+
+        print(f"chktok: {chktok}")
+        print(f"chkhsh: {chkhsh}")
+
+        res = None
 
-    def _set_vocab_gpt2(self, tokenizer_model:str = "gpt2") -> None:
-        tokens, toktypes = self.get_basic_vocab()
-        self.gguf_writer.add_tokenizer_model(tokenizer_model)
+        # NOTE: if you get an error here, you need to add the model to the if-elif chain below
+        #       observe the stdout for the chkhsh value and add it to the chain
+        if self.model_arch == gguf.MODEL_ARCH.LLAMA:
+            if chkhsh == -3290901550109860290:
+                # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct/blob/main/tokenizer.json
+                res = "llama3"
+            if chkhsh ==  4190561703949727616:
+                # ref: https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct/blob/main/tokenizer.json
+                res = "deepseek-coder"
+
+        if res is None:
+            raise NotImplementedError(f"BPE pre-tokenizer was not recognized - update get_vocab_base_pre()")
+
+    def _set_vocab_gpt2(self) -> None:
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
 
@@ -394,6 +426,8 @@ def _set_vocab_qwen(self):
         vocab_size = hparams["vocab_size"]
         assert max(tokenizer.get_vocab().values()) < vocab_size
 
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
         merges = []
         vocab = {}
         mergeable_ranks = tokenizer.mergeable_ranks
@@ -421,6 +455,7 @@ def _set_vocab_qwen(self):
                 toktypes.append(gguf.TokenType.NORMAL)
 
         self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
 
@@ -493,6 +528,7 @@ def _set_vocab_sentencepiece(self):
         assert len(tokens) == vocab_size
 
         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)
@@ -514,6 +550,7 @@ def _set_vocab_llama_hf(self):
         assert len(tokens) == vocab.vocab_size
 
         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)
@@ -957,6 +994,7 @@ def set_vocab(self):
             toktypes.append(toktype)
 
         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_types(toktypes)
 
@@ -2174,6 +2212,7 @@ def set_vocab(self):
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
 
         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)
@@ -2416,6 +2455,7 @@ def set_vocab(self):
                     toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
 
         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)
@@ -2565,7 +2605,7 @@ def set_gguf_parameters(self):
             self.gguf_writer.add_pooling_type(pooling_type)
 
     def set_vocab(self):
-        tokens, toktypes = self.get_basic_vocab()
+        tokens, toktypes, tokpre = self.get_vocab_base()
         self.vocab_size = len(tokens)
 
         # we need this to validate the size of the token_type embeddings
@@ -2583,6 +2623,7 @@ def phantom(tok):
 
         # add vocab to gguf
         self.gguf_writer.add_tokenizer_model("bert")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
 
@@ -2760,6 +2801,9 @@ def set_vocab(self):
             field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
             self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
 
+            field = neox_reader.get_field(gguf.Keys.Tokenizer.PRE)
+            self.gguf_writer.add_tokenizer_pre(bytes(field.parts[-1]))
+
             field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
             self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
 

convert-llama-ggml-to-gguf.py

@@ -281,6 +281,7 @@ def add_params(self, gguf_writer):
     def add_vocab(self, gguf_writer):
         hp = self.model.hyperparameters
         gguf_writer.add_tokenizer_model('llama')
+        gguf_writer.add_tokenizer_pre('default')
         tokens = []
         scores = []
         toktypes = []

convert-persimmon-to-gguf.py

@@ -99,6 +99,7 @@ def main():
 
     tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
     gguf_writer.add_tokenizer_model('llama')
+    gguf_writer.add_tokenizer_pre('default')
     gguf_writer.add_token_list(tokens)
     gguf_writer.add_token_scores(scores)
     gguf_writer.add_token_types(toktypes)

gguf-py/gguf/constants.py

@@ -72,6 +72,7 @@ class SSM:
 
     class Tokenizer:
         MODEL            = "tokenizer.ggml.model"
+        PRE              = "tokenizer.ggml.pre"
         LIST             = "tokenizer.ggml.tokens"
         TOKEN_TYPE       = "tokenizer.ggml.token_type"
         TOKEN_TYPE_COUNT = "tokenizer.ggml.token_type_count"  # for BERT-style token types
@@ -940,6 +941,7 @@ def get_type(val: Any) -> GGUFValueType:
 
 # tokenization
 KEY_TOKENIZER_MODEL      = Keys.Tokenizer.MODEL
+KEY_TOKENIZER_PRE        = Keys.Tokenizer.PRE
 KEY_TOKENIZER_LIST       = Keys.Tokenizer.LIST
 KEY_TOKENIZER_TOKEN_TYPE = Keys.Tokenizer.TOKEN_TYPE
 KEY_TOKENIZER_SCORES     = Keys.Tokenizer.SCORES

gguf-py/gguf/gguf_reader.py

@@ -140,7 +140,7 @@ def _get(
     def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
         if field.name in self.fields:
             # TODO: add option to generate error on duplicate keys
-            #raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
+            # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
 
             print(f'Warning: Duplicate key {field.name} at offset {field.offset}')
             self.fields[field.name + '_{}'.format(field.offset)] = field

gguf-py/gguf/gguf_writer.py

@@ -422,6 +422,9 @@ def add_ssm_time_step_rank(self, value: int) -> None:
     def add_tokenizer_model(self, model: str) -> None:
         self.add_string(Keys.Tokenizer.MODEL, model)
 
+    def add_tokenizer_pre(self, pre: str) -> None:
+        self.add_string(Keys.Tokenizer.PRE, pre)
+
     def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None:
         self.add_array(Keys.Tokenizer.LIST, tokens)