fix tokenizer

Author: cebtenzzre

convert_hf_to_gguf.py

@@ -695,8 +695,6 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e":
             # ref: https://huggingface.co/Xenova/gpt-4o
             res = "gpt-4o"
-        if chkhsh == "a81863d07e75497e2194eb1a1574d5e5cd4d5f85a87a0728b922bf2bed6fb327":
-            res = "bert"
 
         if res is None:
             logger.warning("\n")
@@ -3088,6 +3086,97 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
 
         return [(self.map_tensor_name(name), data_torch)]
 
+    def _xlmroberta_tokenizer_init(self) -> None:
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def _xlmroberta_set_vocab(self) -> None:
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(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.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        # realign tokens (see HF tokenizer code)
+        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
+        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+        toktypes = [
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.CONTROL,
+            SentencePieceTokenTypes.UNKNOWN,
+        ] + toktypes[3:-1]
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        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)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
 
 @Model.register("RobertaModel")
 class RobertaModel(BertModel):
@@ -3154,6 +3243,10 @@ def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path,
                          eager, metadata_override, model_name, split_max_tensors,
                          split_max_size, dry_run, small_first_shard, hparams)
 
+        self._tokenizer_is_xlmroberta = self._is_tokenizer_xlmroberta()
+        if self._tokenizer_is_xlmroberta:
+            self._xlmroberta_tokenizer_init()
+
         # the HF config claims n_ctx=8192, but it uses RoPE scaling
         self.hparams["n_ctx"] = 2048
 
@@ -3181,6 +3274,21 @@ def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path,
         assert self.hparams["rotary_emb_interleaved"] is False
         assert self.hparams["rotary_emb_scale_base"] is None
 
+    def _is_tokenizer_xlmroberta(self) -> bool:
+        with open(self.dir_model / "tokenizer.json") as f:
+            tokenizer_json = json.load(f)
+        toktyp = tokenizer_json["model"]["type"]
+        if toktyp == "Unigram":
+            return True
+        if toktyp == "WordPiece":
+            return False
+        raise ValueError(f"unknown tokenizer: {toktyp}")
+
+    def set_vocab(self) -> None:
+        if self._tokenizer_is_xlmroberta:
+            return self._xlmroberta_set_vocab()
+        return super().set_vocab()
+
     def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
         # If the tensor is an experts bias tensor, skip it by returning an empty list.
         if "mlp.experts.bias" in name:
@@ -3212,96 +3320,10 @@ class XLMRobertaModel(BertModel):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
+        self._xlmroberta_tokenizer_init()
 
-        # we need the pad_token_id to know how to chop down position_embd matrix
-        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
-            self._position_offset = 1 + pad_token_id
-            if "max_position_embeddings" in self.hparams:
-                self.hparams["max_position_embeddings"] -= self._position_offset
-        else:
-            self._position_offset = None
-
-    def set_vocab(self):
-        # to avoid TypeError: Descriptors cannot be created directly
-        # exception when importing sentencepiece_model_pb2
-        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
-        from sentencepiece import SentencePieceProcessor
-        from sentencepiece import sentencepiece_model_pb2 as model
-
-        tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
-        if not tokenizer_path.is_file():
-            raise FileNotFoundError(f"File not found: {tokenizer_path}")
-
-        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
-        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
-        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
-
-        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
-        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
-        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
-
-        tokenizer = SentencePieceProcessor()
-        tokenizer.LoadFromFile(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.UNUSED] * vocab_size
-
-        for token_id in range(tokenizer.vocab_size()):
-            piece = tokenizer.IdToPiece(token_id)
-            text = piece.encode("utf-8")
-            score = tokenizer.GetScore(token_id)
-
-            toktype = SentencePieceTokenTypes.NORMAL
-            if tokenizer.IsUnknown(token_id):
-                toktype = SentencePieceTokenTypes.UNKNOWN
-            elif tokenizer.IsControl(token_id):
-                toktype = SentencePieceTokenTypes.CONTROL
-            elif tokenizer.IsUnused(token_id):
-                toktype = SentencePieceTokenTypes.UNUSED
-            elif tokenizer.IsByte(token_id):
-                toktype = SentencePieceTokenTypes.BYTE
-
-            tokens[token_id] = text
-            scores[token_id] = score
-            toktypes[token_id] = toktype
-
-        if vocab_size > len(tokens):
-            pad_count = vocab_size - len(tokens)
-            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
-            for i in range(1, pad_count + 1):
-                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
-                scores.append(-1000.0)
-                toktypes.append(SentencePieceTokenTypes.UNUSED)
-
-        # realign tokens (see HF tokenizer code)
-        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
-        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
-        toktypes = [
-            SentencePieceTokenTypes.CONTROL,
-            SentencePieceTokenTypes.CONTROL,
-            SentencePieceTokenTypes.CONTROL,
-            SentencePieceTokenTypes.UNKNOWN,
-        ] + toktypes[3:-1]
-
-        self.gguf_writer.add_tokenizer_model("t5")
-        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)
-        self.gguf_writer.add_add_space_prefix(add_prefix)
-        self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
-        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
-        if precompiled_charsmap:
-            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
-
-        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
-        special_vocab.add_to_gguf(self.gguf_writer)
-
-        self.gguf_writer.add_add_bos_token(True)
-        self.gguf_writer.add_add_eos_token(True)
+    def set_vocab(self) -> None:
+        self._xlmroberta_set_vocab()
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # if name starts with "roberta.", remove the prefix

convert_hf_to_gguf_update.py

@@ -110,7 +110,6 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "deepseek-v3",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
     {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"},
     {"name": "gpt-4o",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", },
-    {"name": "nomic-embed-text-v2-moe",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe", },
 ]