feat: add custom tokenizer with multi-char Chinese token splitting

tokenization_voxcpm2.py

@@ -0,0 +1,72 @@
+"""Custom tokenizer for VoxCPM2 that splits multi-character Chinese tokens.
+
+VoxCPM2 was trained with ``mask_multichar_chinese_tokens`` which splits
+multi-character Chinese tokens (e.g. "你好" -> ["你", "好"]) into individual
+character IDs before embedding.  The base LlamaTokenizerFast produces
+multi-character Chinese tokens that the model has never seen during training,
+yielding garbled Chinese audio output in downstream inference frameworks.
+
+This module provides ``VoxCPM2Tokenizer`` which transparently applies the
+character splitting inside ``encode()`` and ``__call__()``, so any downstream
+consumer (vLLM, vLLM-Omni, Nano-vLLM, etc.) gets correct single-character
+IDs without code changes.
+"""
+
+from transformers import LlamaTokenizerFast
+
+
+class VoxCPM2Tokenizer(LlamaTokenizerFast):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._split_map = self._build_split_map()
+
+    def _build_split_map(self) -> dict[int, list[int]]:
+        vocab = self.get_vocab()
+        split_map: dict[int, list[int]] = {}
+        for token, tid in vocab.items():
+            clean = token.replace("\u2581", "")
+            if len(clean) >= 2 and all(self._is_cjk(c) for c in clean):
+                char_ids = self.convert_tokens_to_ids(list(clean))
+                if all(c != self.unk_token_id for c in char_ids):
+                    split_map[tid] = char_ids
+        return split_map
+
+    @staticmethod
+    def _is_cjk(c: str) -> bool:
+        return (
+            "\u4e00" <= c <= "\u9fff"
+            or "\u3400" <= c <= "\u4dbf"
+            or "\uf900" <= c <= "\ufaff"
+            or "\U00020000" <= c <= "\U0002a6df"
+        )
+
+    def _expand_ids(self, ids: list[int]) -> list[int]:
+        result: list[int] = []
+        for tid in ids:
+            expansion = self._split_map.get(tid)
+            if expansion is not None:
+                result.extend(expansion)
+            else:
+                result.append(tid)
+        return result
+
+    def encode(self, text, *args, **kwargs):
+        ids = super().encode(text, *args, **kwargs)
+        return self._expand_ids(ids)
+
+    def __call__(self, text, *args, **kwargs):
+        result = super().__call__(text, *args, **kwargs)
+        if hasattr(result, "input_ids"):
+            ids = result["input_ids"]
+            if isinstance(ids, list) and ids and isinstance(ids[0], list):
+                result["input_ids"] = [self._expand_ids(x) for x in ids]
+                if "attention_mask" in result:
+                    result["attention_mask"] = [
+                        [1] * len(x) for x in result["input_ids"]
+                    ]
+            elif isinstance(ids, list):
+                result["input_ids"] = self._expand_ids(ids)
+                if "attention_mask" in result:
+                    result["attention_mask"] = [1] * len(result["input_ids"])
+        return result

tokenizer_config.json

@@ -97,7 +97,7 @@
       "rstrip": false,
       "single_word": false,
       "special": true
-    },   
+    },
     "110": {
       "content": "<|/speaker_id|>",
       "lstrip": false,
@@ -205,8 +205,14 @@
   "pad_token": null,
   "sp_model_kwargs": {},
   "spaces_between_special_tokens": false,
-  "tokenizer_class": "LlamaTokenizer",
+  "tokenizer_class": "VoxCPM2Tokenizer",
   "unk_token": "<unk>",
   "use_default_system_prompt": false,
-  "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
+  "chat_template": "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_voxcpm2.VoxCPM2Tokenizer",
+      null
+    ]
+  }
 }