[Model] Support Hy3 token suffix and JSON Schema array types (#47192)

Signed-off-by: stevenkuang-tencent <stevenkuang@tencent.com>

vllm/reasoning/hy_v3_reasoning_parser.py

@@ -26,6 +26,8 @@ class HYV3ReasoningParser(BaseThinkingReasoningParser):
     """
 
     def __init__(self, tokenizer: TokenizerLike, *args, **kwargs):
+        init_kwargs = getattr(tokenizer, "init_kwargs", None) or {}
+        self.suffix: str = init_kwargs.get("token_suffix") or ""
         super().__init__(tokenizer, *args, **kwargs)
 
         # First, If there is reasoning_effort in chat_kwargs,
@@ -52,12 +54,12 @@ class HYV3ReasoningParser(BaseThinkingReasoningParser):
     @property
     def start_token(self) -> str:
         """The token that starts reasoning content."""
-        return "<think>"
+        return f"<think{self.suffix}>"
 
     @property
     def end_token(self) -> str:
         """The token that ends reasoning content."""
-        return "</think>"
+        return f"</think{self.suffix}>"
 
     def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
         if self._identity_parser is not None:

vllm/tool_parsers/hy_v3_tool_parser.py

@@ -108,6 +108,14 @@ class HYV3ToolParser(ToolParser):
         Note: single ``type`` has the highest priority.
         """
         if "type" in arg_schema:
+            type_val = arg_schema["type"]
+            # JSON Schema allows "type" to be an array to represent union types,
+            # e.g. "type": ["string", "object"].
+            # Expand it into an anyOf-equivalent format:
+            #   [{"type": "string"}, {"type": "object"}]
+            # so that _get_types / _parse_value can handle it uniformly later.
+            if isinstance(type_val, list):
+                return [{"type": t} for t in type_val]
             return [arg_schema]
         if "anyOf" in arg_schema:
             return arg_schema["anyOf"]
@@ -261,19 +269,22 @@ class HYV3ToolParser(ToolParser):
         self._current_arg_is_string: bool = False  # is current arg pure string?
         self._streamed_json_len: int = 0  # bytes of JSON already sent
 
-        self.tool_calls_start_token: str = "<tool_calls>"
-        self.tool_calls_end_token: str = "</tool_calls>"
+        init_kwargs = getattr(tokenizer, "init_kwargs", None) or {}
+        self.suffix: str = init_kwargs.get("token_suffix") or ""
 
-        self.tool_call_start_token: str = "<tool_call>"
-        self.tool_call_end_token: str = "</tool_call>"
+        self.tool_calls_start_token: str = f"<tool_calls{self.suffix}>"
+        self.tool_calls_end_token: str = f"</tool_calls{self.suffix}>"
 
-        self.tool_sep_token: str = "<tool_sep>"
+        self.tool_call_start_token: str = f"<tool_call{self.suffix}>"
+        self.tool_call_end_token: str = f"</tool_call{self.suffix}>"
 
-        self.arg_key_start_token: str = "<arg_key>"
-        self.arg_key_end_token: str = "</arg_key>"
+        self.tool_sep_token: str = f"<tool_sep{self.suffix}>"
 
-        self.arg_value_start_token: str = "<arg_value>"
-        self.arg_value_end_token: str = "</arg_value>"
+        self.arg_key_start_token: str = f"<arg_key{self.suffix}>"
+        self.arg_key_end_token: str = f"</arg_key{self.suffix}>"
+
+        self.arg_value_start_token: str = f"<arg_value{self.suffix}>"
+        self.arg_value_end_token: str = f"</arg_value{self.suffix}>"
 
         self.tool_call_regex = re.compile(
             rf"{self.tool_call_start_token}(.*?){self.tool_sep_token}"