b4增加模型原生内置tool_schema传参

.gitignore

@@ -2,5 +2,6 @@ __pycache__/
 *.py[cod]
 outputs/_test_*/
 outputs/tmp*/
+outputs/*
 models/*
 !models/.gitkeep

code/b4_local_agent_llm.py

@@ -256,7 +256,7 @@ def _load_model_bundle(
     return tokenizer, model
 
 
-def _build_prompt_messages(messages: list[dict], tools_schema: list[dict]) -> list[dict]:
+def _build_prompt_messages(messages: list[dict], tools_schema: list[dict],include_tools_schema: bool = True) -> list[dict]:
     prompt_messages = deepcopy(messages)
     format_instruction = (
         "IMPORTANT OUTPUT FORMAT:\n"
@@ -284,12 +284,17 @@ def _build_prompt_messages(messages: list[dict], tools_schema: list[dict]) -> li
         "Choose exactly one schema: final content with an empty tool_calls array, or empty content with tool calls. "
         'Never put tool_calls inside content. Never output {"content":"tool_calls": ...}.'
     )
+
+    system_instruction = "\n\n" + format_instruction
+
+    if include_tools_schema:
         system_instruction = (
             "\n\nAvailable tools JSON schema:\n"
             + json.dumps(tools_schema, ensure_ascii=False)
             + "\n"
             + format_instruction
         )
+
     if prompt_messages and prompt_messages[0].get("role") == "system":
         prompt_messages[0]["content"] += system_instruction
     else:
@@ -344,7 +349,42 @@ def _prompt_json_generate(config_path: Path, config: dict, messages: list[dict],
         model_config.get("device_map", "auto"),
         model_config.get("max_memory"),
     )
-    prompt_messages = _build_prompt_messages(messages, tools_schema)
+    tool_calling_config = config.get("tool_calling", {})
+    schema_injection = tool_calling_config.get("schema_injection", "prompt")
+
+    if schema_injection not in {"prompt", "native"}:
+        raise ValueError(
+            "tool_calling.schema_injection must be prompt or native"
+        )
+
+    template_kwargs = {}
+
+    if schema_injection == "prompt":
+        # 原有方案：schema 转成文字后拼接到 system prompt。
+        prompt_messages = _build_prompt_messages(
+            messages,
+            tools_schema,
+            include_tools_schema=True,
+        )
+
+    else:
+        # 新方案：schema 不写入 prompt，通过 tokenizer 的 tools 参数注入。
+        prompt_messages = _build_prompt_messages(
+            messages,
+            tools_schema,
+            include_tools_schema=False,
+        )
+
+        # 内部临时移除B3 的 x-returns 字段，
+        native_tools = deepcopy(tools_schema)
+
+        for tool in native_tools:
+            function = tool.get("function")
+            if isinstance(function, dict):
+                function.pop("x-returns", None)
+
+        template_kwargs["tools"] = native_tools
+
     inputs = tokenizer.apply_chat_template(
         prompt_messages,
         tokenize=True,
@@ -352,6 +392,7 @@ def _prompt_json_generate(config_path: Path, config: dict, messages: list[dict],
         return_tensors="pt",
         return_dict=True,
         enable_thinking=False,
+        **template_kwargs,
     )
     device = next(model.parameters()).device
     inputs = inputs.to(device)

configs/model.yaml

@@ -21,6 +21,7 @@ context:
 
 tool_calling:
   mode: prompt_json
+  schema_injection: prompt 
   
 outputs:
   save_raw_output: true