Tool choice required

debug_gym/agents/utils.py

@@ -6,7 +6,6 @@
 
 import yaml
 
-from debug_gym.agents.base_agent import AGENT_REGISTRY
 from debug_gym.logger import DebugGymLogger
 
 

debug_gym/llms/base.py

@@ -222,17 +222,16 @@ def instantiate(
         name: str | None = None,
         llm_config_file_path: str | None = None,
         logger: DebugGymLogger | None = None,
-        temperature: float | None = None,
-        max_tokens: int | None = None,
+        **runtime_generate_kwargs,
     ) -> "LLM":
         """Creates an instance of the appropriate LLM class based on the configuration.
 
         Args:
             name: Name of the LLM model (e.g., "gpt-4o", "claude-3.7").
             llm_config_file_path: Optional path to the LLM configuration file.
             logger: Optional DebugGymLogger for logging.
-            temperature: Optional temperature for generation.
-            max_tokens: Optional max tokens for generation.
+            **runtime_generate_kwargs: Additional generation kwargs to pass to the LLM.
+                e.g. temperature, max_tokens, tool_choice ("auto", "required", "none")
 
         Returns:
             An instance of the appropriate LLM class.
@@ -243,13 +242,6 @@ def instantiate(
         if not name:
             return None
 
-        # Build runtime generation kwargs from explicit args
-        runtime_generate_kwargs = {}
-        if temperature is not None:
-            runtime_generate_kwargs["temperature"] = temperature
-        if max_tokens is not None:
-            runtime_generate_kwargs["max_tokens"] = max_tokens
-
         if name == "human":
             from debug_gym.llms import Human
 

debug_gym/llms/huggingface.py

@@ -2,6 +2,7 @@
 
 from transformers import AutoTokenizer
 
+from debug_gym.llms.base import LLMResponse
 from debug_gym.llms.openai import OpenAILLM
 
 
@@ -71,3 +72,12 @@ def tokenize(self, messages: list[dict]) -> list[list[str]]:
                 tokens = tokenizer.tokenize(content)
                 result.append(tokens)
             return result
+
+    def generate(
+        self, messages, tools, tool_choice="required", **kwargs
+    ) -> LLMResponse:
+        """Override default tool_choice parameter to "required"."""
+        llm_response = super().generate(
+            messages, tools, tool_choice=tool_choice, **kwargs
+        )
+        return llm_response

debug_gym/llms/openai.py

@@ -283,7 +283,7 @@ def convert_observation_to_message(
                 "content": filter_non_utf8(observation),
             }
 
-    def generate(self, messages, tools, **kwargs) -> LLMResponse:
+    def generate(self, messages, tools, tool_choice="auto", **kwargs) -> LLMResponse:
         # set max tokens if not provided
         kwargs["max_tokens"] = kwargs.get("max_tokens", NOT_GIVEN)
         api_call = retry_on_exception(
@@ -296,7 +296,7 @@ def generate(self, messages, tools, **kwargs) -> LLMResponse:
                     model=self.config.model,
                     messages=messages,
                     tools=self.define_tools(tools),
-                    tool_choice="auto",
+                    tool_choice=tool_choice,
                     **kwargs,
                 )
             else:

tests/llms/test_base.py

@@ -103,10 +103,10 @@ def test_instantiate_llm(mock_open, logger_mock):
     assert llm.runtime_generate_kwargs == {"temperature": 0.5, "max_tokens": 1000}
 
     # Test with **kwargs unpacking (like config)
-    llm_config = {"name": "gpt-4o-mini", "temperature": 0.7}
+    llm_config = {"name": "gpt-4o-mini", "temperature": 0.7, "tool_choice": "auto"}
     llm = LLM.instantiate(**llm_config, logger=logger_mock)
     assert isinstance(llm, OpenAILLM)
-    assert llm.runtime_generate_kwargs == {"temperature": 0.7}
+    assert llm.runtime_generate_kwargs == {"temperature": 0.7, "tool_choice": "auto"}
 
 
 class Tool1(EnvironmentTool):