Enabled native function calling for O1 + added support for reasoning_effort config in the config. (#6256)

Co-authored-by: Engel Nyst <enyst@users.noreply.github.com>

config.template.toml

@@ -23,6 +23,9 @@ workspace_base = "./workspace"
 # Cache directory path
 #cache_dir = "/tmp/cache"
 
+# Reasoning effort for o1 models (low, medium, high, or not set)
+#reasoning_effort = "medium"
+
 # Debugging enabled
 #debug = false
 

openhands/core/config/llm_config.py

@@ -40,6 +40,7 @@ class LLMConfig:
         drop_params: Drop any unmapped (unsupported) params without causing an exception.
         modify_params: Modify params allows litellm to do transformations like adding a default message, when a message is empty.
         disable_vision: If model is vision capable, this option allows to disable image processing (useful for cost reduction).
+        reasoning_effort: The effort to put into reasoning. This is a string that can be one of 'low', 'medium', 'high', or 'none'. Exclusive for o1 models.
         caching_prompt: Use the prompt caching feature if provided by the LLM and supported by the provider.
         log_completions: Whether to log LLM completions to the state.
         log_completions_folder: The folder to log LLM completions to. Required if log_completions is True.
@@ -79,6 +80,7 @@ class LLMConfig:
     # Note: this setting is actually global, unlike drop_params
     modify_params: bool = True
     disable_vision: bool | None = None
+    reasoning_effort: str | None = None
     caching_prompt: bool = True
     log_completions: bool = False
     log_completions_folder: str = os.path.join(LOG_DIR, 'completions')

openhands/llm/async_llm.py

@@ -6,7 +6,11 @@ from litellm import acompletion as litellm_acompletion
 
 from openhands.core.exceptions import UserCancelledError
 from openhands.core.logger import openhands_logger as logger
-from openhands.llm.llm import LLM, LLM_RETRY_EXCEPTIONS
+from openhands.llm.llm import (
+    LLM,
+    LLM_RETRY_EXCEPTIONS,
+    REASONING_EFFORT_SUPPORTED_MODELS,
+)
 from openhands.utils.shutdown_listener import should_continue
 
 
@@ -55,6 +59,10 @@ class AsyncLLM(LLM):
             elif 'messages' in kwargs:
                 messages = kwargs['messages']
 
+            # Set reasoning effort for models that support it
+            if self.config.model.lower() in REASONING_EFFORT_SUPPORTED_MODELS:
+                kwargs['reasoning_effort'] = self.config.reasoning_effort
+
             # ensure we work with a list of messages
             messages = messages if isinstance(messages, list) else [messages]
 

openhands/llm/llm.py

@@ -70,7 +70,15 @@ FUNCTION_CALLING_SUPPORTED_MODELS = [
     'claude-3.5-haiku',
     'claude-3-5-haiku-20241022',
     'gpt-4o-mini',
-    'gpt-4o',
+    'o1-2024-12-17',
+]
+
+REASONING_EFFORT_SUPPORTED_MODELS = [
+    'o1-2024-12-17',
+]
+
+MODELS_WITHOUT_STOP_WORDS = [
+    'o1-mini',
 ]
 
 
@@ -186,7 +194,8 @@ class LLM(RetryMixin, DebugMixin):
                     messages, kwargs['tools']
                 )
                 kwargs['messages'] = messages
-                kwargs['stop'] = STOP_WORDS
+                if self.config.model not in MODELS_WITHOUT_STOP_WORDS:
+                    kwargs['stop'] = STOP_WORDS
                 mock_fncall_tools = kwargs.pop('tools')
 
             # if we have no messages, something went very wrong
@@ -205,6 +214,10 @@ class LLM(RetryMixin, DebugMixin):
                         'anthropic-beta': 'prompt-caching-2024-07-31',
                     }
 
+            # Set reasoning effort for models that support it
+            if self.config.model.lower() in REASONING_EFFORT_SUPPORTED_MODELS:
+                kwargs['reasoning_effort'] = self.config.reasoning_effort
+
             # set litellm modify_params to the configured value
             # True by default to allow litellm to do transformations like adding a default message, when a message is empty
             # NOTE: this setting is global; unlike drop_params, it cannot be overridden in the litellm completion partial
@@ -213,7 +226,6 @@ class LLM(RetryMixin, DebugMixin):
             try:
                 # Record start time for latency measurement
                 start_time = time.time()
-
                 # we don't support streaming here, thus we get a ModelResponse
                 resp: ModelResponse = self._completion_unwrapped(*args, **kwargs)
 

openhands/llm/streaming_llm.py

@@ -5,6 +5,7 @@ from typing import Any
 from openhands.core.exceptions import UserCancelledError
 from openhands.core.logger import openhands_logger as logger
 from openhands.llm.async_llm import LLM_RETRY_EXCEPTIONS, AsyncLLM
+from openhands.llm.llm import REASONING_EFFORT_SUPPORTED_MODELS
 
 
 class StreamingLLM(AsyncLLM):
@@ -61,6 +62,10 @@ class StreamingLLM(AsyncLLM):
                     'The messages list is empty. At least one message is required.'
                 )
 
+            # Set reasoning effort for models that support it
+            if self.config.model.lower() in REASONING_EFFORT_SUPPORTED_MODELS:
+                kwargs['reasoning_effort'] = self.config.reasoning_effort
+
             self.log_prompt(messages)
 
             try: