fix: add llm drop_params parameter to LLMConfig (OpenHands#2471)

tobitege · SmartManoj · web-flow · commit d50a8447addc · 2024-07-24T16:25:36.000+02:00
* feat: add drop_params to LLMConfig

* Update opendevin/llm/llm.py

Fix use of unknown method.

Co-authored-by: மனோஜ்குமார் பழனிச்சாமி &lt;smartmanoj42857@gmail.com&gt;

---------

Co-authored-by: மனோஜ்குமார் பழனிச்சாமி &lt;smartmanoj42857@gmail.com&gt;
diff --git a/docs/modules/usage/llms/llms.md b/docs/modules/usage/llms/llms.md
@@ -15,13 +15,14 @@ OpenDevin will issue many prompts to the LLM you configure. Most of these LLMs c
 The `LLM_MODEL` environment variable controls which model is used in programmatic interactions.
 But when using the OpenDevin UI, you'll need to choose your model in the settings window.
 
-The following environment variables might be necessary for some LLMs:
+The following environment variables might be necessary for some LLMs/providers:
 
 - `LLM_API_KEY`
 - `LLM_BASE_URL`
 - `LLM_EMBEDDING_MODEL`
 - `LLM_EMBEDDING_DEPLOYMENT_NAME`
 - `LLM_API_VERSION`
+- `LLM_DROP_PARAMS`
 
 We have a few guides for running OpenDevin with specific model providers:
 
diff --git a/opendevin/core/config.py b/opendevin/core/config.py
@@ -49,6 +49,7 @@ class LLMConfig:
         input_cost_per_token: The cost per input token. This will available in logs for the user to check.
         output_cost_per_token: The cost per output token. This will available in logs for the user to check.
         ollama_base_url: The base URL for the OLLAMA API.
+        drop_params: Drop any unmapped (unsupported) params without causing an exception.
     """
 
     model: str = 'gpt-4o'
@@ -75,6 +76,7 @@ class LLMConfig:
     input_cost_per_token: float | None = None
     output_cost_per_token: float | None = None
     ollama_base_url: str | None = None
+    drop_params: bool | None = None
 
     def defaults_to_dict(self) -> dict:
         """Serialize fields to a dict for the frontend, including type hints, defaults, and whether it's optional."""
diff --git a/opendevin/core/schema/config.py b/opendevin/core/schema/config.py
@@ -4,6 +4,7 @@
 class ConfigType(str, Enum):
     # For frontend
     LLM_CUSTOM_LLM_PROVIDER = 'LLM_CUSTOM_LLM_PROVIDER'
+    LLM_DROP_PARAMS = 'LLM_DROP_PARAMS'
     LLM_MAX_INPUT_TOKENS = 'LLM_MAX_INPUT_TOKENS'
     LLM_MAX_OUTPUT_TOKENS = 'LLM_MAX_OUTPUT_TOKENS'
     LLM_TOP_P = 'LLM_TOP_P'
diff --git a/opendevin/llm/llm.py b/opendevin/llm/llm.py
@@ -52,18 +52,19 @@ def __init__(
         Args:
             config: The LLM configuration
         """
-
         self.config = copy.deepcopy(config)
         self.metrics = metrics if metrics is not None else Metrics()
         self.cost_metric_supported = True
 
         # litellm actually uses base Exception here for unknown model
         self.model_info = None
         try:
-            if not config.model.startswith('openrouter'):
-                self.model_info = litellm.get_model_info(config.model.split(':')[0])
+            if self.config.model.startswith('openrouter'):
+                self.model_info = litellm.get_model_info(self.config.model)
             else:
-                self.model_info = litellm.get_model_info(config.model)
+                self.model_info = litellm.get_model_info(
+                    self.config.model.split(':')[0]
+                )
         # noinspection PyBroadException
         except Exception:
             logger.warning(f'Could not get model info for {config.model}')
@@ -91,6 +92,9 @@ def __init__(
                 # Max output tokens for gpt3.5, so this is a safe fallback for any potentially viable model
                 self.config.max_output_tokens = 1024
 
+        if self.config.drop_params:
+            litellm.drop_params = self.config.drop_params
+
         self._completion = partial(
             litellm_completion,
             model=self.config.model,
