Revert "Simplify max_output_tokens handling in LLM classes (#9296)"

This reverts commit 1e33624.

Author: enyst

openhands/llm/llm.py

@@ -163,7 +163,6 @@ def __init__(
             'temperature': self.config.temperature,
             'max_completion_tokens': self.config.max_output_tokens,
         }
-
         if self.config.top_k is not None:
             # openai doesn't expose top_k
             # litellm will handle it a bit differently than the openai-compatible params
@@ -493,6 +492,26 @@ def init_model_info(self) -> None:
                 # Safe fallback for any potentially viable model
                 self.config.max_input_tokens = 4096
 
+        if self.config.max_output_tokens is None:
+            # Safe default for any potentially viable model
+            self.config.max_output_tokens = 4096
+            if self.model_info is not None:
+                # max_output_tokens has precedence over max_tokens, if either exists.
+                # litellm has models with both, one or none of these 2 parameters!
+                if 'max_output_tokens' in self.model_info and isinstance(
+                    self.model_info['max_output_tokens'], int
+                ):
+                    self.config.max_output_tokens = self.model_info['max_output_tokens']
+                elif 'max_tokens' in self.model_info and isinstance(
+                    self.model_info['max_tokens'], int
+                ):
+                    self.config.max_output_tokens = self.model_info['max_tokens']
+            if any(
+                model in self.config.model
+                for model in ['claude-3-7-sonnet', 'claude-3.7-sonnet']
+            ):
+                self.config.max_output_tokens = 64000  # litellm set max to 128k, but that requires a header to be set
+
         # Initialize function calling capability
         # Check if model name is in our supported list
         model_name_supported = (

tests/unit/test_llm.py

@@ -132,7 +132,7 @@ def test_llm_init_with_model_info(mock_get_model_info, default_config):
     llm = LLM(default_config)
     llm.init_model_info()
     assert llm.config.max_input_tokens == 8000
-    assert llm.config.max_output_tokens is None
+    assert llm.config.max_output_tokens == 2000
 
 
 @patch('openhands.llm.llm.litellm.get_model_info')
@@ -141,7 +141,7 @@ def test_llm_init_without_model_info(mock_get_model_info, default_config):
     llm = LLM(default_config)
     llm.init_model_info()
     assert llm.config.max_input_tokens == 4096
-    assert llm.config.max_output_tokens is None
+    assert llm.config.max_output_tokens == 4096
 
 
 def test_llm_init_with_custom_config():
@@ -260,7 +260,7 @@ def test_llm_init_with_openrouter_model(mock_get_model_info, default_config):
     llm = LLM(default_config)
     llm.init_model_info()
     assert llm.config.max_input_tokens == 7000
-    assert llm.config.max_output_tokens is None
+    assert llm.config.max_output_tokens == 1500
     mock_get_model_info.assert_called_once_with('openrouter:gpt-4o-mini')