[LLM Translation] Fix Model Usage not having text tokens (#13234)

* fix + test

* remove test comments

* fix mypy

* fix mypy

* fix tests

litellm/cost_calculator.py

@@ -1265,7 +1265,7 @@ class BaseTokenUsageProcessor:
         Combine multiple Usage objects into a single Usage object, checking model keys for nested values.
         """
         from litellm.types.utils import (
-            CompletionTokensDetails,
+            CompletionTokensDetailsWrapper,
             PromptTokensDetailsWrapper,
             Usage,
         )
@@ -1320,7 +1320,7 @@ class BaseTokenUsageProcessor:
                     not hasattr(combined, "completion_tokens_details")
                     or not combined.completion_tokens_details
                 ):
-                    combined.completion_tokens_details = CompletionTokensDetails()
+                    combined.completion_tokens_details = CompletionTokensDetailsWrapper()
 
                 # Check what keys exist in the model's completion_tokens_details
                 for attr in usage.completion_tokens_details.model_fields:

litellm/litellm_core_utils/streaming_chunk_builder_utils.py

@@ -527,7 +527,12 @@ class ChunkProcessor:
                 returned_usage, "cache_read_input_tokens", cache_read_input_tokens
             )  # for anthropic
         if completion_tokens_details is not None:
-            returned_usage.completion_tokens_details = completion_tokens_details
+            if isinstance(completion_tokens_details, CompletionTokensDetails):
+                returned_usage.completion_tokens_details = CompletionTokensDetailsWrapper(
+                    **completion_tokens_details.model_dump()
+                )
+            else:
+                returned_usage.completion_tokens_details = completion_tokens_details
 
         if reasoning_tokens is not None:
             if returned_usage.completion_tokens_details is None:

litellm/types/utils.py

@@ -888,6 +888,9 @@ class Usage(CompletionUsage):
 
     server_tool_use: Optional[ServerToolUse] = None
 
+    completion_tokens_details: Optional[CompletionTokensDetailsWrapper] = None
+    """Breakdown of tokens used in a completion."""
+
     prompt_tokens_details: Optional[PromptTokensDetailsWrapper] = None
     """Breakdown of tokens used in the prompt."""
 

tests/local_testing/test_caching.py

@@ -2155,8 +2155,8 @@ async def test_caching_kwargs_input(sync_mode):
         Message,
         ModelResponse,
         Usage,
-        CompletionTokensDetails,
-        PromptTokensDetails,
+        CompletionTokensDetailsWrapper,
+        PromptTokensDetailsWrapper,
     )
     from datetime import datetime
 
@@ -2187,10 +2187,10 @@ async def test_caching_kwargs_input(sync_mode):
                 completion_tokens=31,
                 prompt_tokens=16,
                 total_tokens=47,
-                completion_tokens_details=CompletionTokensDetails(
+                completion_tokens_details=CompletionTokensDetailsWrapper(
                     audio_tokens=None, reasoning_tokens=0
                 ),
-                prompt_tokens_details=PromptTokensDetails(
+                prompt_tokens_details=PromptTokensDetailsWrapper(
                     audio_tokens=None, cached_tokens=0
                 ),
             ),

tests/test_litellm/types/test_types_utils.py

@@ -73,3 +73,56 @@ def test_usage_dump():
 
     new_usage = Usage(**current_usage.model_dump())
     assert new_usage.prompt_tokens_details.web_search_requests == 1
+
+
+def test_usage_completion_tokens_details_text_tokens():
+    from litellm.types.utils import Usage
+
+    # Test data from the reported issue
+    usage_data = {
+        'completion_tokens': 77,
+        'prompt_tokens': 11937,
+        'total_tokens': 12014,
+        'completion_tokens_details': {
+            'accepted_prediction_tokens': None,
+            'audio_tokens': None,
+            'reasoning_tokens': 65,
+            'rejected_prediction_tokens': None,
+            'text_tokens': 12
+        },
+        'prompt_tokens_details': {
+            'audio_tokens': None,
+            'cached_tokens': None,
+            'text_tokens': 11937,
+            'image_tokens': None
+        }
+    }
+
+    # Create Usage object
+    u = Usage(**usage_data)
+    
+    # Verify the object has the text_tokens field
+    assert hasattr(u.completion_tokens_details, 'text_tokens')
+    assert u.completion_tokens_details.text_tokens == 12
+    
+    # Get model_dump output
+    dump_result = u.model_dump()
+    
+    # Verify text_tokens is present in the model_dump output
+    assert 'completion_tokens_details' in dump_result
+    assert 'text_tokens' in dump_result['completion_tokens_details']
+    assert dump_result['completion_tokens_details']['text_tokens'] == 12
+    
+    # Verify the full completion_tokens_details structure
+    expected_completion_details = {
+        'accepted_prediction_tokens': None,
+        'audio_tokens': None,
+        'reasoning_tokens': 65,
+        'rejected_prediction_tokens': None,
+        'text_tokens': 12
+    }
+    assert dump_result['completion_tokens_details'] == expected_completion_details
+    
+    # Verify round-trip serialization works
+    new_usage = Usage(**dump_result)
+    assert new_usage.completion_tokens_details.text_tokens == 12