[Frontend] Fixes anthropic streaming message_start usage nesting (#30266)

Signed-off-by: bbartels <benjamin@bartels.dev>

tests/entrypoints/openai/test_messages.py

@@ -79,9 +79,12 @@ async def test_anthropic_streaming(client: anthropic.AsyncAnthropic):
 
     assert chunk_count > 0
     assert first_chunk is not None, "message_start chunk was never observed"
-    assert first_chunk.usage is not None, "first chunk should include usage stats"
-    assert first_chunk.usage["output_tokens"] == 0
-    assert first_chunk.usage["input_tokens"] > 5
+    assert first_chunk.message is not None, "first chunk should include message"
+    assert first_chunk.message.usage is not None, (
+        "first chunk should include usage stats"
+    )
+    assert first_chunk.message.usage.output_tokens == 0
+    assert first_chunk.message.usage.input_tokens > 5
 
 
 @pytest.mark.asyncio

vllm/entrypoints/anthropic/serving_messages.py

@@ -324,13 +324,13 @@ class AnthropicServingMessages(OpenAIServingChat):
                                     id=origin_chunk.id,
                                     content=[],
                                     model=origin_chunk.model,
-                                ),
                                     usage=AnthropicUsage(
                                         input_tokens=origin_chunk.usage.prompt_tokens
                                         if origin_chunk.usage
                                         else 0,
                                         output_tokens=0,
                                     ),
+                                ),
                             )
                             first_item = False
                             data = chunk.model_dump_json(exclude_unset=True)