From 410441189ff7b642252d1815f49d48990fccff43 Mon Sep 17 00:00:00 2001
From: nightcityblade <nightcityblade@gmail.com>
Date: Fri, 14 Aug 2026 00:06:53 +0800
Subject: [PATCH] fix(bedrock): drop trailing empty Converse chunk

---
 litellm/llms/bedrock/chat/invoke_handler.py   |  5 +++-
 .../llms/bedrock/chat/test_invoke_handler.py  | 29 ++++++++++++++-----
 2 files changed, 26 insertions(+), 8 deletions(-)

diff --git a/litellm/llms/bedrock/chat/invoke_handler.py b/litellm/llms/bedrock/chat/invoke_handler.py
index 8d2b3dae71b..82232ccf42c 100644
--- a/litellm/llms/bedrock/chat/invoke_handler.py
+++ b/litellm/llms/bedrock/chat/invoke_handler.py
@@ -565,6 +565,9 @@ class AWSEventStreamDecoder:
             if "trace" in chunk_data:
                 trace: Final = chunk_data.get("trace")
                 model_response_provider_specific_fields["trace"] = trace
+            is_content_event: Final = any(
+                key in chunk_data for key in ("role", "start", "delta", "contentBlockIndex", "stopReason")
+            )
             response: Final = ModelResponseStream(
                 choices=[
                     StreamingChoices(
@@ -572,7 +575,7 @@ class AWSEventStreamDecoder:
                         index=0,  # Always 0 - Bedrock never returns multiple choices
                         delta=Delta(
                             content=text,
-                            role="assistant",
+                            role="assistant" if is_content_event else None,
                             tool_calls=[tool_use] if tool_use else None,
                             provider_specific_fields=(provider_specific_fields if provider_specific_fields else None),
                             thinking_blocks=thinking_blocks,
diff --git a/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py b/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py
index ee50b9db015..f0cf833b968 100644
--- a/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py
+++ b/tests/test_litellm/llms/bedrock/chat/test_invoke_handler.py
@@ -4,10 +4,9 @@ from unittest.mock import AsyncMock, MagicMock
 
 import pytest
 
-sys.path.insert(
-    0, os.path.abspath("../../../../..")
-)  # Adds the parent directory to the system path
+sys.path.insert(0, os.path.abspath("../../../../.."))  # Adds the parent directory to the system path
 
+from litellm.litellm_core_utils.model_response_utils import is_model_response_stream_empty
 from litellm.llms.bedrock.chat.invoke_handler import (
     AWSEventStreamDecoder,
     make_call,
@@ -203,9 +202,26 @@ def test_bedrock_converse_streaming_consistent_id():
     expected_id = f"chatcmpl-{native_conversation_id}"
 
     for response in parsed_responses:
-        assert (
-            response.id == expected_id
-        ), "All chunk IDs must match the one captured from the messageStart event"
+        assert response.id == expected_id, "All chunk IDs must match the one captured from the messageStart event"
+
+
+def test_converse_metadata_chunk_is_empty_after_usage_is_stripped():
+    decoder = AWSEventStreamDecoder(model="test")
+
+    content_chunk = decoder.converse_chunk_parser({"delta": {"text": "Hello"}})
+    assert content_chunk.choices[0].delta.role == "assistant"
+
+    metadata_chunk = decoder.converse_chunk_parser(
+        {
+            "usage": {"inputTokens": 10, "outputTokens": 5, "totalTokens": 15},
+            "metrics": {"latencyMs": 100},
+        }
+    )
+    assert metadata_chunk.usage is not None
+    assert metadata_chunk.choices[0].delta.role is None
+
+    stripped_chunk = metadata_chunk.model_copy(update={"usage": None})
+    assert is_model_response_stream_empty(stripped_chunk)
 
 
 @pytest.mark.asyncio
@@ -292,4 +308,3 @@ def test_make_sync_call_honors_explicit_stream_chunk_size():
     )
 
     response.iter_bytes.assert_called_once_with(chunk_size=2048)
-