Merge 03990d1570 into 44d84360fb

litellm/llms/bedrock_mantle/chat/handler.py

@@ -0,0 +1,32 @@
+"""
+Streaming handler for Amazon Bedrock Mantle (OpenAI-compatible API).
+
+Bedrock Mantle emits a fresh ``id`` on every SSE chunk, which violates the
+OpenAI streaming contract that all chunks in one response share a single
+``id``. Clients that validate this (e.g. the openai-go SDK's
+``ChatCompletionAccumulator``) drop every chunk after the first, losing
+content and tool-call arguments. We pin the id from the first chunk that
+carries one and reuse it for the rest of the stream.
+"""
+
+from typing import Optional
+
+from litellm.llms.openai.chat.gpt_transformation import (
+    OpenAIChatCompletionStreamingHandler,
+)
+from litellm.types.utils import ModelResponseStream
+
+
+class BedrockMantleChatCompletionStreamingHandler(OpenAIChatCompletionStreamingHandler):
+    def __init__(self, streaming_response, sync_stream: bool, json_mode: Optional[bool] = False):
+        super().__init__(streaming_response=streaming_response, sync_stream=sync_stream, json_mode=json_mode)
+        self._response_id: Optional[str] = None
+
+    def chunk_parser(self, chunk: dict) -> ModelResponseStream:
+        parsed = super().chunk_parser(chunk)
+        raw_id = chunk.get("id")
+        if self._response_id is None and raw_id:
+            self._response_id = raw_id
+        if self._response_id is not None:
+            parsed.id = self._response_id
+        return parsed

litellm/llms/bedrock_mantle/chat/transformation.py

@@ -111,11 +111,11 @@ class BedrockMantleChatConfig(BedrockMantleAuthMixin, OpenAILikeChatConfig):
         sync_stream: bool,
         json_mode: bool | None = False,
     ) -> Any:
-        from litellm.llms.openai.chat.gpt_transformation import (
-            OpenAIChatCompletionStreamingHandler,
+        from litellm.llms.bedrock_mantle.chat.handler import (
+            BedrockMantleChatCompletionStreamingHandler,
         )
 
-        return OpenAIChatCompletionStreamingHandler(
+        return BedrockMantleChatCompletionStreamingHandler(
             streaming_response=streaming_response,
             sync_stream=sync_stream,
             json_mode=json_mode,

tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_transformation.py

@@ -666,6 +666,56 @@ def test_gemma_4_bedrock_mantle_model_metadata(
     )
 
 
+def test_streaming_handler_pins_first_chunk_id():
+    """Bedrock Mantle emits a fresh id on every SSE chunk; the OpenAI contract
+    requires all chunks in a response to share one id, so the handler must pin
+    the first chunk's id onto every subsequent chunk."""
+    from litellm.llms.bedrock_mantle.chat.handler import (
+        BedrockMantleChatCompletionStreamingHandler,
+    )
+
+    handler = BedrockMantleChatCompletionStreamingHandler(
+        streaming_response=iter([]), sync_stream=True
+    )
+
+    raw_chunks = [
+        {
+            "id": f"chatcmpl-{i}",
+            "created": 1,
+            "model": "openai.gpt-5.5",
+            "choices": [{"index": 0, "delta": {"content": token}}],
+        }
+        for i, token in enumerate(["Hel", "lo", "!"])
+    ]
+
+    parsed_ids = [handler.chunk_parser(chunk).id for chunk in raw_chunks]
+
+    assert parsed_ids == ["chatcmpl-0", "chatcmpl-0", "chatcmpl-0"]
+
+
+def test_streaming_handler_pins_id_when_first_chunk_lacks_one():
+    """Some backends omit id on an initial role-only chunk; the handler should
+    pin the first id it actually sees and keep it stable afterwards."""
+    from litellm.llms.bedrock_mantle.chat.handler import (
+        BedrockMantleChatCompletionStreamingHandler,
+    )
+
+    handler = BedrockMantleChatCompletionStreamingHandler(
+        streaming_response=iter([]), sync_stream=True
+    )
+
+    raw_chunks = [
+        {"created": 1, "model": "openai.gpt-5.5", "choices": [{"index": 0, "delta": {"role": "assistant"}}]},
+        {"id": "chatcmpl-real", "created": 1, "model": "openai.gpt-5.5", "choices": [{"index": 0, "delta": {"content": "Hi"}}]},
+        {"id": "chatcmpl-other", "created": 1, "model": "openai.gpt-5.5", "choices": [{"index": 0, "delta": {"content": "!"}}]},
+    ]
+
+    parsed_ids = [handler.chunk_parser(chunk).id for chunk in raw_chunks]
+
+    assert parsed_ids[1] == "chatcmpl-real"
+    assert parsed_ids[2] == "chatcmpl-real"
+
+
 @pytest.mark.parametrize(
     "model_id",
     [