Merge 26f563f47e into 44d84360fb

litellm/llms/bedrock_mantle/chat/streaming_handler.py

@@ -0,0 +1,37 @@
+"""Streaming handler for the Amazon Bedrock Mantle OpenAI-compatible surface.
+
+Mantle emits streaming ``tool_calls[].index`` values that start at 1 for the
+first tool call, whereas the OpenAI spec (and every client that aggregates the
+deltas, including the OpenAI SDK) expects a 0-based index. This handler shifts
+the provider's indices down so the first tool call reported in a stream is
+index 0, leaving already-0-based streams untouched.
+"""
+
+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, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        # Offset of the first tool_call index seen in the stream; every index is
+        # rebased against it so the first tool call becomes index 0.
+        self._tool_call_index_offset: Optional[int] = None
+
+    def chunk_parser(self, chunk: dict) -> ModelResponseStream:
+        parsed = super().chunk_parser(chunk)
+        for choice in parsed.choices:
+            tool_calls = choice.delta.tool_calls
+            if not tool_calls:
+                continue
+            offset = self._tool_call_index_offset
+            if offset is None:
+                offset = min(tool_call.index for tool_call in tool_calls)
+                self._tool_call_index_offset = offset
+            for tool_call in tool_calls:
+                tool_call.index -= offset
+        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.streaming_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 +666,85 @@ def test_gemma_4_bedrock_mantle_model_metadata(
     )
 
 
+def _mantle_tool_call_chunk(index: int, *, tool_id: str, name: str, args: str) -> str:
+    return "data: " + json.dumps(
+        {
+            "id": "chatcmpl-test",
+            "object": "chat.completion.chunk",
+            "created": 1733529600,
+            "model": "openai.gpt-5.5",
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {
+                        "tool_calls": [
+                            {
+                                "index": index,
+                                "id": tool_id,
+                                "type": "function",
+                                "function": {"name": name, "arguments": args},
+                            }
+                        ]
+                    },
+                    "finish_reason": None,
+                }
+            ],
+        }
+    )
+
+
+def _collect_tool_call_indices(chunks):
+    from litellm.llms.bedrock_mantle.chat.streaming_handler import (
+        BedrockMantleChatCompletionStreamingHandler,
+    )
+
+    from litellm.types.utils import ModelResponseStream
+
+    handler = BedrockMantleChatCompletionStreamingHandler(
+        streaming_response=iter(chunks),
+        sync_stream=True,
+    )
+    return [
+        tool_call.index
+        for parsed in handler
+        if isinstance(parsed, ModelResponseStream)
+        for choice in parsed.choices
+        for tool_call in (choice.delta.tool_calls or [])
+    ]
+
+
+class TestBedrockMantleStreamingToolCallIndex:
+    """Regression tests for https://github.com/BerriAI/litellm/issues/32759
+
+    Mantle streams the first tool call with index 1; the OpenAI spec (and any
+    client aggregating the deltas) expects the first index to be 0.
+    """
+
+    def test_first_tool_call_index_rebased_to_zero(self):
+        chunks = [
+            _mantle_tool_call_chunk(1, tool_id="call_a", name="get_weather", args=""),
+            _mantle_tool_call_chunk(1, tool_id="call_a", name="get_weather", args='{"city": "NYC"}'),
+            "data: [DONE]",
+        ]
+        assert _collect_tool_call_indices(chunks) == [0, 0]
+
+    def test_parallel_tool_call_indices_rebased_contiguously(self):
+        chunks = [
+            _mantle_tool_call_chunk(1, tool_id="call_a", name="get_weather", args="{}"),
+            _mantle_tool_call_chunk(2, tool_id="call_b", name="get_time", args="{}"),
+            "data: [DONE]",
+        ]
+        assert _collect_tool_call_indices(chunks) == [0, 1]
+
+    def test_already_zero_based_stream_is_untouched(self):
+        chunks = [
+            _mantle_tool_call_chunk(0, tool_id="call_a", name="get_weather", args="{}"),
+            _mantle_tool_call_chunk(1, tool_id="call_b", name="get_time", args="{}"),
+            "data: [DONE]",
+        ]
+        assert _collect_tool_call_indices(chunks) == [0, 1]
+
+
 @pytest.mark.parametrize(
     "model_id",
     [