refactor(bedrock): centralize stream_chunk_size drop via typed control-param set

#30240 dropped stream_chunk_size with two hardcoded pops, one per invoke
transform entry point. Replace those literals with a single authoritative
LiteLLMControlParams type whose field names drive the filter both entry
points already apply for aws_authentication_params, so the set of
LiteLLM-internal control params lives in one place and the strip happens
pre-nesting (covering providers like Titan that nest inference params under
textGenerationConfig, which a post-serialization top-level strip would miss).

The regression test now asserts on the actual signed wire body that both the
base and Anthropic-override transforms funnel through, across four
provider/model combinations; all four fail without the filter.

litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py

@@ -24,6 +24,7 @@ from litellm.llms.bedrock.common_utils import (
     remove_custom_field_from_tools,
 )
 from litellm.types.llms.anthropic import ANTHROPIC_TOOL_SEARCH_BETA_HEADER
+from litellm.types.llms.bedrock import LITELLM_CONTROL_PARAM_KEYS
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import ModelResponse
 from litellm.utils import _supports_factory
@@ -194,6 +195,7 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
             k: v
             for k, v in optional_params.items()
             if k not in self.aws_authentication_params
+            and k not in LITELLM_CONTROL_PARAM_KEYS
         }
         output_config = filtered_params.get("output_config")
         if isinstance(output_config, dict):
@@ -215,7 +217,6 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig):
 
         anthropic_request.pop("model", None)
         anthropic_request.pop("stream", None)
-        anthropic_request.pop("stream_chunk_size", None)
         output_format = anthropic_request.pop("output_format", None)
         output_config_format = pop_bedrock_invoke_output_config_format(
             anthropic_request

litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py

@@ -24,6 +24,7 @@ from litellm.llms.custom_httpx.http_handler import (
     HTTPHandler,
     _get_httpx_client,
 )
+from litellm.types.llms.bedrock import LITELLM_CONTROL_PARAM_KEYS
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import ModelResponse, Usage
 from litellm.utils import CustomStreamWrapper
@@ -150,7 +151,6 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM):
     ) -> dict:
         ## SETUP ##
         stream = optional_params.pop("stream", None)
-        optional_params.pop("stream_chunk_size", None)
         custom_prompt_dict: dict = litellm_params.pop("custom_prompt_dict", None) or {}
         hf_model_name = litellm_params.get("hf_model_name", None)
 
@@ -167,6 +167,7 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM):
             k: v
             for k, v in inference_params.items()
             if k not in self.aws_authentication_params
+            and k not in LITELLM_CONTROL_PARAM_KEYS
         }
         request_data: dict = {}
         if provider == "cohere":

litellm/types/llms/bedrock.py

@@ -1,11 +1,24 @@
 import json
 from typing import Any, Dict, List, Literal, Optional, Union
 
+from pydantic import BaseModel
 from typing_extensions import TYPE_CHECKING, Required, TypedDict, override
 
 from .openai import ChatCompletionToolCallChunk
 
 
+class LiteLLMControlParams(BaseModel):
+    """LiteLLM-internal control parameters that must never be serialized into a
+    provider request body. They live in optional_params for convenience but
+    govern client-side behavior (e.g. how the HTTP response stream is
+    re-chunked), so Bedrock rejects them as unknown fields."""
+
+    stream_chunk_size: Optional[int] = None
+
+
+LITELLM_CONTROL_PARAM_KEYS = frozenset(LiteLLMControlParams.model_fields)
+
+
 class CachePointBlock(TypedDict, total=False):
     type: Literal["default"]
     ttl: str

tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_base_invoke_transformation.py

@@ -1,12 +1,9 @@
-import json
 import os
 import sys
 
 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.llms.bedrock.chat.invoke_transformations.anthropic_claude3_transformation import (
     AmazonAnthropicClaudeConfig,
@@ -25,12 +22,15 @@ from litellm.llms.bedrock.chat.invoke_transformations.base_invoke_transformation
         (AmazonAnthropicClaudeConfig, "anthropic.claude-sonnet-4-6"),
     ],
 )
-def test_transform_request_drops_stream_chunk_size(config, model):
+def test_signed_invoke_body_drops_stream_chunk_size(config, model):
     """stream_chunk_size is a LiteLLM-internal knob for re-chunking the HTTP
     response stream. Leaking it into the provider request body makes Bedrock
     reject the whole request: ValidationException 'stream_chunk_size: Extra
-    inputs are not permitted'."""
-    request_body = config().transform_request(
+    inputs are not permitted'. The two invoke transform entry points build the
+    body differently, so this asserts on the actual signed wire bytes that both
+    funnel through, regardless of which transform produced them."""
+    cfg = config()
+    request_body = cfg.transform_request(
         model=model,
         messages=[{"role": "user", "content": "hi"}],
         optional_params={"stream": True, "stream_chunk_size": 2048, "max_tokens": 10},
@@ -38,4 +38,16 @@ def test_transform_request_drops_stream_chunk_size(config, model):
         headers={},
     )
 
-    assert "stream_chunk_size" not in json.dumps(request_body)
+    _, signed_body = cfg.sign_request(
+        headers={},
+        optional_params={},
+        request_data=request_body,
+        api_base="https://bedrock-runtime.us-east-1.amazonaws.com/model/{}/invoke".format(model),
+        api_key="test-bearer-token",
+        model=model,
+        stream=True,
+    )
+
+    assert signed_body is not None
+    assert "stream_chunk_size" not in signed_body.decode()
+    assert "max_tokens" in signed_body.decode()