From da1c77be2c0f80dbc99e29d117927cb724ae29e4 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Sat, 26 Sep 2026 17:43:22 +0000
Subject: [PATCH] fix(bedrock): keep native chat completions when only top_k is
 set

Converse only forwards top_k for Anthropic and Nova, so treating it as a
converse-only key sent gpt-oss, GPT-5.6, and Grok off the native route and
still dropped the value
---
 litellm/llms/bedrock/common_utils.py          |  5 ++--
 ...bedrock_chat_completions_transformation.py | 28 +++++++++++++++----
 2 files changed, 24 insertions(+), 9 deletions(-)

diff --git a/litellm/llms/bedrock/common_utils.py b/litellm/llms/bedrock/common_utils.py
index 098be7082d3..cb7cfdf9640 100644
--- a/litellm/llms/bedrock/common_utils.py
+++ b/litellm/llms/bedrock/common_utils.py
@@ -898,7 +898,6 @@ BEDROCK_CONVERSE_ONLY_REQUEST_KEYS: Final = frozenset(
         "outputConfig",
         "thinking",
         "additionalModelRequestFields",
-        "top_k",
     )
 )
 
@@ -919,8 +918,8 @@ def bedrock_request_needs_converse(model: str, request_params: Mapping[str, obje
     """Whether a request on a runtime-Chat-Completions model must still be served by Converse.
 
     Converse-shaped body keys (``BEDROCK_CONVERSE_ONLY_REQUEST_KEYS``, the Anthropic-style ``thinking``
-    block and the ``additionalModelRequestFields`` / ``top_k`` extension params included, which only Converse
-    forwards as ``additionalModelRequestFields`` and ``inferenceConfig``) have no field on
+    block and ``additionalModelRequestFields`` included, which only Converse
+    forwards as ``additionalModelRequestFields``) have no field on
     AWS's native OpenAI surface, operator-owned request metadata is only written onto the Converse body,
     function tools (``tools`` or legacy ``functions``) on a model without
     ``supports_bedrock_runtime_chat_completions_tools_with_reasoning`` are rejected there unless
diff --git a/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py b/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py
index 17c0f8d3c1d..3f1499cda8b 100644
--- a/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py
+++ b/tests/unit/llms/bedrock/chat/chat_completions/test_bedrock_chat_completions_transformation.py
@@ -259,17 +259,33 @@ def test_guardrail_config_falls_back_to_converse(local_cost_map, model):
 
 
 @pytest.mark.parametrize("model", ["openai.gpt-oss-20b-1:0", "us.xai.grok-4.6"])
-@pytest.mark.parametrize(
-    "request_params",
-    [{"additionalModelRequestFields": {"reasoning_effort": "high"}}, {"top_k": 40}],
-    ids=["additionalModelRequestFields", "top_k"],
-)
-def test_converse_extension_params_fall_back_to_converse(local_cost_map, model, request_params):
+def test_additional_model_request_fields_fall_back_to_converse(local_cost_map, model):
+    request_params = {"additionalModelRequestFields": {"reasoning_effort": "high"}}
     assert bedrock_request_needs_converse(model, request_params) is True
     assert BedrockModelInfo.get_bedrock_route(model, request_params) == "converse"
     assert BedrockModelInfo.get_bedrock_route(model, {key: None for key in request_params}) == "chat_completions"
 
 
+@pytest.mark.parametrize("model", ["openai.gpt-oss-20b-1:0", "us.xai.grok-4.6", "global.openai.gpt-5.6-sol"])
+def test_top_k_stays_on_chat_completions(local_cost_map, fake_aws_env, model):
+    request_params = {"top_k": 40}
+    assert bedrock_request_needs_converse(model, request_params) is False
+    assert BedrockModelInfo.get_bedrock_route(model, request_params) == "chat_completions"
+
+    requests, client = _recording_client(json=_chat_completion_json("ok", model))
+    litellm.completion(
+        model=f"bedrock/{model}",
+        messages=[{"role": "user", "content": "hello"}],
+        top_k=40,
+        client=client,
+    )
+
+    assert str(requests[0].url) == "https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1/chat/completions"
+    body = json.loads(requests[0].content)
+    assert body["top_k"] == 40
+    assert "inferenceConfig" not in body
+
+
 @pytest.mark.parametrize(
     "request_params, expected_route",
     [