Merge remote-tracking branch 'origin/lts/1.108.x' into litellm_backport_bedrock_converse_s3_sources_lts_1_108_x_w716

litellm/llms/anthropic/pass_through/messages/handler.py

@@ -11,7 +11,7 @@ from collections.abc import AsyncIterator, Coroutine, Iterator, Mapping, Sequenc
 from functools import partial
 from typing import Final, cast
 
-from pydantic import TypeAdapter
+from pydantic import TypeAdapter, ValidationError
 
 import litellm
 from litellm.integrations.anthropic_cache_control_hook import AnthropicCacheControlHook
@@ -56,6 +56,15 @@ __all__ = ("anthropic_messages", "anthropic_messages_handler")
 _RESPONSES_API_PROVIDERS: Final = frozenset({"openai"})
 _SAFEGUARD_KWARGS: Final = TypeAdapter(Mapping[str, object])
 _SAFEGUARD_MESSAGES: Final = TypeAdapter(Sequence[Mapping[str, object]])
+_DROP_PATHS: Final = TypeAdapter(list[object])
+
+
+def additional_drop_paths(kwargs: Mapping[str, object]) -> tuple[str, ...]:
+    try:
+        configured: Final = _DROP_PATHS.validate_python(kwargs.get("additional_drop_params"))
+    except ValidationError:
+        return ()
+    return tuple(path for path in configured if isinstance(path, str))
 
 
 def _bridges_to_responses_api(model: str, custom_llm_provider: str) -> bool:
@@ -679,6 +688,7 @@ def anthropic_messages_handler(
             model=model,
             drop_params=litellm_params.get("drop_params") is True,
             custom_llm_provider=custom_llm_provider,
+            additional_drop_params=additional_drop_paths(request_kwargs),
         )
     )
     resolved_api_base: Final = (

litellm/llms/anthropic/pass_through/messages/utils.py

@@ -182,6 +182,7 @@ class AnthropicMessagesRequestUtils:
         model: str | None = None,
         drop_params: bool = False,
         custom_llm_provider: str | None = None,
+        additional_drop_params: Iterable[str] = (),
     ) -> AnthropicMessagesRequestOptionalParams:
         """
         Filter parameters to only include those defined in AnthropicMessagesRequestOptionalParams.
@@ -191,11 +192,13 @@ class AnthropicMessagesRequestUtils:
             model: Resolved model id; when set, unsupported params may be dropped
             drop_params: Per-request drop_params flag (also respects litellm.drop_params)
             custom_llm_provider: Routed provider; fast mode is gated to direct Anthropic
+            additional_drop_params: Paths the deployment drops. A listed top-level key is left out
+                before the unsupported-param checks run, so it never raises
 
         Returns:
             AnthropicMessagesRequestOptionalParams instance with only the valid parameters
         """
-        valid_keys: Final = anthropic_messages_optional_param_keys()
+        valid_keys: Final = anthropic_messages_optional_param_keys().difference(additional_drop_params)
         filtered_params: Final = {k: v for k, v in params.items() if k in valid_keys and v is not None}
         if model is not None:
             from litellm.llms.anthropic.chat.transformation import AnthropicConfig

litellm/llms/anthropic/prompt_cache_prediction.py

@@ -650,17 +650,15 @@ def prepare_native_baseline_body(request: Mapping[str, object], model: str) -> M
             custom_llm_provider="anthropic",
             tools=_MESSAGES.validate_python(owned.get("tools") or []),
         )
+        drop_paths: Final = TypeAdapter(tuple[str, ...]).validate_python(owned.get("additional_drop_params") or ())
         options: Final = AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
             {**owned, "system": system},
             model=resolved_model,
             custom_llm_provider="anthropic",
             drop_params=owned.get("drop_params") is True,
+            additional_drop_params=drop_paths,
         )
-        filtered: Final = reduce(
-            delete_nested_value,
-            TypeAdapter(tuple[str, ...]).validate_python(owned.get("additional_drop_params") or ()),
-            dict(options),
-        )
+        filtered: Final = reduce(delete_nested_value, drop_paths, dict(options))
         body: Final = AnthropicMessagesConfig().transform_anthropic_messages_request(
             model=resolved_model,
             messages=strip_provider_specific_fields_from_anthropic_messages(messages),

tests/integration/messages_endpoint/test_messages_additional_drop_params_wire.py

@@ -0,0 +1,119 @@
+import json
+from collections.abc import Mapping
+from dataclasses import dataclass
+from typing import Final
+
+import pytest
+from integration._support.client import Gateway, list_value, object_value
+from integration._support.wire import Reply, Request, Wire, wire_server
+from pydantic import JsonValue, TypeAdapter
+
+_JSON: Final = TypeAdapter(dict[str, JsonValue])
+_ANSWER: Final = "sampling params dropped"
+_SAMPLING_SPELLINGS: Final = ("top_k", "top_p", "topK", "topP")
+_AWS: Final[dict[str, JsonValue]] = {
+    "aws_access_key_id": "AKIASCRIPTEDPROVIDER",
+    "aws_secret_access_key": "scripted-secret",
+    "aws_region_name": "us-east-1",
+}
+
+
+@dataclass(frozen=True, slots=True)
+class Backend:
+    model: str
+    path_suffix: str
+    credentials: Mapping[str, JsonValue]
+
+
+_BEDROCK_INVOKE: Final = Backend("bedrock/invoke/us.anthropic.claude-opus-4-8", "/invoke", _AWS)
+_ANTHROPIC: Final = Backend("anthropic/claude-opus-4-8", "/v1/messages", {})
+_BEDROCK_CONVERSE: Final = Backend("bedrock/converse/us.anthropic.claude-opus-4-8", "/converse", _AWS)
+_NATIVE_BACKENDS: Final = (
+    pytest.param(_BEDROCK_INVOKE, id="bedrock_invoke"),
+    pytest.param(_ANTHROPIC, id="anthropic"),
+)
+
+
+def _answer(request: Request) -> Reply:
+    if request.target.endswith("/converse"):
+        return Reply(
+            body=json.dumps(
+                {
+                    "output": {"message": {"role": "assistant", "content": [{"text": _ANSWER}]}},
+                    "stopReason": "end_turn",
+                    "usage": {"inputTokens": 5, "outputTokens": 3, "totalTokens": 8},
+                    "metrics": {"latencyMs": 1},
+                }
+            ).encode()
+        )
+    return Reply(
+        body=json.dumps(
+            {
+                "id": "msg_dropped_sampling_params",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-opus-4-8",
+                "content": [{"type": "text", "text": _ANSWER}],
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+                "usage": {"input_tokens": 5, "output_tokens": 3},
+            }
+        ).encode()
+    )
+
+
+def _messages(gateway: Gateway, model: str, **sampling: JsonValue) -> tuple[int, str]:
+    response: Final = gateway.request(
+        "POST",
+        "/v1/messages",
+        {"model": model, "max_tokens": 32, "messages": [{"role": "user", "content": "hi"}], **sampling},
+    )
+    return response.status_code, response.text
+
+
+def _answered(status: int, text: str) -> None:
+    assert status == 200, text
+    content: Final = list_value(_JSON.validate_json(text)["content"])
+    assert object_value(content[0])["text"] == _ANSWER, text
+
+
+def _sent_without_sampling_params(wire: Wire, backend: Backend) -> None:
+    (request,) = wire.drain()
+    assert request.target.endswith(backend.path_suffix), request.target
+    sent: Final = request.body.decode()
+    assert not any(spelling in sent for spelling in _SAMPLING_SPELLINGS), sent
+
+
+@pytest.mark.parametrize(
+    "backend", (*_NATIVE_BACKENDS, pytest.param(_BEDROCK_CONVERSE, id="bedrock_converse_chat_bridge"))
+)
+def test_messages_drops_the_sampling_params_a_deployment_lists_on_a_model_without_them(
+    gateway: Gateway, backend: Backend
+) -> None:
+    with wire_server(_answer) as wire, gateway.scenario() as scenario:
+        model: Final = scenario.model(
+            model=backend.model,
+            api_base=wire.url,
+            additional_drop_params=["top_k", "top_p"],
+            **backend.credentials,
+        )
+        _answered(*_messages(gateway, model, top_k=5, top_p=0.9))
+        _sent_without_sampling_params(wire, backend)
+
+
+@pytest.mark.parametrize("backend", _NATIVE_BACKENDS)
+def test_messages_rejects_a_sampling_param_the_deployment_does_not_list_before_calling_the_provider(
+    gateway: Gateway, backend: Backend
+) -> None:
+    with wire_server(_answer) as wire, gateway.scenario() as scenario:
+        model: Final = scenario.model(
+            model=backend.model,
+            api_base=wire.url,
+            additional_drop_params=["top_k"],
+            **backend.credentials,
+        )
+        rejected_status, rejected_text = _messages(gateway, model, top_k=5, top_p=0.9)
+        assert rejected_status == 400, rejected_text
+        assert "does not support top_p=0.9" in rejected_text, rejected_text
+        _answered(*_messages(gateway, model, top_k=5))
+        _sent_without_sampling_params(wire, backend)

tests/unit/llms/anthropic/pass_through/messages/test_anthropic_experimental_pass_through_messages_handler.py

@@ -1635,6 +1635,89 @@ _CLIENT_BETA_HEADERS: Final = (
 )
 
 
+_OPUS_4_8_ON_BEDROCK_INVOKE: Final = "bedrock/invoke/us.anthropic.claude-opus-4-8"
+
+
+def _bedrock_invoke_recording(bodies: list[dict[str, object]]) -> AsyncHTTPHandler:
+    def record_and_answer(request: httpx.Request) -> httpx.Response:
+        bodies.append(json.loads(request.content))
+        return httpx.Response(
+            200,
+            json={
+                "id": "msg_drop_top_k",
+                "type": "message",
+                "role": "assistant",
+                "model": "claude-opus-4-8",
+                "content": [{"type": "text", "text": "ok"}],
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+                "usage": {"input_tokens": 3, "output_tokens": 1},
+            },
+            request=request,
+        )
+
+    upstream = AsyncHTTPHandler()
+    upstream.client = httpx.AsyncClient(transport=httpx.MockTransport(record_and_answer))
+    return upstream
+
+
+async def _opus_4_8_messages_with_top_k(bodies: list[dict[str, object]], **deployment: object) -> object:
+    return await messages.acreate(
+        max_tokens=16,
+        messages=[{"role": "user", "content": "hi"}],
+        model=_OPUS_4_8_ON_BEDROCK_INVOKE,
+        top_k=5,
+        aws_access_key_id="test-access-key",
+        aws_secret_access_key="test-secret-key",
+        aws_region_name="us-east-1",
+        client=_bedrock_invoke_recording(bodies),
+        **deployment,
+    )
+
+
+@pytest.mark.asyncio
+async def test_bedrock_invoke_messages_drops_top_k_the_deployment_lists_on_a_model_without_sampling_params(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(litellm, "drop_params", False)
+    bodies: list[dict[str, object]] = []
+
+    response = await _opus_4_8_messages_with_top_k(bodies, additional_drop_params=["top_k"])
+
+    assert response["content"] == [{"type": "text", "text": "ok"}]
+    (body,) = bodies
+    assert "top_k" not in body
+    assert body["max_tokens"] == 16
+
+
+@pytest.mark.asyncio
+async def test_bedrock_invoke_messages_still_rejects_top_k_the_deployment_does_not_list(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(litellm, "drop_params", False)
+    bodies: list[dict[str, object]] = []
+
+    with pytest.raises(litellm.UnsupportedParamsError, match="does not support top_k=5") as raised:
+        await _opus_4_8_messages_with_top_k(bodies, additional_drop_params=["top_p"])
+
+    assert raised.value.status_code == 400
+    assert bodies == []
+
+
+@pytest.mark.asyncio
+async def test_bedrock_invoke_messages_answers_the_top_k_400_for_a_drop_list_that_is_not_a_list(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(litellm, "drop_params", False)
+    bodies: list[dict[str, object]] = []
+
+    with pytest.raises(litellm.UnsupportedParamsError, match="does not support top_k=5") as raised:
+        await _opus_4_8_messages_with_top_k(bodies, additional_drop_params="top_k")
+
+    assert raised.value.status_code == 400
+    assert bodies == []
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("client_headers", _CLIENT_BETA_HEADERS)
 async def test_anthropic_messages_forwards_safeguards_and_dangerous_tool_use_beta_to_bedrock_invoke(

tests/unit/llms/anthropic/pass_through/messages/test_request_optional_param_utils.py

@@ -133,3 +133,37 @@ def test_sampling_param_raises_clean_400_without_drop_params(monkeypatch):
             model="claude-opus-4-7",
             drop_params=False,
         )
+
+
+@pytest.mark.parametrize(
+    ("model", "param", "value"),
+    [
+        ("claude-opus-4-8", "temperature", 0.3),
+        ("claude-opus-4-8", "top_p", 0.9),
+        ("claude-opus-4-8", "top_k", 5),
+        ("claude-sonnet-4-6", "speed", "fast"),
+    ],
+)
+def test_deployment_listed_param_is_dropped_before_the_unsupported_param_gate(monkeypatch, model, param, value):
+    monkeypatch.setattr(litellm, "drop_params", False)
+    result = AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+        params={param: value, "stream": True},
+        model=model,
+        drop_params=False,
+        additional_drop_params=(param,),
+    )
+
+    assert result == {"stream": True}
+
+
+def test_param_missing_from_the_drop_list_still_raises_the_unsupported_param_400(monkeypatch):
+    monkeypatch.setattr(litellm, "drop_params", False)
+    with pytest.raises(litellm.utils.UnsupportedParamsError, match="does not support top_p") as raised:
+        AnthropicMessagesRequestUtils.get_requested_anthropic_messages_optional_param(
+            params={"top_k": 5, "top_p": 0.9},
+            model="claude-opus-4-8",
+            drop_params=False,
+            additional_drop_params=("top_k", "top_p.nested"),
+        )
+
+    assert raised.value.status_code == 400

tests/unit/llms/anthropic/test_anthropic_prompt_cache_prediction.py

@@ -26,6 +26,7 @@ from litellm.llms.anthropic.prompt_cache_prediction import (
     parse_cache_plan,
     parse_observed_cache,
     parse_prompt,
+    prepare_native_baseline_body,
     resolve_baseline_prediction_target,
     resolve_prediction_target,
     supported_prediction_headers,
@@ -452,3 +453,23 @@ async def test_native_count_preserves_settings_and_requires_every_prefix(
         assert tuple(marker.prefix_tokens for marker in result.breakpoints) == (5000, 5800)
     assert tuple(route.call_count for route in routes) == (1, 1, 1)
     assert unexpected.call_count == 0
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+def test_native_baseline_body_leaves_out_a_listed_sampling_param_the_model_does_not_support(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setattr(litellm, "drop_params", False)
+    request: Final = {
+        "model": "claude-opus-4-8",
+        "max_tokens": 64,
+        "messages": [{"role": "user", "content": "hi"}],
+        "top_k": 5,
+        "additional_drop_params": ["top_k"],
+    }
+
+    body: Final = prepare_native_baseline_body(request, "anthropic/claude-opus-4-8")
+
+    assert body is not None
+    assert "top_k" not in body
+    assert body["max_tokens"] == 64