Merge 4819f686f9 into ab3b12b951

litellm/proxy/common_utils/http_parsing_utils.py

@@ -1,6 +1,6 @@
 import json
 import re
-from collections.abc import Collection, Mapping
+from collections.abc import Collection, Mapping, Sequence
 from types import MappingProxyType, UnionType
 from typing import Annotated, Any, Final, Literal, Protocol, Union, get_args, get_origin
 
@@ -32,10 +32,14 @@ _FORM_CONTENT_TYPES: Final[frozenset[str]] = frozenset({"application/x-www-form-
 # Binary bodies (e.g. OTLP trace exports on POST /v1/traces) are not JSON: arbitrary bytes used to
 # hit the JSON surrogate-repair path and fail auth with a 400. JSON under these types still parses.
 _BINARY_CONTENT_TYPES: Final[frozenset[str]] = frozenset({"application/x-protobuf", "application/protobuf"})
+_FORM_FALLBACK_FIELDS: Final[frozenset[str]] = frozenset(
+    {"fallbacks", "context_window_fallbacks", "content_policy_fallbacks"}
+)
 
 _ANNOTATION_QUALIFIERS: Final[frozenset[object]] = frozenset({Annotated, NotRequired, ReadOnly, Required})
 _DECODED_JSON: Final = TypeAdapter(object)
 _JSON_REQUEST_BODY: Final = TypeAdapter(Mapping[str, object])
+_FORM_FALLBACK_LIST: Final = TypeAdapter(list[object])
 
 
 class _ReadableUpload(Protocol):
@@ -233,6 +237,28 @@ def _parse_form_metadata(raw_metadata: str) -> object:
     return parsed
 
 
+def _parse_form_fallbacks(field: str, raw_fallbacks: str) -> Sequence[object]:
+    try:
+        return _FORM_FALLBACK_LIST.validate_json(raw_fallbacks)
+    except ValidationError as e:
+        raise ProxyException(
+            message=f"Invalid form payload: {field} must be a JSON array",
+            type="invalid_request_error",
+            param=field,
+            code=status.HTTP_400_BAD_REQUEST,
+        ) from e
+
+
+def _parse_form_field(field: str, value: object) -> object:
+    if not isinstance(value, str):
+        return value
+    if field == "metadata":
+        return _parse_form_metadata(value)
+    if field in _FORM_FALLBACK_FIELDS:
+        return _parse_form_fallbacks(field, value)
+    return value
+
+
 def _reraise_request_size_error(error: Exception) -> None:
     if isinstance(error, HTTPException) and error.status_code == status.HTTP_413_CONTENT_TOO_LARGE:
         raise error
@@ -292,12 +318,7 @@ async def read_request_body(request: Request | None) -> dict:
             parsed_body = _parse_binary_body(binary_body)
         elif _is_form_content_type(content_type):
             form_body: Final = dict(await _read_form_data(request))
-            raw_metadata: Final = form_body.get("metadata")
-            parsed_body = (
-                {**form_body, "metadata": _parse_form_metadata(raw_metadata)}
-                if isinstance(raw_metadata, str)
-                else form_body
-            )
+            parsed_body = {field: _parse_form_field(field, value) for field, value in form_body.items()}
         else:
             # Read the request body
             body: Final = await request.body()

tests/unit/proxy/common_utils/test_http_parsing_utils.py

@@ -423,6 +423,37 @@ async def test_form_data_with_invalid_json_metadata() -> None:
     assert exc_info.value.param == "metadata"
 
 
+def _form_request(fields: Mapping[str, str]) -> MagicMock:
+    request: Final = MagicMock()
+    request.form = AsyncMock(return_value=FormData(fields))
+    request.headers = {"content-type": "multipart/form-data"}
+    request.scope = {}
+    request.state._cached_headers = None
+    return request
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("field", "raw", "expected"),
+    [
+        ("fallbacks", '["backup"]', ["backup"]),
+        ("context_window_fallbacks", '[{"primary": ["backup"]}]', [{"primary": ["backup"]}]),
+        ("content_policy_fallbacks", '[{"primary": ["backup"]}]', [{"primary": ["backup"]}]),
+    ],
+)
+async def test_form_data_fallback_fields_are_decoded_to_lists(field: str, raw: str, expected: list[object]) -> None:
+    result: Final = await read_request_body(_form_request({"model": "primary", field: raw}))
+    assert result == {"model": "primary", field: expected}, result
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("raw", ['"backup"', "backup", '{"primary": ["backup"]}'])
+async def test_form_data_fallbacks_that_are_not_a_json_array_are_rejected(raw: str) -> None:
+    with pytest.raises(ProxyException) as exc_info:
+        await read_request_body(_form_request({"model": "primary", "fallbacks": raw}))
+    assert (exc_info.value.code, exc_info.value.param) == ("400", "fallbacks"), exc_info.value.message
+
+
 @pytest.mark.asyncio
 async def test_form_data_without_metadata():
     """