fix(proxy): keep every repeated form field value in get_form_data

dict(form) collapses duplicate multipart keys to the last one, so a request with
timestamp_granularities[]=segment and timestamp_granularities[]=word reached the
provider as a single-element list. Iterate form.multi_items() instead.

Fixes #35937

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/proxy/common_utils/http_parsing_utils.py

@@ -274,9 +274,8 @@ async def get_form_data(request: Request) -> dict[str, Any]:
     Handles when OpenAI SDKs pass form keys as `timestamp_granularities[]="word"` instead of `timestamp_granularities=["word", "sentence"]`
     """
     form: Final = await request.form()
-    form_data: Final = dict(form)
     parsed_form_data: Final[dict[str, Any]] = {}
-    for key, value in form_data.items():
+    for key, value in form.multi_items():
         # OpenAI SDKs pass form keys as `timestamp_granularities[]="word"` instead of `timestamp_granularities=["word", "sentence"]`
         if key.endswith("[]"):
             clean_key = key[:-2]

tests/test_litellm/proxy/common_utils/test_http_parsing_utils.py

@@ -7,6 +7,7 @@ import orjson
 import pytest
 from fastapi import Request
 from fastapi.testclient import TestClient
+from starlette.datastructures import FormData
 
 sys.path.insert(
     0, os.path.abspath("../../../..")
@@ -506,19 +507,20 @@ async def test_get_form_data():
     # Create a mock request with transcription form data
     mock_request = MagicMock()
 
-    # Create mock form data with array notation for timestamp_granularities
-    mock_form_data = {
-        "file": "file_object",  # In a real request this would be an UploadFile
-        "model": "gpt-4o-transcribe",
-        "include[]": "logprobs",  # Array notation
-        "language": "en",
-        "prompt": "Transcribe this audio file",
-        "response_format": "json",
-        "stream": "false",
-        "temperature": "0.2",
-        "timestamp_granularities[]": "word",  # First array item
-        "timestamp_granularities[]": "segment",  # Second array item (would overwrite in dict, but handled by the function)
-    }
+    mock_form_data = FormData(
+        [
+            ("file", "file_object"),  # In a real request this would be an UploadFile
+            ("model", "gpt-4o-transcribe"),
+            ("include[]", "logprobs"),
+            ("language", "en"),
+            ("prompt", "Transcribe this audio file"),
+            ("response_format", "json"),
+            ("stream", "false"),
+            ("temperature", "0.2"),
+            ("timestamp_granularities[]", "word"),
+            ("timestamp_granularities[]", "segment"),
+        ]
+    )
 
     # Mock the form method to return the test data
     mock_request.form = AsyncMock(return_value=mock_form_data)
@@ -540,11 +542,23 @@ async def test_get_form_data():
     assert isinstance(result["include"], list)
     assert "logprobs" in result["include"]
 
-    assert "timestamp_granularities" in result
-    assert isinstance(result["timestamp_granularities"], list)
-    # Note: In a real MultiDict, both values would be present
-    # But in our mock dictionary the second value overwrites the first
-    assert "segment" in result["timestamp_granularities"]
+    assert result["timestamp_granularities"] == ["word", "segment"]
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "granularities",
+    [["segment", "word"], ["word", "segment"], ["segment"], ["word"]],
+)
+async def test_get_form_data_preserves_every_repeated_array_value(granularities):
+    mock_request = MagicMock()
+    mock_request.form = AsyncMock(
+        return_value=FormData([("model", "whisper-1"), *(("timestamp_granularities[]", g) for g in granularities)])
+    )
+
+    result = await get_form_data(mock_request)
+
+    assert result["timestamp_granularities"] == granularities
 
 
 def test_get_tags_from_request_body_with_metadata_tags():
@@ -969,7 +983,7 @@ class TestReadRequestBodyNonCanonicalContentType:
     @pytest.mark.asyncio
     async def test_real_form_post_still_parsed_as_form(self):
         mock_request = MagicMock()
-        mock_request.form = AsyncMock(return_value={"k": "v"})
+        mock_request.form = AsyncMock(return_value=FormData([("k", "v")]))
         mock_request.body = AsyncMock(return_value=b"")
         mock_request.headers = {"content-type": "application/x-www-form-urlencoded"}
         mock_request.scope = {}
@@ -1025,7 +1039,7 @@ class TestGetRequestBody:
         mock_request = MagicMock()
         mock_request.method = "POST"
         mock_request.headers = {"content-type": "multipart/form-data; boundary=x"}
-        mock_request.form = AsyncMock(return_value={"k": "v"})
+        mock_request.form = AsyncMock(return_value=FormData([("k", "v")]))
         mock_request.scope = {}
 
         result = await get_request_body(mock_request)

tests/test_litellm/proxy/pass_through_endpoints/test_llm_pass_through_endpoints.py

@@ -9,6 +9,7 @@ import httpx
 import pytest
 from fastapi import Request, Response
 from fastapi.testclient import TestClient
+from starlette.datastructures import FormData
 
 sys.path.insert(
     0, os.path.abspath("../../../..")
@@ -1259,7 +1260,7 @@ async def test_is_streaming_request_fn():
     mock_request = Mock()
     mock_request.method = "POST"
     mock_request.headers = {"content-type": "multipart/form-data"}
-    mock_request.form = AsyncMock(return_value={"stream": "true"})
+    mock_request.form = AsyncMock(return_value=FormData([("stream", "true")]))
     assert await is_streaming_request_fn(mock_request) is True