Merge 8c59d24d78 into 72228db5c6

litellm/llms/custom_httpx/http_handler.py

@@ -2,6 +2,7 @@ import asyncio
 import concurrent.futures
 import inspect
 import os
+import re
 import socket
 import ssl
 import sys
@@ -272,6 +273,18 @@ _STREAMING_ERROR_BODY_READ_EXECUTOR: Final = concurrent.futures.ThreadPoolExecut
 )
 
 
+_SURROGATE_RE: Final = re.compile(r"[\ud800-\udfff]")
+
+
+def _repair_surrogates(value: str) -> str:
+    """Repair malformed UTF-16 text without copying ordinary Unicode strings."""
+    if not _SURROGATE_RE.search(value):
+        return value
+    # Preserve valid surrogate pairs as their scalar value; replace only lone
+    # surrogates. Byte payloads and caller-owned mappings are never rewritten.
+    return value.encode("utf-16-le", errors="surrogatepass").decode("utf-16-le", errors="replace")
+
+
 def _prepare_request_data_and_content(
     data: dict | str | bytes | None = None,
     content: _RequestContent | None = None,
@@ -309,7 +322,7 @@ def _prepare_request_data_and_content(
             # dict/Mapping stays in data= parameter
             request_data = data
 
-    return request_data, request_content
+    return request_data, _repair_surrogates(request_content) if isinstance(request_content, str) else request_content
 
 
 # Cache for SSL contexts to avoid creating duplicate contexts with the same configuration

tests/unit/llms/custom_httpx/test_raw_body_surrogates.py

@@ -0,0 +1,31 @@
+import httpx
+import pytest
+
+from litellm.llms.custom_httpx.http_handler import _prepare_request_data_and_content
+
+
+@pytest.mark.parametrize("argument", ["data", "content"])
+@pytest.mark.parametrize(
+    ("raw", "expected"),
+    [("hi\ud800!", "hi\ufffd!"), ("\udc00", "\ufffd"), ("\ud83d\ude00", "😀"), ("漢字😀", "漢字😀")],
+)
+def test_raw_text_can_be_encoded_as_utf8(argument, raw, expected):
+    data, content = _prepare_request_data_and_content(**{argument: raw})
+    request = httpx.Request("POST", "http://localhost/", data=data, content=content)
+    assert request.content == expected.encode()
+
+
+def test_valid_text_and_bytes_keep_identity():
+    text = "valid 😀"
+    payload = b"\xed\xa0\x80"
+    assert _prepare_request_data_and_content(data=text) == (None, text)
+    assert _prepare_request_data_and_content(content=text)[1] is text
+    assert _prepare_request_data_and_content(data=payload)[1] is payload
+
+
+def test_explicit_content_wins_without_mutating_form_fields():
+    fields = {"field": "value"}
+    data, content = _prepare_request_data_and_content(data=fields, content="\ud800")
+    assert data is fields
+    assert content == "\ufffd"
+    assert _prepare_request_data_and_content(data="ignored\ud800", content=b"chosen") == (None, b"chosen")