[Bugfix][Frontend] Ignore reused prompt token ids for media in /v1/responses (#59173)

Signed-off-by: Shijie Lyu <lshjhf@gmail.com>
This commit is contained in:
Shijie Lyu
2026-09-30 05:21:09 +00:00
committed by GitHub
parent a3af78384a
commit ac68c30872
4 changed files with 146 additions and 33 deletions
+26
View File
@@ -1814,6 +1814,32 @@ PART_TYPES_TO_SKIP_NONE_CONTENT = (
TEXT_PART_TYPES = frozenset(
{"text", "input_text", "output_text", "refusal", "thinking"}
)
# Content part types that carry no multimodal data.
_TEXT_CONTENT_PART_TYPES = TEXT_PART_TYPES | {"tool_reference"}
# Keys that mark a content part as multimodal, whatever its ``type``.
_MEDIA_CONTENT_PART_KEYS = frozenset(MM_PARSER_MAP) - _TEXT_CONTENT_PART_TYPES
def has_non_text_content(messages: Any) -> bool:
"""Whether any message in ``messages`` has a non-text content part.
Only list content is inspected, so validated chat content that is a
one-shot iterator is never consumed.
"""
if not isinstance(messages, list):
return False
for msg in messages:
content = msg.get("content") if isinstance(msg, dict) else None
if not isinstance(content, list):
continue
for part in content:
if isinstance(part, dict) and (
any(key in part for key in _MEDIA_CONTENT_PART_KEYS)
or not isinstance(part_type := part.get("type", "text"), str)
or part_type not in _TEXT_CONTENT_PART_TYPES
):
return True
return False
def _parse_chat_message_content_parts(