[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
@@ -34,6 +34,7 @@ from openai_harmony import Message as OpenAIHarmonyMessage
from openai_harmony import Role
import vllm.envs as envs
from tests.entrypoints.openai.chat_completion.test_serving_chat import MockModelConfig
from vllm.entrypoints.generate.base.protocol import (
DeltaFunctionCall,
DeltaMessage,
@@ -294,6 +295,102 @@ async def test_online_renderer_renders_non_harmony_responses_with_explicit_histo
assert preprocess_kwargs["default_template_kwargs"]["request_default"] == "kept"
_REUSED_IDS = [10, 20, 30]
_IMAGE_PART = {
"type": "input_image",
"image_url": "https://example.com/a.png",
"detail": "auto",
}
async def _render_responses_with_reuse(request, previous_messages=None):
online_renderer = OnlineRenderer(
model_config=MockModelConfig(),
renderer=MagicMock(),
request_logger=None,
chat_template=None,
chat_template_content_format="auto",
)
render_chat = AsyncMock(return_value=([[]], [tokens_input([1, 2])]))
online_renderer.renderer.render_chat_async = render_chat
result = await online_renderer.render_responses(
request, previous_messages=previous_messages
)
assert not isinstance(result, ErrorResponse)
return render_chat, result.engine_input
@pytest.mark.asyncio
@pytest.mark.parametrize("ids", [_REUSED_IDS, [1.5]])
@pytest.mark.parametrize(
"request_input",
[
[
{
"role": "user",
"content": [{"type": "input_text", "text": "look"}, _IMAGE_PART],
}
],
[
{
"type": "function_call",
"call_id": "call_1",
"name": "shot",
"arguments": "{}",
},
{
"type": "function_call_output",
"call_id": "call_1",
"output": [_IMAGE_PART],
},
],
],
ids=["input_image", "tool_output_image"],
)
async def test_responses_kv_transfer_prompt_token_ids_ignored_with_non_text_input(
request_input, ids
):
request = ResponsesRequest(
input=request_input,
kv_transfer_params={"do_remote_prefill": True, "prompt_token_ids": ids},
)
render_chat, _ = await _render_responses_with_reuse(request)
render_chat.assert_awaited_once()
assert request.kv_transfer_params == {"do_remote_prefill": True}
@pytest.mark.asyncio
async def test_responses_kv_transfer_prompt_token_ids_ignored_with_media_in_history():
request = ResponsesRequest(
input="next", kv_transfer_params={"prompt_token_ids": _REUSED_IDS}
)
render_chat, _ = await _render_responses_with_reuse(
request, previous_messages=[{"role": "user", "content": [_IMAGE_PART]}]
)
render_chat.assert_awaited_once()
@pytest.mark.asyncio
async def test_responses_kv_transfer_prompt_token_ids_kept_with_text_input():
request = ResponsesRequest(
input=[
{"role": "user", "content": [{"type": "input_text", "text": "hi"}]},
{"role": "assistant", "content": "ok"},
],
kv_transfer_params={"prompt_token_ids": _REUSED_IDS},
)
_, engine_input = await _render_responses_with_reuse(request)
assert engine_input["prompt_token_ids"] == _REUSED_IDS
@pytest.mark.asyncio
async def test_online_renderer_rejects_untrusted_responses_chat_template():
renderer = _new_online_renderer()
+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(
@@ -20,10 +20,9 @@ from pydantic import (
from vllm.config import ModelConfig
from vllm.entrypoints.chat_utils import (
MM_PARSER_MAP,
TEXT_PART_TYPES,
ChatCompletionMessageParam,
ChatTemplateContentFormatOption,
has_non_text_content,
)
from vllm.entrypoints.generate.base.protocol import (
AnyResponseFormat,
@@ -60,11 +59,6 @@ logger = init_logger(__name__)
_INT64_MIN = -(2**63)
_INT64_MAX = 2**63 - 1
# 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
class ChatMessage(OpenAIBaseModel):
role: str
@@ -991,29 +985,16 @@ class ChatCompletionRequest(OpenAIBaseModel):
if not isinstance(data, dict):
return data
kv_transfer_params = data.get("kv_transfer_params")
messages = data.get("messages")
if (
not isinstance(kv_transfer_params, dict)
or kv_transfer_params.get("prompt_token_ids") is None
or not isinstance(messages, list)
isinstance(kv_transfer_params, dict)
and kv_transfer_params.get("prompt_token_ids") is not None
and has_non_text_content(data.get("messages"))
):
return data
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
):
logger.debug(
"Ignoring kv_transfer_params['prompt_token_ids']: "
"messages have non-text content and are rendered instead."
)
kv_transfer_params.pop("prompt_token_ids")
return data
logger.debug(
"Ignoring kv_transfer_params['prompt_token_ids']: "
"messages have non-text content and are rendered instead."
)
kv_transfer_params.pop("prompt_token_ids")
return data
@model_validator(mode="before")
+14 -5
View File
@@ -16,6 +16,7 @@ from vllm.entrypoints.chat_utils import (
ChatCompletionMessageParam,
ChatTemplateContentFormatOption,
ConversationMessage,
has_non_text_content,
)
from vllm.entrypoints.generate.base.protocol import validate_request_mm_kwargs
from vllm.entrypoints.openai.chat_completion.protocol import (
@@ -101,16 +102,18 @@ def _extract_allowed_tools_from_mcp_requests(
return allowed_tools_map
def _reused_prompt_token_ids(request: Any) -> list[int] | None:
def _reused_prompt_token_ids(
request: Any, messages: list[Any] | None = None
) -> list[int] | None:
"""Pop prompt token ids forwarded for decode-side reuse, if any.
Disaggregated serving carries the prefill stage's ids in
``kv_transfer_params`` so the decode stage can skip re-tokenizing. Removing
the key keeps the id list out of the engine's sampling metadata.
Returns None without checking the ids when ``echo`` is set, since echo
needs ``messages`` to be rendered. Otherwise raises VLLMValidationError if
the ids are malformed.
Returns None without checking the ids when ``echo`` is set or ``messages``
has non-text content, since both need ``messages`` to be rendered.
Otherwise raises VLLMValidationError if the ids are malformed.
"""
kv = getattr(request, "kv_transfer_params", None)
if not isinstance(kv, dict):
@@ -124,6 +127,12 @@ def _reused_prompt_token_ids(request: Any) -> list[int] | None:
"echo is set, so messages are rendered instead."
)
return None
if has_non_text_content(messages):
logger.debug(
"Ignoring kv_transfer_params['prompt_token_ids']: "
"messages have non-text content and are rendered instead."
)
return None
# bool is an int subclass, hence the exact type check.
if (
not isinstance(ids, list)
@@ -772,7 +781,7 @@ class OnlineRenderer:
default_mm_processor_kwargs=getattr(request, "mm_processor_kwargs", None),
)
reuse_ids = _reused_prompt_token_ids(request)
reuse_ids = _reused_prompt_token_ids(request, messages)
if reuse_ids:
# Decode-side token reuse: feed the forwarded ids straight to the
# engine, skipping templating and tokenization. ``messages`` are not