[Bugfix] Gracefully handle unsupported reasoning_effort in chat templates (#54022)

Signed-off-by: frankie <wangyongsheng686@gmail.com>
Co-authored-by: Trae Agent <noreply@trae.ai>
This commit is contained in:
frankie
2026-09-06 21:11:35 -07:00
committed by GitHub
co-authored by Trae Agent
parent f43ef1531d
commit 9cc7793e32
2 changed files with 121 additions and 4 deletions
+20 -4
View File
@@ -27,6 +27,7 @@ from vllm.entrypoints.chat_utils import (
parse_chat_messages,
parse_chat_messages_async,
)
from vllm.exceptions import VLLMValidationError
from vllm.inputs import EmbedsPrompt
from vllm.inputs.engine import MultiModalInput
from vllm.logger import init_logger
@@ -684,6 +685,18 @@ def resolve_chat_template_kwargs(
return {k: v for k, v in chat_template_kwargs.items() if k in accept_vars}
def _template_error_reason(exc: BaseException) -> str:
# Extract the most specific reason from a chat template error chain.
seen: set[int] = set()
current: BaseException | None = exc
while current is not None and id(current) not in seen:
seen.add(id(current))
if isinstance(current, jinja2.TemplateError):
return str(current)
current = current.__cause__ or current.__context__
return str(exc)
@overload
def safe_apply_chat_template(
model_config: ModelConfig,
@@ -806,10 +819,13 @@ def safe_apply_chat_template(
**resolved_kwargs,
)
except Exception as e:
logger.exception(
"An error occurred in `transformers` while applying chat template"
)
raise ValueError(str(e)) from e
# Chat templates reject invalid user input (e.g. an unsupported
# `reasoning_effort` value) by raising from within the template.
# Surface those as a 400 Bad Request carrying the template's own
# reason (which typically lists the supported values) instead of a
# 500 or any generic upstream wrapper message.
logger.warning("Chat template rejected the request: %s", e)
raise VLLMValidationError(_template_error_reason(e)) from e
if return_assistant_tokens_mask:
assert isinstance(plain, list), f"Expected list[int], got {type(plain)}"