[Bugfix][Frontend] Validate reused prompt token ids; render messages for media and echo (#55771)

Signed-off-by: Shijie Lyu <sjlyu@amazon.com>
Co-authored-by: Shijie Lyu <sjlyu@amazon.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Shijie Lyu
2026-09-29 10:36:22 +08:00
committed by GitHub
co-authored by Shijie Lyu Claude
parent 0af34418e9
commit c6843d397c
6 changed files with 239 additions and 2 deletions
+25 -1
View File
@@ -107,11 +107,35 @@ def _reused_prompt_token_ids(request: Any) -> list[int] | None:
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.
"""
kv = getattr(request, "kv_transfer_params", None)
if not isinstance(kv, dict):
return None
return kv.pop("prompt_token_ids", None) or None
ids = kv.pop("prompt_token_ids", None)
if ids is None:
return None
if getattr(request, "echo", False):
logger.debug(
"Ignoring kv_transfer_params['prompt_token_ids']: "
"echo is set, so messages are rendered instead."
)
return None
# bool is an int subclass, hence the exact type check.
if (
not isinstance(ids, list)
or not ids
or any(type(x) is not int or x < 0 for x in ids)
):
raise VLLMValidationError(
"`kv_transfer_params['prompt_token_ids']` must be a non-empty list "
"of non-negative integers.",
parameter="kv_transfer_params.prompt_token_ids",
)
return ids
class OnlineRenderer: