[Feature] support stop_token_ids (#5399)

* support stop_token_ids

* fix

* delete chinese

* support both

* delete print
This commit is contained in:
lizexu123
2025-12-09 17:49:12 +08:00
committed by GitHub
parent df67379bc3
commit 95eab9f9ee
18 changed files with 377 additions and 127 deletions
+31
View File
@@ -19,3 +19,34 @@ __all__ = [
]
IDS_TYPE_FLAG = {"text": 0, "image": 1, "video": 2, "audio": 3}
from typing import Any, Callable, Dict, List, Tuple
def process_stop_token_ids(
request: Dict[str, Any],
update_stop_seq_fn: Callable[[List[str]], Tuple[List[List[int]], List[int]]],
) -> None:
stop_token_ids_final = []
if request.get("stop_token_ids") is not None:
stop_token_ids = request.get("stop_token_ids")
if isinstance(stop_token_ids, list) and len(stop_token_ids) > 0:
if isinstance(stop_token_ids[0], int):
# List[int] -> List[List[int]]
stop_token_ids_final.extend([[t] for t in stop_token_ids])
elif isinstance(stop_token_ids[0], list):
# Already List[List[int]]
stop_token_ids_final.extend(stop_token_ids)
stop_sequences = request.get("stop", [])
if stop_sequences:
stop_seqs, _ = update_stop_seq_fn(stop_sequences)
stop_token_ids_final.extend(stop_seqs)
# Update request
if stop_token_ids_final:
stop_seqs_len = [len(seq) for seq in stop_token_ids_final]
request["stop_token_ids"] = stop_token_ids_final
request["stop_seqs_len"] = stop_seqs_len