diff --git a/python/sglang/srt/managers/schedule_batch.py b/python/sglang/srt/managers/schedule_batch.py index b40f93002..77472f97b 100644 --- a/python/sglang/srt/managers/schedule_batch.py +++ b/python/sglang/srt/managers/schedule_batch.py @@ -878,6 +878,8 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin): # If req.input_embeds is already a list, append its content directly input_embeds.extend(req.input_embeds) # Use extend to avoid nesting + if req.is_retracted: + req.already_computed = 0 req.cached_tokens += pre_len - req.already_computed req.already_computed = seq_len req.is_retracted = False