上下文会不断累加，最终超过模型的context length然后报错

显卡是24G 的A5000
启动命令：
ENABLE_SUMMARIZER=false\
CHUNK=6 T_s=6\
MAX_MODEL_LEN=12000\
ADAPTER_HOST="0.0.0.0" WEBINFER_ARGS="--gpu-memory-utilization 0.9 \
--quantization fp8 \
--limit-mm-per-prompt '{\"image\":8,\"video\":1}'" bash services/webinfer/scripts/run.sh all

在运行400多轮后报错
(APIServer pid=314978) ValueError: Input length (12272) exceeds model's maximum context length (12000).
2026-07-02 06:31:30,816 INFO httpx: HTTP Request: POST http://127.0.0.1:7060/v1/chat/completions "HTTP/1.1 400 Bad Request"
2026-07-02 06:31:30,817 ERROR streaming_infer_adapter: chat completion failed
Traceback (most recent call last):

由于显存比较小，最大只能设置MAX_MODEL_LEN=12000
但是上下文会不断累加最终超过12000
请问下这种问题怎么解决。
我有个疑问，假设是80G的显存，那上下文不断累加的话是不是最终也会报错？

Start with services/webinfer/scripts/run.sh and reproduce the failure using the reported settings, especially ENABLE_SUMMARIZER=false and MAX_MODEL_LEN=12000. Trace how conversation context is accumulated across requests and identify where the 12,272-token input exceeds the configured limit. Done means the behavior for long-running conversations is defined and verified without exceeding the model context limit.

Written by the indexing model from the issue text.