[Bug] v0.12.2-cu12.8部署qwen3.5的时候请，求任何内容出现 illegal memory

[huangtao2999]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.

Describe the bug

latest-cu128 在单卡2080TI（ 22G显存 ） + 8卡成功部署了Qwen3.5-35B-A3B。但是用请求测试的时候出现illegal memory 错误

Reproduction

docker run -d
--runtime nvidia
--gpus '"device=0,1,2,3,4,5,6,7"'
--name lmdeploy_ht
-v /data/model/hugginface:/models
-v ~/.cache/huggingface:/root/.cache/huggingface
-p 23333:23333
--ipc=host
--shm-size=128g
docker.1ms.run/openmmlab/lmdeploy:v0.12.2-cu12.8
tail -f /dev/null
docker exec -it lmdeploy_ht bash

lmdeploy serve api_server /models/Qwen/Qwen3.5-35B-A3B
--tp 8
--cache-max-entry-count 0.8
--log-level INFO
--max-concurrent-requests 4
--model-name pkumlm_txt
--backend turbomind
--max_batch_size 64
--cache-block-seq-len 32

Environment

v0.12.2-cu12.8
单卡2080TI（  22G显存 ） * 8卡
Qwen3.5-35B-A3B

Error traceback

curl  -H "Accept: application/json"  -H "Content-type: application/json"   -X POST  -d '{"model": "pkumlm_txt","messages":[{"role":"user","content":"你是谁"} ]}'   http://192.168.1.36:23333/v1/chat/completions

报错
2026-03-19 12:52:50,954 - lmdeploy - INFO - turbomind.py:687 - [async_stream_infer] session 1 start
[TM][INFO] [SeqMgr][Create] ID 1
[TM][WARNING] [ProcessInferRequests] [1] total sequence length (11 + 262133) exceeds `session_len` (258368), `max_new_tokens` is truncated to 258357
invalid argument

[KevinLiuMY]

遇到类似的问题

同时我还进行了如下测试

1. v0.10 + Qwen3-Coder-30B-A3B-Instruct-AWQ --> 工作正常
2. v0.12.2 + Qwen3-Coder-30B-A3B-Instruct-AWQ --> 工作异常

发送任何请求就立刻 coredump