Qwen3.8-27B-FP8 部署兼容性问题咨询

前提

目前有一台机器，里面有4张L20的卡，其中有三张已经有其他用途不能动，驱动无法升级，现在想剩下其中的一个卡来试试部署 Qwen3.8-27B-FP8，部署步骤参考：

recipes.vllm.ai

Qwen/Qwen3.8-27B — 27B · DENSE · 256K ctx

27B-parameter dense hybrid-attention model with linear attention on 48 of 64 layers, a vision tower, a built-in MTP draft head, 262K native context window and extensible to 1M context

目前遇到的问题

我的驱动是 Driver Version: 535.216.03 CUDA Version: 12.2

目标模型需要 vllm >= 0.17.0，而该版本的 vLLM 官方预编译包依赖 CUDA 12.9 或 CUDA 13.0 的 PyTorch 环境，与现有的 CUDA 12.2 环境不兼容 。

由于不能升级驱动，咨询一下各位大佬有什么解决思路或者其他部署的方式吗

报错日志```
(vllm_fix1) [root@gpu-l20-10 ~]# CUDA_VISIBLE_DEVICES=1 vllm serve /data/model/qwen/Qwen--Qwen3.8-27B-FP8/snapshots/master \
--served-model-name Qwen3.8-27B-FP8 \
--tensor-parallel-size 1 \
--max-model-len 65314 \
--kv-cache-dtype fp8 \
--reasoning-parser qwen3 \
--port 8088
Traceback (most recent call last):
File "/root/miniconda3/envs/vllm_fix1/bin/vllm", line 4, in <module>
from vllm.entrypoints.cli.main import main
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/__init__.py", line 14, in <module>
import vllm.env_override # noqa: F401
^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/env_override.py", line 90, in <module>
from vllm.utils.torch_utils import is_torch_equal, is_torch_equal_or_newer
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/utils/torch_utils.py", line 72, in <module>
PIN_MEMORY = is_pin_memory_available()
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/utils/platform_utils.py", line 45, in is_pin_memory_available
from vllm.platforms import current_platform
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/platforms/__init__.py", line 277, in __getattr__
_current_platform = resolve_obj_by_qualname(platform_cls_qualname)()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/utils/import_utils.py", line 109, in resolve_obj_by_qualname
module = importlib.import_module(module_name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/importlib/__init__.py", line 126, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/root/miniconda3/envs/vllm_fix1/lib/python3.11/site-packages/vllm/platforms/cuda.py", line 23, in <module>
import vllm._C_stable_libtorch # noqa
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: libcudart.so.13: cannot open shared object file: No such file or directory