[Bug] torch_npu与torch_cuda环境冲突导致包导入失败

### 检查清单

- [x] 1. 我已经搜索过相关问题，但未能获得预期的帮助
- [x] 2. 该问题在最新版本中尚未修复
- [x] 3. 请注意，如果您提交的BUG相关 issue 缺少对应环境信息和最小可复现示例，我们将难以复现和定位问题，降低获得反馈的可能性
- [x] 4. 如果您提出的不是bug而是问题，请在讨论区发起讨论 https://github.com/kvcache-ai/ktransformers/discussions。否则该 issue 将被关闭
- [x] 5. 为方便社区交流，我将使用中文/英文或附上中文/英文翻译（如使用其他语言）。未附带翻译的非中文/英语内容可能会被关闭

### 问题描述

显卡A800 80g
环境：
Python 3.11.14
ktransformers 0.3.2+cu126torch29fancy
torch 2.9.0+cu126
torchaudio 2.9.0+cu126
torchvision 0.24.0+cu126
分支是20251103 16：00左右拉取，先后两次拉取最新代码，仍然无法运行，报错没有setup_model_parallel，报错信息如下：

(ktransformers)_** user@user-R8488-G12:/data/wmh/ktransformers$ python -m ktransformers.local_chat --model_path /data/wmh/model/qwen14b --gguf_path ./qwen14b-GGUF
no balance_serve
2025-11-03 17:04:55,768 - INFO - flashinfer.jit: Prebuilt kernels not found, using JIT backend
found flashinfer
Traceback (most recent call last):
File "", line 198, in _run_module_as_main
File "", line 88, in _run_code
File "/data/wmh/ktransformers/ktransformers/local_chat.py", line 258, in
fire.Fire(local_chat)
File "/home/user/miniconda3/envs/ktransformers/lib/python3.11/site-packages/fire/core.py", line 135, in Fire
component_trace = _Fire(component, args, parsed_flag_args, context, name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/envs/ktransformers/lib/python3.11/site-packages/fire/core.py", line 468, in _Fire
component, remaining_args = _CallAndUpdateTrace(
^^^^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/envs/ktransformers/lib/python3.11/site-packages/fire/core.py", line 684, in _CallAndUpdateTrace
component = fn(*varargs, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^
File "/data/wmh/ktransformers/ktransformers/local_chat.py", line 92, in local_chat
local_rank, world_size = setup_model_parallel(tp=tp)
^^^^^^^^^^^^^^^^^^^^
NameError: name 'setup_model_parallel' is not defined

经过排查，local_chat.py中，第16行import torch_npu报错，被异常捕捉跳过了setup_model_parallel导入，对于import torch_npu报错信息如下：

Traceback (most recent call last):
File "/data/wmh/ktransformers/ktransformers/local_chat.py", line 18, in
from ktransformers.util.ascend.ascend_utils import get_absort_weight, setup_model_parallel, get_tensor_parallel_group
File "/data/wmh/ktransformers/ktransformers/util/ascend/ascend_utils.py", line 5, in
import torch_npu
ModuleNotFoundError: No module named 'torch_npu'

对于torch_npu，搜索得知这个是华为显卡的torch，跟cuda版本torch是冲突的， 现在卡在这里了

### 复现步骤

python -m ktransformers.local_chat --model_path /data/wmh/model/qwen14b --gguf_path ./qwen14b-GGUF

用的模型是qwen2.5 14b

### 环境信息

==================== 系统信息 ====================
操作系统: Ubuntu 22.04.2 LTS
内核版本: 6.8.0-60-generic
架构: x86_64

==================== CPU 信息 ====================
Model name: Intel(R) Xeon(R) Platinum 8358 CPU @ 2.60GHz
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
逻辑 CPU 核心数: 128

==================== GPU 信息 ====================
NVIDIA A800-SXM4-80GB, 560.35.05, 81920, 4, 16

Start with ktransformers/local_chat.py around lines 16, 92, and 258, then inspect ktransformers/util/ascend/ascend_utils.py line 5. Reproduce the failure with the provided local_chat command in the listed CUDA environment and trace how the torch_npu import exception affects setup_model_parallel. Done means the CUDA environment no longer fails with the reported undefined name, with the relevant behavior covered by tests or a verified reproduction.

Written by the indexing model from the issue text.