📚 文档问题描述

文档链接：https://gitcode.com/Ascend/MindSpeed-MM/blob/26.0.0/examples/qwen3_5/README.md
1、该readme中明确提示：torch和torch_npu版本推荐2.7.1版本，CANN推荐使用8.5.2版本；
2、但是随后的“‼️MoE部分的加速特性依赖较新版本的torch_npu和CANN，推荐使用以下版本”处附带的链接是cann9.0.0和Torch-npu 2.7.1.post2的
3、昇腾镜像仓库中26.0.0的镜像预置的是cann9.0.0 和 torch-npu 2.7.1.post4，这是符合torch官网的配套推荐的：https://gitcode.com/Ascend/pytorch/blob/master/COMPATIBILITY.md
4、一键安装脚本bash scripts/install.sh --msid eb10b92 && bash examples/qwen3_5/install_extensions.sh执行后提示会安装torch-npu 2.7.1，提示信息如下
=== torch_npu Version Mismatch ===
Currently installed torch_npu version: 2.7.1.post4
Target version: 2.7.1
Reinstall torch_npu to match PyTorch version? (y/n)

以上4点提到的cann版本和torch-npu版本无法对齐，令人疑惑，看不懂仓上推荐的到底是哪个版本的cann和torch-npu

建议的替代/修复方案

请前后统一，明确具体推荐的版本，并充分考虑torch官网推荐的配套关系

**信息补充：**安装readme给的cann链接中的cann9.0.0后，执行一键式安装脚本会安装triton-ascend==3.2.0，拉起会报错，报错信息如下：
“““
Traceback (most recent call last):
File "/workspace/MindSpeed-MM/mindspeed_mm/fsdp/models/qwen3_5/triton/utils.py", line 132, in get_available_device
return triton.runtime.driver.active.get_current_target().backend
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/conda/lib/python3.11/site-packages/triton/runtime/driver.py", line 23, in getattr
self._initialize_obj()
File "/opt/conda/lib/python3.11/site-packages/triton/runtime/driver.py", line 20, in _initialize_obj
self._obj = self._init_fn()
^^^^^^^^^^^^^^^
File "/opt/conda/lib/python3.11/site-packages/triton/runtime/driver.py", line 9, in _create_driver
return actives0
^^^^^^^^^^^^
File "/opt/conda/lib/python3.11/site-packages/triton/backends/ascend/driver.py", line 162, in init
self.utils = NPUUtils()
^^^^^^^^^^
File "/opt/conda/lib/python3.11/site-packages/triton/backends/ascend/driver.py", line 65, in init
so = _build_npu_ext("npu_utils", None, tmp_src_path)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/conda/lib/python3.11/site-packages/triton/backends/ascend/utils.py", line 413, in _build_npu_ext
raise RuntimeError(f"Failed to compile {src_path}, error: {result.stderr}")
RuntimeError: Failed to compile /tmp/tmpp5tkumuk/npu_utils.cpp, error: /tmp/tmpp5tkumuk/npu_utils.cpp:321:44: error: ‘RT_LIMIT_TYPE_SIMT_WARP_STACK_SIZE’ is not a member of ‘rtLimitType_t’ {aka ‘tagRtLimitType’}
321 | {"WARP_STACK_SIZE", rtLimitType_t::RT_LIMIT_TYPE_SIMT_WARP_STACK_SIZE},
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/tmp/tmpp5tkumuk/npu_utils.cpp:324:3: error: could not convert ‘{{"LOW_POWER_TIMEOUT", RT_LIMIT_TYPE_LOW_POWER_TIMEOUT}, {"WARP_STACK_SIZE", }, {"DVG_WARP_STACK_SIZE", RT_LIMIT_TYPE_SIMT_DVG_WARP_STACK_SIZE}, {"STACK_SIZE", RT_LIMIT_TYPE_STACK_SIZE}}’ from ‘’ to ‘const std::unordered_map<std::__cxx11::basic_string, tagRtLimitType>’
324 | };
| ^
| |
|

”””
这是因为triton-ascend 3.2.0与cann9.0.0版本不匹配，升级triton-ascend会引入参数不匹配的报错，报错如下：
“”“”
File "/workspace/MindSpeed-MM/mindspeed_mm/fsdp/models/qwen3_5/triton/utils.py", line 330, in get_autotune_config
triton.Config(
triton.Config(
TypeError: TypeErrorConfig.init() got an unexpected keyword argument 'multibuffer'
: Config.init() got an unexpected keyword argument 'multibuffer'
“”“”
将版本修改为cann8.5.0+triton-ascend3.2.0后才能拉起。