Lada在AMD 7000系显卡win10下的运行

asas99

  本文基于@fish313331（https://javpcn.com/forum.php?mod=viewthread&tid=1324）、@gbywreyuerh@out、@fish313331Amd卡在Win11下运用显卡原生加速运行Lada教程，简单易懂！！ - JavPlayer - JavPlayer 中文交流论坛 -教程讨论lada在win10，7900xtx平台下的安装与使用情况。

@gbywreyuerh@out大佬在其教程中已写明具体操作，但lada在github删库以及大佬的安装代码是针对amd9000系，现打算针对现阶段7000系显卡的运行进行补充。
步骤1、2、3不变，注意将需要的程序添加到环境变量，否则无法正常运行。库由https://github.com/ladaapp/lada替换为https://codeberg.org/ladaapp/lada
步骤4 7000系使用以下代码

pip install `

https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/torch-2.9.0a0%2Brocm7.0.0rc20250908-cp313-cp313-win_amd64.whl`

https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/torchaudio-2.8.0a0%2Brocm7.0.0rc20250908-cp313-cp313-win_amd64.whl`

https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/torchvision-0.24.0a0%2Brocm7.0.0rc20250908-cp313-cp313-win_amd64.whl`

--extra-index-url https://rocm.nightlies.amd.com/v2/gfx110X-dgpu/

步骤5 将模型文件放入lada目录下的model_weights文件夹中。模型文件下载地址：

https://pan.quark.cn/s/6db69da7b91b

最后完成全部教程步骤，到测试阶段都不会报错。直到运行时可能会弹出以下报错：

Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "D:\Lada\lada\.venv\Scripts\lada-cli.exe\__main__.py", line 6, in <module>
    sys.exit(main())
             ~~~~^^
  File "D:\Lada\lada\lada\cli\main.py", line 151, in main
    mosaic_detection_model, mosaic_restoration_model, preferred_pad_mode = load_models(
                                                                           ~~~~~~~~~~~^
        args.device, args.mosaic_restoration_model, args.mosaic_restoration_model_path, args.mosaic_restoration_config_path,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
        args.mosaic_detection_model_path
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "D:\Lada\lada\lada\lib\frame_restorer.py", line 26, in load_models
    from lada.basicvsrpp.inference import load_model, get_default_gan_inference_config
  File "D:\Lada\lada\lada\basicvsrpp\inference.py", line 9, in <module>
    from mmengine.runner import load_checkpoint
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\runner\__init__.py", line 2, in <module>
    from ._flexible_runner import FlexibleRunner
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\runner\_flexible_runner.py", line 14, in <module>
    from mmengine._strategy import BaseStrategy
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\_strategy\__init__.py", line 3, in <module>
    from mmengine.utils.dl_utils import TORCH_VERSION
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\utils\dl_utils\__init__.py", line 8, in <module>
    from .time_counter import TimeCounter
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\utils\dl_utils\time_counter.py", line 8, in <module>
    from mmengine.dist.utils import master_only
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\dist\__init__.py", line 2, in <module>
    from .dist import (all_gather_object, all_reduce, all_gather, all_reduce_dict,
    ...<2 lines>...
                       collect_results_cpu, collect_results_gpu, all_reduce_params)
  File "D:\Lada\lada\.venv\Lib\site-packages\mmengine\dist\dist.py", line 26, in <module>
    def _get_reduce_op(name: str) -> torch_dist.ReduceOp:
                                     ^^^^^^^^^^^^^^^^^^^
AttributeError: module 'torch.distributed' has no attribute 'ReduceOp'

根据@fish313331大佬的方法

在.venv\Lib\site-packages\mmengine\model\wrappers\__init__.py开头加上

• import torch, types
• if not hasattr(torch, "distributed") or not hasattr(torch.distributed, "fsdp"):
•     torch.distributed = types.SimpleNamespace()
•     torch.distributed.fsdp = types.SimpleNamespace()
•     torch.distributed.fsdp.fully_sharded_data_parallel = types.SimpleNamespace()
  

然後後面這段

• if digit_version(TORCH_VERSION) >= digit_version('2.0.0'):
•     from .fully_sharded_distributed import \
•         MMFullyShardedDataParallel  # noqa:F401
•     __all__.append('MMFullyShardedDataParallel')
• 
  

改為

• if digit_version(TORCH_VERSION) >= digit_version('2.0.0'):
•     try:
•         from .fully_sharded_distributed import MMFullyShardedDataParallel  # noqa:F401
•     except Exception as e:
•         import warnings
•         warnings.warn(f"FSDP disabled: {e}")
•         MMFullyShardedDataParallel = None
•     __all__.append('MMFullyShardedDataParallel')
  

[color=rgb(51, 102, 153) !important]复制代码
另一個檔案.venv\Lib\site-packages\mmengine\dist\dist.py
在

• def _get_reduce_op(name: str) -> torch_dist.ReduceOp:
  

之前加上

• if not hasattr(torch.distributed, "ReduceOp"):
•     class DummyReduceOp:
•         SUM = None
•         MEAN = None
•     torch.distributed.ReduceOp = DummyReduceOp
  

至此lada即可在win10，7900xtx平台下正常使用

Moy

感谢哥

gbywreyuerh@out

不错，恭喜你也站在巨人的肩膀上二创了！

我就喜欢这种自己动脑子，会学习的人发布的贴子，而不是光一个劲只知道问问问。

老弟加油！

nanjogege

来学习了，a卡9070xt终于有用武之地了

jkzz123

学习一下

spicy

来学习了

Clerk1480

学习学习

rina5566

学习学习

clarks

大神，6000及以下显卡有解吗？

laoda0414

有点东西