[Bug] Windows (RTX 50 / one-click + RTX50 env): KeyError: 'None-triton' in MM_WEIGHT_REGISTER during T5 offload attention init

[GOJiong]

Describe the bug
When using the Windows one-click install package combined with the "50 series environment package", launching the Gradio interface and attempting to generate results in a KeyError: 'None-triton'. The crash occurs specifically during the T5 offloaded attention initialization for the Wan model.

To Reproduce
Steps to reproduce the behavior:

1. Use Windows with an RTX 50-series GPU
2. Download and extract the Windows one-click install package.
3. Replace the env folder with the "50 series environment package".
4. Launch the web UI via start_lightx2v.bat.
5. Attempt a generation using a model that triggers the Wan T5 text encoder.
6. See the error below.

Error Traceback

File "D:\AIgradio_app_v2\gradio_app_v2\LightX2V\lightx2v\models\input_encoders\hf\wan\t5\model.py", line 478, in __init__
  self.blocks_weights = T5offloadBlocksweights(num_layers, quant_scheme, lazy_load, lazy_load_path)

File "D:\AIgradio_app_v2\gradio_app_v2\LightX2V\lightx2v\models\input_encoders\hf\wan\t5\model.py", line 56, in __init__
  [T5offloadselfAttention(i, mm_type, create_cuda_buffer=True, create_cpu_buffer=False, lazy_load=lazy_load, lazy_load_path=lazy_load_path) for i in range(1)]

File "D:\AIgradio_app_v2\gradio_app_v2\LightX2V\lightx2v\models\input_encoders\hf\wan\t5\model.py", line 56, in <listcomp>
  [T5offloadselfAttention(i, mm_type, create_cuda_buffer=True, create_cpu_buffer=False, lazy_load=lazy_load, lazy_load_path=lazy_load_path) for i in range(1)]

File "D:\AIgradio_app_v2\gradio_app_v2\LightX2V\lightx2v\models\input_encoders\hf\wan\t5\model.py", line 92, in __init__
  T5offloadAttention(block_index, block_prefix, mm_type, create_cuda_buffer, create_cpu_buffer, lazy_load, lazy_load_path)

File "D:\AIgradio_app_v2\gradio_app_v2\LightX2V\lightx2v\models\input_encoders\hf\wan\t5\model.py", line 106, in __init__
  MM_WEIGHT_REGISTER[self.mm_type](f"{block_prefix}.self.block_index.attn.q.Weight", create_cuda_buffer, create_cpu_buffer, lazy_load, lazy_load_path), None

File "D:\AIgradio_app_v2\gradio_app_v2\LightX2V\lightx2v\utils\registry_factory.py", line 32, in __getitem__
  return self._dict[key]

KeyError: 'None-triton'

Environment details:

• OS: Windows 11
• GPU: RTX 5070
• Installation Method: Windows One-Click + 50 Series env package
• Likely Missing Dependency: triton-windows might be missing or failing to initialize in the provided 50-series environment.

[gushiqiao]

Run pip install triton-windows, and replace the LightX2V directory in the installed package with the latest code.