lisa_grounding单卡顺畅，多卡就会卡在模型初始化

[GGbond-study]

单卡A800运行的时候就正常的，能够进行训练。但是转化为2张卡时候，就会卡在模型初始化这一步，也不报错。输出：
[2025-09-03 17:03:50,412] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)
INFO 09-03 17:03:52 init.py:190] Automatically detected platform cuda.
INFO 09-03 17:03:52 init.py:190] Automatically detected platform cuda.
[2025-09-03 17:03:53,038] [INFO] [comm.py:652:init_distributed] cdb=None
[2025-09-03 17:03:53,038] [INFO] [comm.py:683:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl
[2025-09-03 17:03:53,103] [INFO] [config.py:733:init] Config mesh_device None world_size = 2
[2025-09-03 17:03:53,164] [INFO] [comm.py:652:init_distributed] cdb=None
[2025-09-03 17:03:53,398] [INFO] [config.py:733:init] Config mesh_device None world_size = 2
经过print大法就会卡在
if "Qwen2-VL" in model_id:
----》model = Qwen2VLForConditionalGeneration.from_pretrained(model, **model_init_kwargs)
elif "Qwen2.5-VL" in model_id:

[tanjjjjrr]

没设置分布式训练吧应该，有个distributed的参数

[Re-dot-art]

单卡A800运行的时候就正常的，能够进行训练。但是转化为2张卡时候，就会卡在模型初始化这一步，也不报错。输出： [2025-09-03 17:03:50,412] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect) INFO 09-03 17:03:52 init.py:190] Automatically detected platform cuda. INFO 09-03 17:03:52 init.py:190] Automatically detected platform cuda. [2025-09-03 17:03:53,038] [INFO] [comm.py:652:init_distributed] cdb=None [2025-09-03 17:03:53,038] [INFO] [comm.py:683:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl [2025-09-03 17:03:53,103] [INFO] [config.py:733:init] Config mesh_device None world_size = 2 [2025-09-03 17:03:53,164] [INFO] [comm.py:652:init_distributed] cdb=None [2025-09-03 17:03:53,398] [INFO] [config.py:733:init] Config mesh_device None world_size = 2 经过print大法就会卡在 if "Qwen2-VL" in model_id: ----》model = Qwen2VLForConditionalGeneration.from_pretrained(model, **model_init_kwargs) elif "Qwen2.5-VL" in model_id:

请问您lisa_train_sft.json 这个数据集文件是怎么构建的呀，感谢

[fei-chang]

单卡A800运行的时候就正常的，能够进行训练。但是转化为2张卡时候，就会卡在模型初始化这一步，也不报错。输出： [2025-09-03 17:03:50,412] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect) INFO 09-03 17:03:52 init.py:190] Automatically detected platform cuda. INFO 09-03 17:03:52 init.py:190] Automatically detected platform cuda. [2025-09-03 17:03:53,038] [INFO] [comm.py:652:init_distributed] cdb=None [2025-09-03 17:03:53,038] [INFO] [comm.py:683:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl [2025-09-03 17:03:53,103] [INFO] [config.py:733:init] Config mesh_device None world_size = 2 [2025-09-03 17:03:53,164] [INFO] [comm.py:652:init_distributed] cdb=None [2025-09-03 17:03:53,398] [INFO] [config.py:733:init] Config mesh_device None world_size = 2 经过print大法就会卡在 if "Qwen2-VL" in model_id: ----》model = Qwen2VLForConditionalGeneration.from_pretrained(model, **model_init_kwargs) elif "Qwen2.5-VL" in model_id:

你好，请问最后解决了么？如何解决的呢？我这里似乎遇到了同样的问题。