[BUG]: applications/Chat/coati/trainer/strategies /colossalai.py is not compatible with the gemini_plugin.py

[alphanlp]

🐛 Describe the bug

  File "/data/llmodel/huap/ColossalAI/applications/Chat/train_sft.py", line 205, in <module>
    train(args)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/train_sft.py", line 32, in train
    strategy = GeminiStrategy(placement_policy='cuda')
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/colossalai.py", line 194, in __init__
    super().__init__(seed, plugin_initializer)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/ddp.py", line 42, in __init__
    super().__init__(plugin_initializer)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/base.py", line 29, in __init__
    self.plugin = plugin_initializer()
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/colossalai.py", line 169, in <lambda>
    plugin_initializer = lambda: GeminiPlugin(
TypeError: __init__() got an unexpected keyword argument 'device'

Can somebody can give a correct Example？

Environment

No response

[flybird11111]

super().init(plugin_initializer)

HIHI, Can you please upgrade Coati？This will resolve your issue.

[cwher]

Some recommended modifications are provided in #4685, yet it may still encounter some errors.

We are working on solving this issue, and @flybird11111 will be in charge.

[flybird11111]

🐛 Describe the bug

  File "/data/llmodel/huap/ColossalAI/applications/Chat/train_sft.py", line 205, in <module>
    train(args)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/train_sft.py", line 32, in train
    strategy = GeminiStrategy(placement_policy='cuda')
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/colossalai.py", line 194, in __init__
    super().__init__(seed, plugin_initializer)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/ddp.py", line 42, in __init__
    super().__init__(plugin_initializer)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/base.py", line 29, in __init__
    self.plugin = plugin_initializer()
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/colossalai.py", line 169, in <lambda>
    plugin_initializer = lambda: GeminiPlugin(
TypeError: __init__() got an unexpected keyword argument 'device'

Can somebody can give a correct Example？

Environment

No response

how about trying 'pip install colossalai==0.3.2'

[alphanlp]

🐛 Describe the bug

  File "/data/llmodel/huap/ColossalAI/applications/Chat/train_sft.py", line 205, in <module>
    train(args)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/train_sft.py", line 32, in train
    strategy = GeminiStrategy(placement_policy='cuda')
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/colossalai.py", line 194, in __init__
    super().__init__(seed, plugin_initializer)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/ddp.py", line 42, in __init__
    super().__init__(plugin_initializer)
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/base.py", line 29, in __init__
    self.plugin = plugin_initializer()
  File "/data/llmodel/huap/ColossalAI/applications/Chat/coati/trainer/strategies/colossalai.py", line 169, in <lambda>
    plugin_initializer = lambda: GeminiPlugin(
TypeError: __init__() got an unexpected keyword argument 'device'

Can somebody can give a correct Example？

Environment

No response

how about trying 'pip install colossalai==0.3.2'

yes , i use colossalai==0.3.2

but the code in colossalai.py is :
plugin_initializer = lambda: GeminiPlugin(
# gemini_config
device=get_current_device(),
placement_policy=placement_policy,
precision='fp16',
pin_memory=pin_memory,
force_outputs_fp32=force_outputs_fp32,
strict_ddp_mode=shard_init,
search_range_m=search_range_m,
hidden_dim=hidden_dim,
min_chunk_size_m=min_chunk_size_m,
# zero_optim_config
gpu_margin_mem_ratio=gpu_margin_mem_ratio,
# optim_config
initial_scale=initial_scale,
growth_factor=growth_factor,

[flybird11111]

hi，This issue is being fixed, please follow this pull request.#4698

[cwher]

This issue is fixed now. You can pull the latest code from main branch to fix it.
However, there still exists some other compatibility issues with Gemini plugin. For instance, Gemini is not compatible with model.from_pretrained yet, and we are working on fixing it.