[BUG]: AttributeError: type object 'ColoParameter' has no attribute 'from_torch_tensor' when run hybrid_parallel example

[ztorchan]

🐛 Describe the bug

I noticed that from_torch_tensor method of class ColoParameter and ColoTensor have been removed in PR #4479
(colossalai/tensor/colo_parameter.py, colossalai/tensor/colo_tensor.py).

But this method was still called under file ColossalAI/colossalai/legacy/pipeline/pipelinable.py

ColossalAI/colossalai/legacy/pipeline/pipelinable.py

Line 120 in 641b1ee

setattr(module, name, ColoParameter.from_torch_tensor(tensor=param.data, requires_grad=param.requires_grad))

which cause error:
AttributeError: type object 'ColoParameter' has no attribute 'from_torch_tensor'

Environment

python 3.10.10
cuda 11.6.1
torch 1.12.1

[char-1ee]

The legacy directory is deprecated now. Some code that uses dependencies from legacy is out-of-date. You can try to refer to the new implementations.

[awer-A]

@char-1ee When I use the latest version of Colossalai, I get an error "ModuleNotFoundError: No module named 'colossalai.context.moe_context'". I don't know which version of Colossalai I should use. Do you have any good suggestions?

[char-1ee]

@awer-A You can try install colossalai from source. I suggest if you want to set up a distributed training environment with ColossalAI on your server, you can follow the instructions in ColossalAI's docker file https://github.com/hpcaitech/ColossalAI/blob/main/docker/Dockerfile, step by step.