[BUG]: Colossalai-OpenMoE-8b : loss value is very large and cannot converge

[hangchen426926]

🐛 Describe the bug

I am currently running the Colossalai/examples/language/openmoe project with the following experimental setup:

datasets: load_dataset("yizhongw/self_instruct/data/finetuning/self_instruct_221203", "super_natural_instructions"),
model: openmoe-8b
A100_num_GPU：8
Epoch：3
BatchSize：4
LR：0.00001
zero_stage：1
precision：bf16
Boost plugin：ep_zero
extra_dp_size：2
max_length：2048

### Issue : loss value is very large and cannot converge
I've encountered challenges during the training process where the convergence seems unachievable. The training loss value persists at an exceptionally high level (exceeding 2.3E+10) even after running 3 epochs. The logged information provided below showcases loss values:

Furthermore, in the default setup of the openmoe project, the lr is 0.00001. Considering this, I suspect the issue might stem from an overfitting problem. Consequently, I attempted to adjustment the lr parameter(from 0.00001 to 0.0000001) ,but still encountered the same problem.

Environment

torch 1.13.1
python 3.8.17
cuda:11.7.17

[Orion-Zheng]

Thank you for your valuable feedback! 😃We are working on this bug and will get back to you later.

[noob-ctrl]

I also encountered this problem

[noob-ctrl]

@Orion-Zheng Has this bug been solved?