Fix CUDA OOM when creating Mixtral checkpoint

[VivekBits2210]

Move w1,w2,w3 tensors to cpu before stacking.

[byshiue]

Thank you for the report. But forcing the tensor on cpu might lead to cpu OOM on other environment. Have you tried using --load_model_on_cpu during converting the checkpoint.

[nv-guomingz]

Hi @VivekBits2210 , do you manage to create mixtral checkpoint with @byshiue 's suggestion?
If so, I'd like to close this PR at current stage.

[ghost]

Have you tried using --load_model_on_cpu during converting the checkpoint.
This won't work for some quantization cases where you need to run the model on some calibration dataset. The model cannot be loaded on CPU.