Support Constrained Server Memory

[drunkcoding]

Colab server T4 has 12GB DRAM, 16GB GPU, quantized mixtral has 26GB in size with single checkpoint, cannot bot be loaded into memory on creating the custom format for offloading