【feature advice】Int8 mode to run original model

[LiuLinyun]

when using LoRA to fine-tune, the original model parameters are locked, which can be converted to INT8 mode to inference, and other trainable parameters with fp16/bf16/fp32/tf32 mode. It is just like what peft library do.

Uing INT8 to run original model will save GPU memory and accelerate training speed.