Dev update transformers

[yaoguany]

1. can't use CUDA_VISIBLE_DEVICES with new DeepSpeed, need to use deepspeed --include localhost:0,1
   CUDA_VISIBLE_DEVICES isn't being respected / hostfile doesn't quite work for one node microsoft/DeepSpeed#662
2. hf scheduler + ds optimizer no longer supported, need to rewrite deepspeed config
   ValueError: Found optimizer configured in the DeepSpeed config, but no scheduler. Please configure a scheduler in the DeepSpeed config. huggingface/transformers#24359 (comment)

[research4pan]

LGTM, we may fix the issues caused by this transformers version update later. Currently main branch is not stable, we recommend everyone to use v0.0.5 for now.

Potential issues

1. CUDA_VISIBLE_DEVICES issue: CUDA_VISIBLE_DEVICES is not respected in latest deepspeed==0.10.0. Please specify gpus with fashion such as deepspeed --include localhost:0,1,2,3.
2. hf/deepspeed scheduler optimizer combo issue: hf scheduler + ds optimizer is no longer supported according to restriction in transformers after version 4.29.2. This update happens in transformers==4.29.2 to transformers==4.30.0
3. qlora multi-GPU issue: qlora currently may not support multiple GPUs.
4. qlora merging issue: qlora currently may not support merging into original model.
5. deepspeed zero3 issue: need modify some parts of configs/ds_config_zero3.json to smoothly utilize zero optimization stage 3 in deepspeed after update.

README.md

• [Style] use bash install.sh is okay. As we have shebang line in install.sh, the original usage is also okay.