Fix save_model

[ypwhs]

The original save_model function saved all parameters, resulting in a file size of 6GB, while the actual Lora model is only 200MB and does not require such a large file. The modified save_model function uses the save_tunable_parameters function, which saves only the tunable parameters and reduces the saved model size to 200MB.

Before:

(base) ➜  ChatGLM-Tuning tree -h exp1
ykx_5k
├── [224M]  chatglm-lora.pt
├── [4.0K]  checkpoint-1000
│   ├── [448M]  optimizer.pt
│   ├── [6.6G]  pytorch_model.bin
│   ├── [ 14K]  rng_state_0.pth
│   ├── [ 14K]  rng_state_1.pth
│   ├── [ 14K]  rng_state_2.pth
│   ├── [ 14K]  rng_state_3.pth
│   ├── [ 14K]  rng_state_4.pth
│   ├── [ 14K]  rng_state_5.pth
│   ├── [ 14K]  rng_state_6.pth
│   ├── [ 14K]  rng_state_7.pth
│   ├── [ 557]  scaler.pt
│   ├── [ 627]  scheduler.pt
│   ├── [2.5K]  trainer_state.json
│   └── [3.5K]  training_args.bin
......

After:

(base) ➜  ChatGLM-Tuning tree -h exp2
.
├── [4.0K]  checkpoint-1000
│   ├── [224M]  chatglm-lora.pt
│   ├── [448M]  optimizer.pt
│   ├── [ 14K]  rng_state_0.pth
│   ├── [ 14K]  rng_state_1.pth
│   ├── [ 14K]  rng_state_2.pth
│   ├── [ 14K]  rng_state_3.pth
│   ├── [ 14K]  rng_state_4.pth
│   ├── [ 14K]  rng_state_5.pth
│   ├── [ 14K]  rng_state_6.pth
│   ├── [ 14K]  rng_state_7.pth
│   ├── [ 557]  scaler.pt
│   ├── [ 627]  scheduler.pt
│   ├── [2.6K]  trainer_state.json
│   └── [3.5K]  training_args.bin
......

[mymusise]

Thanks for this PR, I will merge it after test.