improvement(tools): optimize convert-pth-to-ggml

[tpoisonooo]

optimize convert tool with argparse

$ python3 convert-pth-to-ggml.py -h
usage: convert-pth-to-ggml.py [-h] dir_model {f32,f16} out_dir

Convert ckpt models to ggml models. For example: python3 convert-pth-to-ggml.py ../llama-models/7B/ f32 models/llama-7B

positional arguments:
  dir_model   Directory path of the checkpoint model
  {f32,f16}   Data type of the converted tensor, f32 or f16
  out_dir     Directory path for storing ggml model

options:
  -h, --help  show this help message and exit

Tested on 7B/30B models, it works well.

$ tree models/
models/
├── 7B
├── llama-30B
│   ├── ggml-model-f16.bin
│   ├── ggml-model-f16.bin.1
│   ├── ggml-model-f16.bin.2
│   └── ggml-model-f16.bin.3
└── llama-7B
    ├── ggml-model.bin -> ggml-model-f16.bin
    ├── ggml-model-f16.bin
    └── ggml-model-f32.bin

[tpoisonooo]

cc @ggerganov

[tpoisonooo]

Conflict fixed and tested on 7B/30B.

master version diffs with 3 lines here :

if os.path.exists(fname_out):
    print(f"Skip conversion, it already exists: {fname_out}")
    sys.exit(0)

cc @gjmulder

[sw]

os is already imported

[sw]

can you please update the readme? the choice "1" for ftype will no longer work.

[sw]

This is almost a duplicate of #109

[ggerganov]

Decided to go with #109