Update README.md
Browse files
README.md
CHANGED
@@ -43,4 +43,85 @@ https://twitter.com/ManuelFaysse/status/1706949891358859624
|
|
43 |
|
44 |
### Usage
|
45 |
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
44 |
### Usage
|
45 |
|
46 |
+
Probably something like the Llama2 models from the non-hf release ! Obviously, the model is probably not completely similar to Llama, so conversion to HF will not be so direct.
|
47 |
+
Let's figure this out together !
|
48 |
+
|
49 |
+
```bash
|
50 |
+
torchrun --nproc_per_node 1 example_text_completion.py \
|
51 |
+
--ckpt_dir llama-2-7b/ \
|
52 |
+
--tokenizer_path tokenizer.model \
|
53 |
+
--max_seq_len 128 --max_batch_size 4
|
54 |
+
```
|
55 |
+
|
56 |
+
`example_text_completion.py`
|
57 |
+
```python
|
58 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
59 |
+
# This software may be used and distributed according to the terms of the Llama 2 Community License Agreement.
|
60 |
+
|
61 |
+
import fire
|
62 |
+
|
63 |
+
from llama import Llama
|
64 |
+
from typing import List
|
65 |
+
|
66 |
+
def main(
|
67 |
+
ckpt_dir: str,
|
68 |
+
tokenizer_path: str,
|
69 |
+
temperature: float = 0.6,
|
70 |
+
top_p: float = 0.9,
|
71 |
+
max_seq_len: int = 128,
|
72 |
+
max_gen_len: int = 64,
|
73 |
+
max_batch_size: int = 4,
|
74 |
+
):
|
75 |
+
"""
|
76 |
+
Entry point of the program for generating text using a pretrained model.
|
77 |
+
|
78 |
+
Args:
|
79 |
+
ckpt_dir (str): The directory containing checkpoint files for the pretrained model.
|
80 |
+
tokenizer_path (str): The path to the tokenizer model used for text encoding/decoding.
|
81 |
+
temperature (float, optional): The temperature value for controlling randomness in generation.
|
82 |
+
Defaults to 0.6.
|
83 |
+
top_p (float, optional): The top-p sampling parameter for controlling diversity in generation.
|
84 |
+
Defaults to 0.9.
|
85 |
+
max_seq_len (int, optional): The maximum sequence length for input prompts. Defaults to 128.
|
86 |
+
max_gen_len (int, optional): The maximum length of generated sequences. Defaults to 64.
|
87 |
+
max_batch_size (int, optional): The maximum batch size for generating sequences. Defaults to 4.
|
88 |
+
"""
|
89 |
+
generator = Llama.build(
|
90 |
+
ckpt_dir=ckpt_dir,
|
91 |
+
tokenizer_path=tokenizer_path,
|
92 |
+
max_seq_len=max_seq_len,
|
93 |
+
max_batch_size=max_batch_size,
|
94 |
+
)
|
95 |
+
|
96 |
+
prompts: List[str] = [
|
97 |
+
# For these prompts, the expected answer is the natural continuation of the prompt
|
98 |
+
"I believe the meaning of life is",
|
99 |
+
"Simply put, the theory of relativity states that ",
|
100 |
+
"""A brief message congratulating the team on the launch:
|
101 |
+
|
102 |
+
Hi everyone,
|
103 |
+
|
104 |
+
I just """,
|
105 |
+
# Few shot prompt (providing a few examples before asking model to complete more);
|
106 |
+
"""Translate English to French:
|
107 |
+
|
108 |
+
sea otter => loutre de mer
|
109 |
+
peppermint => menthe poivrée
|
110 |
+
plush girafe => girafe peluche
|
111 |
+
cheese =>""",
|
112 |
+
]
|
113 |
+
results = generator.text_completion(
|
114 |
+
prompts,
|
115 |
+
max_gen_len=max_gen_len,
|
116 |
+
temperature=temperature,
|
117 |
+
top_p=top_p,
|
118 |
+
)
|
119 |
+
for prompt, result in zip(prompts, results):
|
120 |
+
print(prompt)
|
121 |
+
print(f"> {result['generation']}")
|
122 |
+
print("\n==================================\n")
|
123 |
+
|
124 |
+
|
125 |
+
if __name__ == "__main__":
|
126 |
+
fire.Fire(main)
|
127 |
+
```
|