Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,81 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: openrail
|
3 |
+
datasets:
|
4 |
+
- shareAI/ShareGPT-Chinese-English-90k
|
5 |
+
- shareAI/CodeChat
|
6 |
+
language:
|
7 |
+
- zh
|
8 |
+
- en
|
9 |
+
library_name: transformers
|
10 |
+
tags:
|
11 |
+
- code
|
12 |
+
- chat
|
13 |
+
- codellama
|
14 |
+
- copilot
|
15 |
+
- codeAI
|
16 |
+
pipeline_tag: question-answering
|
17 |
+
---
|
18 |
+
|
19 |
+
## CodeLlaMa模型的中文化版本 (支持多轮对话)
|
20 |
+
|
21 |
+
科普:CodeLlaMa是专门用于代码助手的,与ChineseLlaMa不同,适用于代码类问题的回复。
|
22 |
+
用于多轮对话的推理代码:
|
23 |
+
(可以直接复制运行,默认会自动拉取该模型权重)
|
24 |
+
|
25 |
+
关联Github仓库:https://github.com/CrazyBoyM/CodeLLaMA-chat
|
26 |
+
|
27 |
+
```
|
28 |
+
# from Firefly
|
29 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
30 |
+
import torch
|
31 |
+
|
32 |
+
|
33 |
+
def main():
|
34 |
+
model_name = 'shareAI/CodeLLaMA-chat-13b-Chinese'
|
35 |
+
|
36 |
+
device = 'cuda'
|
37 |
+
max_new_tokens = 500 # 每轮对话最多生成多少个token
|
38 |
+
history_max_len = 1000 # 模型记忆的最大token长度
|
39 |
+
top_p = 0.9
|
40 |
+
temperature = 0.35
|
41 |
+
repetition_penalty = 1.0
|
42 |
+
|
43 |
+
model = AutoModelForCausalLM.from_pretrained(
|
44 |
+
model_name,
|
45 |
+
trust_remote_code=True,
|
46 |
+
low_cpu_mem_usage=True,
|
47 |
+
torch_dtype=torch.float16,
|
48 |
+
device_map='auto'
|
49 |
+
).to(device).eval()
|
50 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
51 |
+
model_name,
|
52 |
+
trust_remote_code=True,
|
53 |
+
use_fast=False
|
54 |
+
)
|
55 |
+
|
56 |
+
|
57 |
+
history_token_ids = torch.tensor([[]], dtype=torch.long)
|
58 |
+
|
59 |
+
user_input = input('User:')
|
60 |
+
while True:
|
61 |
+
input_ids = tokenizer(user_input, return_tensors="pt", add_special_tokens=False).input_ids
|
62 |
+
eos_token_id = torch.tensor([[tokenizer.eos_token_id]], dtype=torch.long)
|
63 |
+
user_input_ids = torch.concat([input_ids, eos_token_id], dim=1)
|
64 |
+
history_token_ids = torch.concat((history_token_ids, user_input_ids), dim=1)
|
65 |
+
model_input_ids = history_token_ids[:, -history_max_len:].to(device)
|
66 |
+
with torch.no_grad():
|
67 |
+
outputs = model.generate(
|
68 |
+
input_ids=model_input_ids, max_new_tokens=max_new_tokens, do_sample=True, top_p=top_p,
|
69 |
+
temperature=temperature, repetition_penalty=repetition_penalty, eos_token_id=tokenizer.eos_token_id
|
70 |
+
)
|
71 |
+
model_input_ids_len = model_input_ids.size(1)
|
72 |
+
response_ids = outputs[:, model_input_ids_len:]
|
73 |
+
history_token_ids = torch.concat((history_token_ids, response_ids.cpu()), dim=1)
|
74 |
+
response = tokenizer.batch_decode(response_ids)
|
75 |
+
print("Bot:" + response[0].strip().replace(tokenizer.eos_token, ""))
|
76 |
+
user_input = input('User:')
|
77 |
+
|
78 |
+
|
79 |
+
if __name__ == '__main__':
|
80 |
+
main()
|
81 |
+
```
|