Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,112 @@
|
|
1 |
-
---
|
2 |
-
license: apache-2.0
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
---
|
4 |
+
|
5 |
+
### Inference with Huggingface's Transformers
|
6 |
+
You can directly employ [Huggingface's Transformers](https://github.com/huggingface/transformers) for model inference.
|
7 |
+
|
8 |
+
#### Code Completion
|
9 |
+
```python
|
10 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
11 |
+
import torch
|
12 |
+
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True)
|
13 |
+
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
|
14 |
+
input_text = "#write a quick sort algorithm"
|
15 |
+
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
|
16 |
+
outputs = model.generate(**inputs, max_length=128)
|
17 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
18 |
+
```
|
19 |
+
|
20 |
+
#### Code Insertion
|
21 |
+
```python
|
22 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
23 |
+
import torch
|
24 |
+
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True)
|
25 |
+
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
|
26 |
+
input_text = """<|fim▁begin|>def quick_sort(arr):
|
27 |
+
if len(arr) <= 1:
|
28 |
+
return arr
|
29 |
+
pivot = arr[0]
|
30 |
+
left = []
|
31 |
+
right = []
|
32 |
+
<|fim▁hole|>
|
33 |
+
if arr[i] < pivot:
|
34 |
+
left.append(arr[i])
|
35 |
+
else:
|
36 |
+
right.append(arr[i])
|
37 |
+
return quick_sort(left) + [pivot] + quick_sort(right)<|fim▁end|>"""
|
38 |
+
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
|
39 |
+
outputs = model.generate(**inputs, max_length=128)
|
40 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True)[len(input_text):])
|
41 |
+
```
|
42 |
+
|
43 |
+
#### Chat Completion
|
44 |
+
|
45 |
+
```python
|
46 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
47 |
+
import torch
|
48 |
+
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", trust_remote_code=True)
|
49 |
+
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
|
50 |
+
messages=[
|
51 |
+
{ 'role': 'user', 'content': "write a quick sort algorithm in python."}
|
52 |
+
]
|
53 |
+
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
|
54 |
+
# tokenizer.eos_token_id is the id of <|EOT|> token
|
55 |
+
outputs = model.generate(inputs, max_new_tokens=512, do_sample=False, top_k=50, top_p=0.95, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id)
|
56 |
+
print(tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True))
|
57 |
+
```
|
58 |
+
|
59 |
+
|
60 |
+
|
61 |
+
The complete chat template can be found within `tokenizer_config.json` located in the huggingface model repository.
|
62 |
+
|
63 |
+
An example of chat template is as belows:
|
64 |
+
|
65 |
+
```bash
|
66 |
+
<|begin▁of▁sentence|>User: {user_message_1}
|
67 |
+
|
68 |
+
Assistant: {assistant_message_1}<|end▁of▁sentence|>User: {user_message_2}
|
69 |
+
|
70 |
+
Assistant:
|
71 |
+
```
|
72 |
+
|
73 |
+
You can also add an optional system message:
|
74 |
+
|
75 |
+
```bash
|
76 |
+
<|begin▁of▁sentence|>{system_message}
|
77 |
+
|
78 |
+
User: {user_message_1}
|
79 |
+
|
80 |
+
Assistant: {assistant_message_1}<|end▁of▁sentence|>User: {user_message_2}
|
81 |
+
|
82 |
+
Assistant:
|
83 |
+
```
|
84 |
+
|
85 |
+
### Inference with vLLM (recommended)
|
86 |
+
To utilize [vLLM](https://github.com/vllm-project/vllm) for model inference, please merge this Pull Request into your vLLM codebase: https://github.com/vllm-project/vllm/pull/4650.
|
87 |
+
|
88 |
+
```python
|
89 |
+
from transformers import AutoTokenizer
|
90 |
+
from vllm import LLM, SamplingParams
|
91 |
+
|
92 |
+
max_model_len, tp_size = 8192, 1
|
93 |
+
model_name = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct"
|
94 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
95 |
+
llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True, enforce_eager=True)
|
96 |
+
sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])
|
97 |
+
|
98 |
+
messages_list = [
|
99 |
+
[{"role": "user", "content": "Who are you?"}],
|
100 |
+
[{"role": "user", "content": "write a quick sort algorithm in python."}],
|
101 |
+
[{"role": "user", "content": "Write a piece of quicksort code in C++."}],
|
102 |
+
]
|
103 |
+
|
104 |
+
prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list]
|
105 |
+
|
106 |
+
outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)
|
107 |
+
|
108 |
+
generated_text = [output.outputs[0].text for output in outputs]
|
109 |
+
print(generated_text)
|
110 |
+
```
|
111 |
+
|
112 |
+
|