File size: 4,469 Bytes
787f677
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
---
license: apache-2.0
datasets:
- Open-Orca/OpenOrca
- OpenAssistant/oasst_top1_2023-08-25
language:
- bg
- ca
- cs
- da
- de
- en
- es
- fr
- hr
- hu
- it
- nl
- pl
- pt
- ro
- ru
- sl
- sr
- sv
- uk

library_name: transformers
---


![image/png](https://cdn-uploads.huggingface.co/production/uploads/641b435ba5f876fe30c5ae0a/rJ1RxzuE-3gzgCppx-T8f.png)

```
reference-data-model:

  datasets:
    - OpenAssistant/oasst_top1_2023-08-25:
      lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk"
      link: https://huggingface.co/datasets/OpenAssistant/oasst_top1_2023-08-25

  model:
    - Open-Orca/Mistral-7B-OpenOrca
      Link:
        https://huggingface.co/Open-Orca/Mistral-7B-OpenOrca

  100 examples of generating:
    - Link:
      https://huggingface.co/NickyNicky/Mistral-7B-OpenOrca-oasst_top1_2023-08-25-v2/blob/main/output.xlsx

  Activated training with:
    - Link:
        https://huggingface.co/blog/tomaarsen/attention-sinks
        https://github.com/tomaarsen/attention_sinks
        https://arxiv.org/abs/2309.17453

  Version:
    - Link:
        https://huggingface.co/NickyNicky/Mistral-7B-OpenOrca-oasst_top1_2023-08-25-v1
        https://huggingface.co/NickyNicky/Mistral-7B-OpenOrca-oasst_top1_2023-08-25-v3

```


## 


```py
# attention-sinks
pip install attention_sinks

# flash-attn
!export CUDA_HOME=/usr/local/cuda-11.8
!MAX_JOBS=4 pip install flash-attn --no-build-isolation -qqq
!pip install git+"https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary" -qqq
```


## Version
```py
import torch, transformers,torchvision
torch.__version__,transformers.__version__, torchvision.__version__
#OUTPUTS: ('2.0.1+cu118', '4.34.0.dev0', '0.15.2+cu118')
```

## How to use
```py

from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    HfArgumentParser,
    TrainingArguments,
    pipeline,
    logging,
    GenerationConfig,
    TextIteratorStreamer,
)

from attention_sinks import AutoModelForCausalLM

import torch

# model_id = 'Open-Orca/Mistral-7B-OpenOrca'
model_id='NickyNicky/Mistral-7B-OpenOrca-oasst_top1_2023-08-25-v2'

model = AutoModelForCausalLM.from_pretrained(model_id,
                                             device_map="auto",
                                             trust_remote_code=True,
                                             torch_dtype=torch.bfloat16,
                                             load_in_4bit=True,
                                             low_cpu_mem_usage= True,

                                             attention_sink_size=4,
                                             attention_sink_window_size=1024, #512, # <- Low for the sake of faster generation
                                             )

max_length=2048
print("max_length",max_length)


tokenizer = AutoTokenizer.from_pretrained(model_id,
                                          # use_fast = False,
                                          max_length=max_length,)

tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = 'right'

#EXAMPLE #1
txt="""<|im_start|>user
I'm looking for an efficient Python script to output prime numbers. Can you help me out? I'm interested in a script that can handle large numbers and output them quickly. Also, it would be great if the script could take a range of numbers as input and output all the prime numbers within that range. Can you generate a script that fits these requirements? Thanks!<|im_end|>
<|im_start|>assistant
"""

#EXAMPLE #2
txt="""<|im_start|>user
Estoy desarrollando una REST API con Nodejs, y estoy tratando de aplicar algún sistema de seguridad, ya sea con tokens o algo similar, me puedes ayudar?<|im_end|>
<|im_start|>assistant
"""

inputs = tokenizer.encode(txt, return_tensors="pt").to("cuda")

generation_config = GenerationConfig(
              max_new_tokens=max_new_tokens,
              temperature=0.7,
              top_p=0.9,
              top_k=len_tokens,
              repetition_penalty=1.11, 
              do_sample=True,
              #  pad_token_id=tokenizer.eos_token_id,
              #  eos_token_id=tokenizer.eos_token_id,
              #  use_cache=True,
              # stopping_criteria= StoppingCriteriaList([stopping_criteria]),
          )
outputs = model.generate(generation_config=generation_config,
                                input_ids=inputs,)
tokenizer.decode(outputs[0], skip_special_tokens=False) #True
```