rishiraj commited on
Commit
973a85c
1 Parent(s): 6f629ba

Create app.py

Browse files
Files changed (1) hide show
  1. app.py +100 -0
app.py ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from huggingface_hub import InferenceClient
2
+ import gradio as gr
3
+
4
+ client = InferenceClient(
5
+ "HuggingFaceH4/zephyr-7b-alpha"
6
+ )
7
+
8
+
9
+ def format_prompt(message, history):
10
+ prompt = "<s>"
11
+ for user_prompt, bot_response in history:
12
+ prompt += f"[INST] {user_prompt} [/INST]"
13
+ prompt += f" {bot_response}</s> "
14
+ prompt += f"[INST] {message} [/INST]"
15
+ return prompt
16
+
17
+ def generate(
18
+ prompt, history, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0,
19
+ ):
20
+ temperature = float(temperature)
21
+ if temperature < 1e-2:
22
+ temperature = 1e-2
23
+ top_p = float(top_p)
24
+
25
+ generate_kwargs = dict(
26
+ temperature=temperature,
27
+ max_new_tokens=max_new_tokens,
28
+ top_p=top_p,
29
+ repetition_penalty=repetition_penalty,
30
+ do_sample=True,
31
+ seed=42,
32
+ )
33
+
34
+ formatted_prompt = format_prompt(prompt, history)
35
+
36
+ stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
37
+ output = ""
38
+
39
+ for response in stream:
40
+ output += response.token.text
41
+ yield output
42
+ return output
43
+
44
+
45
+ additional_inputs=[
46
+ gr.Slider(
47
+ label="Temperature",
48
+ value=0.9,
49
+ minimum=0.0,
50
+ maximum=1.0,
51
+ step=0.05,
52
+ interactive=True,
53
+ info="Higher values produce more diverse outputs",
54
+ ),
55
+ gr.Slider(
56
+ label="Max new tokens",
57
+ value=256,
58
+ minimum=0,
59
+ maximum=1048,
60
+ step=64,
61
+ interactive=True,
62
+ info="The maximum numbers of new tokens",
63
+ ),
64
+ gr.Slider(
65
+ label="Top-p (nucleus sampling)",
66
+ value=0.90,
67
+ minimum=0.0,
68
+ maximum=1,
69
+ step=0.05,
70
+ interactive=True,
71
+ info="Higher values sample more low-probability tokens",
72
+ ),
73
+ gr.Slider(
74
+ label="Repetition penalty",
75
+ value=1.2,
76
+ minimum=1.0,
77
+ maximum=2.0,
78
+ step=0.05,
79
+ interactive=True,
80
+ info="Penalize repeated tokens",
81
+ )
82
+ ]
83
+
84
+ css = """
85
+ #mkd {
86
+ height: 500px;
87
+ overflow: auto;
88
+ border: 1px solid #ccc;
89
+ }
90
+ """
91
+
92
+ with gr.Blocks(css=css) as demo:
93
+ gr.HTML("<h1><center><a href='https://rishiraj.github.io/'>Rishiraj</a>'s AI Assistant<h1><center>")
94
+ gr.ChatInterface(
95
+ generate,
96
+ additional_inputs=additional_inputs,
97
+ examples=[["What is the secret to life?"], ["Write me a recipe for pancakes."]]
98
+ )
99
+
100
+ demo.queue(concurrency_count=75, max_size=100).launch(debug=True)