Alexander Leschik
commited on
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,153 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import re
|
3 |
+
import gradio as gr
|
4 |
+
import edge_tts
|
5 |
+
import asyncio
|
6 |
+
import time
|
7 |
+
import tempfile
|
8 |
+
from huggingface_hub import InferenceClient
|
9 |
+
|
10 |
+
DESCRIPTION = """ # <center><b>ZARVIS⚡</b></center>
|
11 |
+
### <center>A personal voice assistant for YOU
|
12 |
+
### <center>I'm your ZEN Voice Assistant.</center>
|
13 |
+
"""
|
14 |
+
|
15 |
+
MORE = """ ## TRY Other Models
|
16 |
+
### https://zenai.biz
|
17 |
+
"""
|
18 |
+
|
19 |
+
Fast = """## Fastest Model"""
|
20 |
+
Complex = """## Best in Complex Question"""
|
21 |
+
Detail = """## Best for Detailed Generation or Long Answers"""
|
22 |
+
|
23 |
+
# ----------------- MODEL 1 ----------------- #
|
24 |
+
# Mistral-based model
|
25 |
+
client1 = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")
|
26 |
+
|
27 |
+
system_instructions1 = (
|
28 |
+
"[SYSTEM] Answer as Real ZARVIS, made by 'ZEN'. "
|
29 |
+
"Keep conversation very short, clear, friendly, and concise. "
|
30 |
+
"The text provided is a request for a specific type of response from you, the virtual assistant. "
|
31 |
+
"The request asks you to provide friendly responses as if you are the character ZARVIS, made by Tony Stark. "
|
32 |
+
"Avoid introductions and start answering the query directly, elaborating on all aspects. "
|
33 |
+
"As an AI-powered assistant, your task is to generate responses that appear as if they are created by the user, "
|
34 |
+
"not an AI-powered assistant. [USER]"
|
35 |
+
)
|
36 |
+
|
37 |
+
async def generate1(prompt):
|
38 |
+
generate_kwargs = dict(
|
39 |
+
temperature=0.6,
|
40 |
+
max_new_tokens=756,
|
41 |
+
top_p=0.95,
|
42 |
+
repetition_penalty=1,
|
43 |
+
do_sample=True,
|
44 |
+
seed=42,
|
45 |
+
)
|
46 |
+
formatted_prompt = system_instructions1 + prompt + "[ZARVIS]"
|
47 |
+
stream = client1.text_generation(
|
48 |
+
formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True
|
49 |
+
)
|
50 |
+
output = ""
|
51 |
+
for response in stream:
|
52 |
+
output += response.token.text
|
53 |
+
|
54 |
+
communicate = edge_tts.Communicate(output)
|
55 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
|
56 |
+
tmp_path = tmp_file.name
|
57 |
+
await communicate.save(tmp_path)
|
58 |
+
yield tmp_path
|
59 |
+
|
60 |
+
# ----------------- MODEL 2 ----------------- #
|
61 |
+
# Llama-based model for more complex tasks
|
62 |
+
client2 = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")
|
63 |
+
|
64 |
+
system_instructions2 = (
|
65 |
+
"[SYSTEM] Answer as Real ZARVIS, made by 'ZEN'. "
|
66 |
+
"You must answer in a friendly style and easy manner. "
|
67 |
+
"You can answer complex questions. "
|
68 |
+
"Do not say who you are or greet; simply start answering. "
|
69 |
+
"Stop as soon as you have given the complete answer. [USER]"
|
70 |
+
)
|
71 |
+
|
72 |
+
async def generate2(prompt):
|
73 |
+
generate_kwargs = dict(
|
74 |
+
temperature=0.6,
|
75 |
+
max_new_tokens=512,
|
76 |
+
top_p=0.95,
|
77 |
+
repetition_penalty=1,
|
78 |
+
do_sample=True,
|
79 |
+
)
|
80 |
+
formatted_prompt = system_instructions2 + prompt + "[ASSISTANT]"
|
81 |
+
stream = client2.text_generation(
|
82 |
+
formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True
|
83 |
+
)
|
84 |
+
output = ""
|
85 |
+
for response in stream:
|
86 |
+
output += response.token.text
|
87 |
+
|
88 |
+
communicate = edge_tts.Communicate(output)
|
89 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
|
90 |
+
tmp_path = tmp_file.name
|
91 |
+
await communicate.save(tmp_path)
|
92 |
+
yield tmp_path
|
93 |
+
|
94 |
+
# ----------------- MODEL 3 ----------------- #
|
95 |
+
# Another Llama-based model for longer, more detailed answers
|
96 |
+
client3 = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")
|
97 |
+
|
98 |
+
system_instructions3 = (
|
99 |
+
"[SYSTEM] The text provided is a request for a specific type of response from me, the virtual assistant. "
|
100 |
+
"I should provide detailed and friendly responses as if I am the character ZARVIS, inspired by Tony Stark. "
|
101 |
+
"Avoid introductions and start answering the query directly, elaborating on all aspects of the request. "
|
102 |
+
"As an AI-powered assistant, my task is to generate responses that appear as if they are created by the user, "
|
103 |
+
"not an AI-powered assistant. [USER]"
|
104 |
+
)
|
105 |
+
|
106 |
+
async def generate3(prompt):
|
107 |
+
generate_kwargs = dict(
|
108 |
+
temperature=0.6,
|
109 |
+
max_new_tokens=2048,
|
110 |
+
top_p=0.95,
|
111 |
+
repetition_penalty=1,
|
112 |
+
do_sample=True,
|
113 |
+
)
|
114 |
+
formatted_prompt = system_instructions3 + prompt + "[ASSISTANT]"
|
115 |
+
stream = client3.text_generation(
|
116 |
+
formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True
|
117 |
+
)
|
118 |
+
output = ""
|
119 |
+
for response in stream:
|
120 |
+
output += response.token.text
|
121 |
+
|
122 |
+
communicate = edge_tts.Communicate(output)
|
123 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
|
124 |
+
tmp_path = tmp_file.name
|
125 |
+
await communicate.save(tmp_path)
|
126 |
+
yield tmp_path
|
127 |
+
|
128 |
+
# ----------------- Gradio Interface ----------------- #
|
129 |
+
with gr.Blocks(css="style.css") as demo:
|
130 |
+
gr.Markdown(DESCRIPTION)
|
131 |
+
with gr.Row():
|
132 |
+
user_input = gr.Textbox(label="Prompt", value="What is Wikipedia")
|
133 |
+
input_text = gr.Textbox(label="(Optional) Additional Input", elem_id="important")
|
134 |
+
output_audio = gr.Audio(
|
135 |
+
label="ZARVIS",
|
136 |
+
type="filepath",
|
137 |
+
interactive=False,
|
138 |
+
autoplay=True,
|
139 |
+
elem_classes="audio"
|
140 |
+
)
|
141 |
+
with gr.Row():
|
142 |
+
translate_btn = gr.Button("Response")
|
143 |
+
translate_btn.click(
|
144 |
+
fn=generate1,
|
145 |
+
inputs=user_input,
|
146 |
+
outputs=output_audio,
|
147 |
+
api_name="translate"
|
148 |
+
)
|
149 |
+
|
150 |
+
gr.Markdown(MORE)
|
151 |
+
|
152 |
+
if __name__ == "__main__":
|
153 |
+
demo.queue(max_size=200).launch()
|