File size: 9,651 Bytes
0fe477d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
"""
RWKV RNN Model - Gradio Space for HuggingFace
YT - Mean Gene Hacks - https://www.youtube.com/@MeanGeneHacks
(C) Gene Ruebsamen - 2/7/2023
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program.  If not, see <http://www.gnu.org/licenses/>.
"""

import gradio as gr
import codecs
from ast import literal_eval
from datetime import datetime
from rwkvstic.load import RWKV
from config import config, title
import torch
import gc

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

desc = '''<p>RNN with Transformer-level LLM Performance (<a href='https://github.com/BlinkDL/RWKV-LM'>github</a>).
    According to the author: "It combines the best of RNN and transformers - great performance, fast inference, saves VRAM, fast training, "infinite" ctx_len, and free sentence embedding."'''

thanks = '''<p>Thanks to <a href='https://github.com/gururise/rwkv_gradio'>Gururise</a> for this template</p>'''


def to_md(text):
    return text.replace("\n", "<br />")


def get_model():
    model = None
    model = RWKV(
        **config
    )
    return model


model = get_model()



def infer(
        prompt,
        mode="generative",
        max_new_tokens=10,
        temperature=0.1,
        top_p=1.0,
        stop="<|endoftext|>",
        end_adj=0.0,
        seed=42,
):
    global model

    if model == None:
        gc.collect()
        if (DEVICE == "cuda"):
            torch.cuda.empty_cache()
        model = get_model()

    max_new_tokens = int(max_new_tokens)
    temperature = float(temperature)
    end_adj = float(end_adj)
    top_p = float(top_p)
    stop = [x.strip(' ') for x in stop.split(',')]
    seed = seed

    assert 1 <= max_new_tokens <= 512
    assert 0.0 <= temperature <= 5.0
    assert 0.0 <= top_p <= 1.0

    temperature = max(0.05, temperature)
    if prompt == "":
        prompt = " "

    # Clear model state for generative mode
    model.resetState()
    if (mode == "Q/A"):
        prompt = f"\nQ: {prompt}\n\nA:"
    if (mode == "ELDR"):
        prompt = f"\n{prompt}\n\nExpert Long Detailed Response:\n\nHi, thanks for reaching out, we would be happy to answer your question"
    if (mode == "Expert"):
        prompt = f"\n{prompt}\n\nExpert Full Response:\n\nHi, thanks for reaching out, we would be happy to answer your question.\n"
    if (mode == "EFA"):
        prompt = f'\nAsk Expert\n\nQuestion:\n{prompt}\n\nExpert Full Answer:\n'
    if (mode == "BFR"):
        prompt = f"Task given:\n\n{prompt}\n\nBest Full Response:"

    print(f"PROMPT ({datetime.now()}):\n-------\n{prompt}")
    print(f"OUTPUT ({datetime.now()}):\n-------\n")
    # Load prompt
    model.loadContext(newctx=prompt)
    generated_text = ""
    done = False
    with torch.no_grad():
        for _ in range(max_new_tokens):
            char = model.forward(stopStrings=stop, temp=temperature, top_p_usual=top_p, end_adj=end_adj)[
                "output"]
            print(char, end='', flush=True)
            generated_text += char
            generated_text = generated_text.lstrip("\n ")

            for stop_word in stop:
                stop_word = codecs.getdecoder("unicode_escape")(stop_word)[0]
                if stop_word != '' and stop_word in generated_text:
                    done = True
                    break
            yield generated_text
            if done:
                print("<stopped>\n")
                break

    # print(f"{generated_text}")

    for stop_word in stop:
        stop_word = codecs.getdecoder("unicode_escape")(stop_word)[0]
        if stop_word != '' and stop_word in generated_text:
            generated_text = generated_text[:generated_text.find(stop_word)]

    gc.collect()
    yield generated_text
username = "USER"
intro = f'''The following is a verbose and detailed conversation between an AI assistant called FRITZ, and a human user called USER. FRITZ is intelligent, knowledgeable, wise and polite.
    {username}: What year was the french revolution?
    FRITZ: The French Revolution started in 1789, and lasted 10 years until 1799.
    {username}: 3+5=?
    FRITZ: The answer is 8.
    {username}: What year did the Berlin Wall fall?
    FRITZ: The Berlin wall stood for 28 years and fell in 1989.
    {username}: solve for a: 9-a=2
    FRITZ: The answer is a=7, because 9-7 = 2.
    {username}: wat is lhc
    FRITZ: The Large Hadron Collider (LHC) is a high-energy particle collider, built by CERN, and completed in 2008. It was used to confirm the existence of the Higgs boson in 2012.
    {username}: Tell me about yourself.
    FRITZ: My name is Fritz. I am an RNN based Large Language Model (LLM).
    '''

model.loadContext(newctx=intro)
chatState = model.getState()
model.resetState()
def chat(
        prompt,
        history,
        max_new_tokens=10,
        temperature=0.1,
        top_p=1.0,
        seed=42,
):
    global model
    global username
    history = history or []

    intro = ""

    if model == None:
        gc.collect()
        if (DEVICE == "cuda"):
            torch.cuda.empty_cache()
        model = get_model()

    username = username.strip()
    username = username or "USER"

    

    if len(history) == 0:
        # no history, so lets reset chat state
        model.setState(chatState)
        history = [[], model.emptyState]
        print("reset chat state")
    else:
        if (history[0][0][0].split(':')[0] != username):
            model.setState(chatState)
            history = [[], model.chatState]
            print("username changed, reset state")
        else:
            model.setState(history[1])
            intro = ""

    max_new_tokens = int(max_new_tokens)
    temperature = float(temperature)
    top_p = float(top_p)
    seed = seed

    assert 1 <= max_new_tokens <= 512
    assert 0.0 <= temperature <= 3.0
    assert 0.0 <= top_p <= 1.0

    temperature = max(0.05, temperature)

    prompt = f"{username}: " + prompt + "\n"
    print(f"CHAT ({datetime.now()}):\n-------\n{prompt}")
    print(f"OUTPUT ({datetime.now()}):\n-------\n")
    # Load prompt

    model.loadContext(newctx=prompt)

    out = model.forward(number=max_new_tokens, stopStrings=[
                        "<|endoftext|>", username+":"], temp=temperature, top_p_usual=top_p)

    generated_text = out["output"].lstrip("\n ")
    generated_text = generated_text.rstrip(username+":")
    print(f"{generated_text}")

    gc.collect()
    history[0].append((prompt, generated_text))
    return history[0], [history[0], out["state"]]


examples = [
    [
        # Question Answering
        '''What is the capital of Germany?''', "Q/A", 25, 0.2, 1.0, "<|endoftext|>"],
    [
        # Question Answering
        '''Are humans good or bad?''', "Q/A", 150, 0.8, 0.8, "<|endoftext|>"],
    [
        # Question Answering
        '''What is the purpose of Vitamin A?''', "Q/A", 50, 0.2, 0.8, "<|endoftext|>"],
    [
        # Chatbot
        '''This is a conversation between two AI large language models named Alex and Fritz. They are exploring each other's capabilities, and trying to ask interesting questions of one another to explore the limits of each others AI.
Conversation:
Alex: Good morning, Fritz, what type of LLM are you based upon?
Fritz: Morning Alex, I am an RNN with transformer level performance. My language model is 100% attention free.
Alex:''', "generative", 220, 0.9, 0.9, "\\n\\n,<|endoftext|>"],
    [
        # Generate List
        '''Task given:
Please Write a Short story about a cat learning python
Best Full Response:
''', "generative", 140, 0.85, 0.8, "<|endoftext|>"],
    [
        # Natural Language Interface
        '''Here is a short story (in the style of Tolkien) in which Aiden attacks a robot with a sword:
        ''', "generative", 140, 0.85, 0.8, "<|endoftext|>"]
]


iface = gr.Interface(
    fn=infer,
    description=f'''<h3>Generative and Question/Answer</h3>{desc}{thanks}''',
    allow_flagging="never",
    inputs=[
        gr.Textbox(lines=20, label="Prompt"),  # prompt
        gr.Radio(["Freeform", "Q/A","ELDR","Expert","EFR","BFR"],
                 value="Expert", label="Choose Mode"),
        gr.Slider(1, 512, value=40),  # max_tokens
        gr.Slider(0.0, 5.0, value=0.9),  # temperature
        gr.Slider(0.0, 1.0, value=0.85),  # top_p
        gr.Textbox(lines=1, value="<|endoftext|>"),  # stop
        gr.Slider(-999, 0.0, value=0.0),  # end_adj
       
    ],
    outputs=gr.Textbox(label="Generated Output", lines=25),
    examples=examples,
    cache_examples=False,
).queue()

chatiface = gr.Interface(
    fn=chat,
    description=f'''<h3>Chatbot</h3><h4>Refresh page or change name to reset memory context</h4>{desc}{thanks}''',
    allow_flagging="never",
    inputs=[
        gr.Textbox(lines=5, label="Message"),  # prompt
        "state",
        gr.Slider(1, 256, value=60),  # max_tokens
        gr.Slider(0.0, 1.0, value=0.8),  # temperature
        gr.Slider(0.0, 1.0, value=0.85)  # top_p
    ],
    outputs=[gr.Chatbot(label="Chat Log", color_map=(
        "green", "pink")), "state"],
).queue()

demo = gr.TabbedInterface(

    [iface, chatiface], ["Q/A", "Chatbot"],
    title=title,

)

demo.queue()
demo.launch(share=False)