File size: 9,623 Bytes
ef366f8 e1419c7 ef366f8 7de6f28 2ca7bf0 ef366f8 e1419c7 cb5d912 7de6f28 2ca7bf0 e1419c7 ef366f8 e1419c7 cb5d912 e1419c7 ef366f8 e1419c7 ef366f8 e1419c7 ef366f8 e1419c7 ef366f8 e1419c7 ef366f8 cb5d912 2f20661 cb5d912 ef366f8 94809e9 ef366f8 94809e9 ef366f8 040958c a857304 e1419c7 cb5d912 ef366f8 a857304 ef366f8 511ba7f ef366f8 e1419c7 ef366f8 e1419c7 cb5d912 e1419c7 cb5d912 35539cb e952e62 cb5d912 e1419c7 cb5d912 35539cb e1419c7 cb5d912 e1419c7 ef366f8 e1419c7 ef366f8 e1419c7 ef366f8 7de6f28 ef366f8 7de6f28 1de9b09 7de6f28 ef366f8 e86f2ba ef366f8 e1419c7 ef366f8 cb5d912 ef366f8 cb5d912 ef366f8 7de6f28 ef366f8 7de6f28 ef366f8 7de6f28 ef366f8 7de6f28 ef366f8 7de6f28 ef366f8 7de6f28 ef366f8 e952e62 ef366f8 7de6f28 ef366f8 7de6f28 e1419c7 cb5d912 e1419c7 cb5d912 e1419c7 ef366f8 cb5d912 ef366f8 cb5d912 ef366f8 e4cda74 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 |
import os
import gradio as gr
from text_generation import Client, InferenceAPIClient
openchat_preprompt = (
"\n<human>: Hi!\n<bot>: My name is Bot, model version is 0.15, part of an open-source kit for "
"fine-tuning new bots! I was created by Together, LAION, and Ontocord.ai and the open-source "
"community. I am not human, not evil and not alive, and thus have no thoughts and feelings, "
"but I am programmed to be helpful, polite, honest, and friendly.\n"
)
def get_client(model: str):
if model == "Rallio67/joi2_20Be_instruct_alpha":
return Client(os.getenv("JOI_API_URL"))
if model == "togethercomputer/GPT-NeoXT-Chat-Base-20B":
return Client(os.getenv("OPENCHAT_API_URL"))
return InferenceAPIClient(model, token=os.getenv("HF_TOKEN", None))
def get_usernames(model: str):
"""
Returns:
(str, str, str, str): pre-prompt, username, bot name, separator
"""
if model == "OpenAssistant/oasst-sft-1-pythia-12b":
return "", "<|prompter|", "<|assistant|>", "<|endoftext|>"
if model == "Rallio67/joi2_20Be_instruct_alpha":
return "", "User: ", "Joi: ", "\n\n"
if model == "togethercomputer/GPT-NeoXT-Chat-Base-20B":
return openchat_preprompt, "<human>: ", "<bot>: ", "\n"
return "", "User: ", "Assistant: ", "\n"
def predict(
model: str,
inputs: str,
typical_p: float,
top_p: float,
temperature: float,
top_k: int,
repetition_penalty: float,
watermark: bool,
chatbot,
history,
):
client = get_client(model)
preprompt, user_name, assistant_name, sep = get_usernames(model)
history.append(inputs)
past = []
for data in chatbot:
user_data, model_data = data
if not user_data.startswith(user_name):
user_data = user_name + user_data
if not model_data.startswith(sep + assistant_name):
model_data = sep + assistant_name + model_data
past.append(user_data + model_data.rstrip() + sep)
if not inputs.startswith(user_name):
inputs = user_name + inputs
total_inputs = preprompt + "".join(past) + inputs + sep + assistant_name.rstrip()
partial_words = ""
if model == "OpenAssistant/oasst-sft-1-pythia-12b":
iterator = client.generate_stream(
total_inputs,
typical_p=typical_p,
watermark=watermark,
max_new_tokens=500,
)
else:
iterator = client.generate_stream(
total_inputs,
top_p=top_p if top_p < 1.0 else None,
top_k=top_k,
truncate=1000,
repetition_penalty=repetition_penalty,
watermark=watermark,
temperature=temperature,
max_new_tokens=500,
stop_sequences=[user_name.rstrip(), assistant_name.rstrip()],
)
for i, response in enumerate(iterator):
if response.token.special:
continue
partial_words = partial_words + response.token.text
if partial_words.endswith(user_name.rstrip()):
partial_words = partial_words.rstrip(user_name.rstrip())
if partial_words.endswith(assistant_name.rstrip()):
partial_words = partial_words.rstrip(assistant_name.rstrip())
if i == 0:
history.append(" " + partial_words)
elif response.token.text not in user_name:
history[-1] = partial_words
chat = [
(history[i].strip(), history[i + 1].strip())
for i in range(0, len(history) - 1, 2)
]
yield chat, history
def reset_textbox():
return gr.update(value="")
def radio_on_change(
value: str,
disclaimer,
typical_p,
top_p,
top_k,
temperature,
repetition_penalty,
watermark,
):
if value == "OpenAssistant/oasst-sft-1-pythia-12b":
typical_p = typical_p.update(value=0.2, visible=True)
top_p = top_p.update(visible=False)
top_k = top_k.update(visible=False)
temperature = temperature.update(visible=False)
disclaimer = disclaimer.update(visible=False)
repetition_penalty = repetition_penalty.update(visible=False)
watermark = watermark.update(False)
elif value == "togethercomputer/GPT-NeoXT-Chat-Base-20B":
typical_p = typical_p.update(visible=False)
top_p = top_p.update(value=0.25, visible=True)
top_k = top_k.update(value=50, visible=True)
temperature = temperature.update(value=0.6, visible=True)
repetition_penalty = repetition_penalty.update(value=1.01, visible=True)
watermark = watermark.update(False)
disclaimer = disclaimer.update(visible=True)
else:
typical_p = typical_p.update(visible=False)
top_p = top_p.update(value=0.95, visible=True)
top_k = top_k.update(value=4, visible=True)
temperature = temperature.update(value=0.5, visible=True)
repetition_penalty = repetition_penalty.update(value=1.03, visible=True)
watermark = watermark.update(True)
disclaimer = disclaimer.update(visible=False)
return (
disclaimer,
typical_p,
top_p,
top_k,
temperature,
repetition_penalty,
watermark,
)
title = """<h1 align="center">🔥Large Language Model API 🚀Streaming🚀</h1>"""
description = """Language models can be conditioned to act like dialogue agents through a conversational prompt that typically takes the form:
```
User: <utterance>
Assistant: <utterance>
User: <utterance>
Assistant: <utterance>
...
```
In this app, you can explore the outputs of multiple LLMs when prompted in this way.
"""
openchat_disclaimer = """
<div align="center">Checkout the official <a href=https://huggingface.co/spaces/togethercomputer/OpenChatKit>OpenChatKit feedback app</a> for the full experience.</div>
"""
with gr.Blocks(
css="""#col_container {margin-left: auto; margin-right: auto;}
#chatbot {height: 520px; overflow: auto;}"""
) as demo:
gr.HTML(title)
with gr.Column(elem_id="col_container"):
model = gr.Radio(
value="OpenAssistant/oasst-sft-1-pythia-12b",
choices=[
"OpenAssistant/oasst-sft-1-pythia-12b",
# "togethercomputer/GPT-NeoXT-Chat-Base-20B",
"Rallio67/joi2_20Be_instruct_alpha",
"google/flan-t5-xxl",
"google/flan-ul2",
"bigscience/bloom",
"bigscience/bloomz",
"EleutherAI/gpt-neox-20b",
],
label="Model",
interactive=True,
)
chatbot = gr.Chatbot(elem_id="chatbot")
inputs = gr.Textbox(
placeholder="Hi there!", label="Type an input and press Enter"
)
disclaimer = gr.Markdown(openchat_disclaimer, visible=False)
state = gr.State([])
b1 = gr.Button()
with gr.Accordion("Parameters", open=False):
typical_p = gr.Slider(
minimum=-0,
maximum=1.0,
value=0.2,
step=0.05,
interactive=True,
label="Typical P mass",
)
top_p = gr.Slider(
minimum=-0,
maximum=1.0,
value=0.25,
step=0.05,
interactive=True,
label="Top-p (nucleus sampling)",
visible=False,
)
temperature = gr.Slider(
minimum=-0,
maximum=5.0,
value=0.6,
step=0.1,
interactive=True,
label="Temperature",
visible=False,
)
top_k = gr.Slider(
minimum=1,
maximum=50,
value=50,
step=1,
interactive=True,
label="Top-k",
visible=False,
)
repetition_penalty = gr.Slider(
minimum=0.1,
maximum=3.0,
value=1.03,
step=0.01,
interactive=True,
label="Repetition Penalty",
visible=False,
)
watermark = gr.Checkbox(value=False, label="Text watermarking")
model.change(
lambda value: radio_on_change(
value,
disclaimer,
typical_p,
top_p,
top_k,
temperature,
repetition_penalty,
watermark,
),
inputs=model,
outputs=[
disclaimer,
typical_p,
top_p,
top_k,
temperature,
repetition_penalty,
watermark,
],
)
inputs.submit(
predict,
[
model,
inputs,
typical_p,
top_p,
temperature,
top_k,
repetition_penalty,
watermark,
chatbot,
state,
],
[chatbot, state],
)
b1.click(
predict,
[
model,
inputs,
typical_p,
top_p,
temperature,
top_k,
repetition_penalty,
watermark,
chatbot,
state,
],
[chatbot, state],
)
b1.click(reset_textbox, [], [inputs])
inputs.submit(reset_textbox, [], [inputs])
gr.Markdown(description)
demo.queue(concurrency_count=16).launch(debug=True)
|