File size: 20,545 Bytes
a85f909 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 |
import dataclasses
import pprint
from functools import partial
import re
import os
from threading import Lock
import urllib
import time
from typing import List, Optional, Union
from pydantic import BaseModel
import absl.logging
from tqdm import tqdm, trange
import numpy as np
import mlxu
from ml_collections import ConfigDict
import uvicorn
from fastapi import FastAPI
import gradio as gr
import requests
from requests.exceptions import Timeout, ConnectionError
class InferenceRequest(BaseModel):
prefix_text: Optional[List[str]] = None
text: Optional[List[str]] = None
until: Optional[Union[List[str], List[List[str]]]] = None
temperature: Optional[float] = None
class ChatRequest(BaseModel):
prompt: str
context: str = ''
temperature: Optional[float] = None
class LMServer(object):
""" HTTP server for serving langauge models. """
@staticmethod
def get_default_config(updates=None):
config = ConfigDict()
config.host = '0.0.0.0'
config.port = 5007
config.batch_size = 1
config.logging = False
config.pre_compile = 'loglikelihood'
config.default_temperature = 1.0
config.greedy_until_max_length = 5000
config.prepend_to_prefix = ''
config.append_to_prefix = ''
config.prepend_to_text = ''
config.append_to_text = ''
config.chat_prepend_text = ''
config.chat_user_prefix = ''
config.chat_user_suffix = ''
config.chat_lm_prefix = ''
config.chat_lm_suffix = ''
config.notes = ''
if updates is not None:
config.update(ConfigDict(updates).copy_and_resolve_references())
return config
def __init__(self, config):
self.config = self.get_default_config(config)
self.lock = Lock()
self.app = FastAPI()
self.app.post('/loglikelihood')(self.serve_loglikelihood)
self.app.post('/loglikelihood-rolling')(self.serve_loglikelihood_rolling)
self.app.post('/generate')(self.serve_generate)
self.app.post('/greedy-until')(self.serve_greedy_until)
self.app.post('/chat')(self.serve_chat)
self.app.get('/ready')(self.serve_ready)
self.app = gr.mount_gradio_app(self.app, self.create_chat_app(), '/')
@staticmethod
def loglikelihood(prefix_text, text):
raise NotImplementedError()
@staticmethod
def loglikelihood_rolling(text):
raise NotImplementedError()
@staticmethod
def generate(text, temperature):
raise NotImplementedError()
@staticmethod
def greedy_until(prefix_text, until, max_length):
raise NotImplementedError()
@staticmethod
def to_list(x):
if isinstance(x, np.ndarray):
return x.tolist()
return x
def serve_ready(self):
return 'Ready!\n'
def serve_loglikelihood(self, data: InferenceRequest):
with self.lock:
if self.config.logging:
absl.logging.info(
'\n========= Serving Log Likelihood Request ========= \n'
+ pprint.pformat(data) + '\n'
)
if data.prefix_text is None:
data.prefix_text = ['' for _ in data.text]
prefix_text = [
self.config.prepend_to_prefix + p + self.config.append_to_prefix
for p in data.prefix_text
]
text = [
self.config.prepend_to_text + t + self.config.append_to_text
for t in data.text
]
log_likelihood = []
is_greedy = []
for i in trange(0, len(text), self.config.batch_size, ncols=0):
batch_prefix_text = prefix_text[i:i + self.config.batch_size]
batch_text = text[i:i + self.config.batch_size]
batch_size = len(batch_text)
if batch_size < self.config.batch_size:
extra = self.config.batch_size - batch_size
batch_prefix_text.extend(['a' for _ in range(extra)])
batch_text.extend(['a' for _ in range(extra)])
batch_log_likelihood, batch_is_greedy = self.loglikelihood(
batch_prefix_text, batch_text
)
batch_log_likelihood = self.to_list(batch_log_likelihood)
batch_is_greedy = self.to_list(batch_is_greedy)
log_likelihood.extend(batch_log_likelihood[:batch_size])
is_greedy.extend(batch_is_greedy[:batch_size])
output = {
'prefix_text': data.prefix_text,
'text': data.text,
'log_likelihood': log_likelihood,
'is_greedy': is_greedy,
}
if self.config.logging:
absl.logging.info(
'\n========= Output ========= \n'
+ pprint.pformat(output) + '\n'
)
return output
def serve_loglikelihood_rolling(self, data: InferenceRequest):
with self.lock:
if self.config.logging:
absl.logging.info(
'\n========= Serving Log Likelihood Request ========= \n'
+ pprint.pformat(data) + '\n'
)
text = [
self.config.prepend_to_text + t + self.config.append_to_text
for t in data.text
]
log_likelihood = []
is_greedy = []
for i in trange(0, len(text), self.config.batch_size, ncols=0):
batch_text = text[i:i + self.config.batch_size]
batch_size = len(batch_text)
if batch_size < self.config.batch_size:
extra = self.config.batch_size - batch_size
batch_text.extend(['a' for _ in range(extra)])
batch_log_likelihood, batch_is_greedy = self.loglikelihood_rolling(
batch_text
)
batch_log_likelihood = self.to_list(batch_log_likelihood)
batch_is_greedy = self.to_list(batch_is_greedy)
log_likelihood.extend(batch_log_likelihood[:batch_size])
is_greedy.extend(batch_is_greedy[:batch_size])
output = {
'text': data.text,
'log_likelihood': log_likelihood,
'is_greedy': is_greedy,
}
if self.config.logging:
absl.logging.info(
'\n========= Output ========= \n'
+ pprint.pformat(output) + '\n'
)
return output
def serve_generate(self, data: InferenceRequest):
with self.lock:
if self.config.logging:
absl.logging.info(
'\n========= Serving Generate Request ========= \n'
+ pprint.pformat(data) + '\n'
)
prefix_text = [
self.config.prepend_to_prefix + p + self.config.append_to_prefix
for p in data.prefix_text
]
if data.temperature is None:
data.temperature = self.config.default_temperature
output_text = []
for i in trange(0, len(prefix_text), self.config.batch_size, ncols=0):
batch_prefix_text = prefix_text[i:i + self.config.batch_size]
batch_size = len(batch_prefix_text)
if batch_size < self.config.batch_size:
extra = self.config.batch_size - batch_size
batch_prefix_text.extend(['a' for _ in range(extra)])
batch_output_text = self.generate(
batch_prefix_text,
temperature=data.temperature,
)
output_text.extend(self.to_list(batch_output_text)[:batch_size])
output = {
'prefix_text': data.prefix_text,
'output_text': output_text,
'temperature': data.temperature,
}
if self.config.logging:
absl.logging.info(
'\n========= Output ========= \n'
+ pprint.pformat(output) + '\n'
)
return output
def serve_greedy_until(self, data: InferenceRequest):
with self.lock:
if self.config.logging:
absl.logging.info(
'\n========= Serving Greedy Until Request ========= \n'
+ pprint.pformat(data) + '\n'
)
prefix_text = [
self.config.prepend_to_prefix + p + self.config.append_to_prefix
for p in data.prefix_text
]
until = data.until
max_length = self.config.greedy_until_max_length
output_text = []
for i in range(0, len(prefix_text), self.config.batch_size):
batch_prefix_text = prefix_text[i:i + self.config.batch_size]
batch_until = until[i:i + self.config.batch_size]
batch_size = len(batch_prefix_text)
batch_output_text = self.greedy_until(batch_prefix_text, batch_until, max_length)
output_text.extend(self.to_list(batch_output_text)[:batch_size])
output = {
'prefix_text': data.prefix_text,
'until': data.until,
'max_length': max_length,
'output_text': output_text,
}
if self.config.logging:
absl.logging.info(
'\n========= Output ========= \n'
+ pprint.pformat(output) + '\n'
)
return output
def process_chat(self, prompt, context, temperature):
context = (
context + self.config.chat_user_prefix
+ prompt + self.config.chat_user_suffix
+ self.config.chat_lm_prefix
)
response = self.generate(
[self.config.chat_prepend_text + context],
temperature=float(temperature),
)[0]
context = context + response + self.config.chat_lm_suffix
return response, context
def serve_chat(self, data: ChatRequest):
if data.temperature is None:
data.temperature = self.config.default_temperature
response, context = self.process_chat(
data.prompt, data.context,
temperature=data.temperature,
)
return {
'response': response,
'context': context,
'temperature': data.temperature,
}
def create_chat_app(self):
with gr.Blocks(analytics_enabled=False, title='EasyLM Chat') as gradio_chatbot:
gr.Markdown('# Chatbot Powered by [EasyLM](https://github.com/young-geng/EasyLM)')
gr.Markdown(self.config.notes)
chatbot = gr.Chatbot(label='Chat history')
msg = gr.Textbox(
placeholder='Type your message here...',
show_label=False
)
with gr.Row():
send = gr.Button('Send')
regenerate = gr.Button('Regenerate', interactive=False)
clear = gr.Button('Reset')
temp_slider = gr.Slider(
label='Temperature', minimum=0, maximum=2.0,
value=self.config.default_temperature
)
context_state = gr.State(['', ''])
def user_fn(user_message, history, context):
return {
msg: gr.update(value='', interactive=False),
clear: gr.update(interactive=False),
send: gr.update(interactive=False),
regenerate: gr.update(interactive=False),
chatbot: history + [[user_message, None]],
context_state: [context[1], context[1]],
}
def model_fn(history, context, temperature):
history[-1][1], new_context = self.process_chat(
history[-1][0], context[0], temperature
)
return {
msg: gr.update(value='', interactive=True),
clear: gr.update(interactive=True),
send: gr.update(interactive=True),
chatbot: history,
context_state: [context[0], new_context],
regenerate: gr.update(interactive=True),
}
def regenerate_fn():
return {
msg: gr.update(value='', interactive=False),
clear: gr.update(interactive=False),
send: gr.update(interactive=False),
regenerate: gr.update(interactive=False),
}
def clear_fn():
return {
chatbot: None,
msg: '',
context_state: ['', ''],
regenerate: gr.update(interactive=False),
}
msg.submit(
user_fn,
inputs=[msg, chatbot, context_state],
outputs=[msg, clear, send, chatbot, context_state, regenerate],
queue=False
).then(
model_fn,
inputs=[chatbot, context_state, temp_slider],
outputs=[msg, clear, send, chatbot, context_state, regenerate],
queue=True
)
send.click(
user_fn,
inputs=[msg, chatbot, context_state],
outputs=[msg, clear, send, chatbot, context_state, regenerate],
queue=False
).then(
model_fn,
inputs=[chatbot, context_state, temp_slider],
outputs=[msg, clear, send, chatbot, context_state, regenerate],
queue=True
)
regenerate.click(
regenerate_fn,
inputs=None,
outputs=[msg, clear, send, regenerate],
queue=False
).then(
model_fn,
inputs=[chatbot, context_state, temp_slider],
outputs=[msg, clear, send, chatbot, context_state, regenerate],
queue=True
)
clear.click(
clear_fn,
inputs=None,
outputs=[chatbot, msg, context_state, regenerate],
queue=False
)
gradio_chatbot.queue(concurrency_count=1)
return gradio_chatbot
def run(self):
if self.config.pre_compile != '':
if self.config.pre_compile == 'all':
pre_compile = ['loglikelihood', 'generate', 'greedy_until', 'chat']
else:
pre_compile = self.config.pre_compile.split(',')
pre_compile_data = ['a' for _ in range(self.config.batch_size)]
for task in pre_compile:
if task == 'loglikelihood':
self.loglikelihood(pre_compile_data, pre_compile_data)
self.loglikelihood_rolling(pre_compile_data)
elif task == 'generate':
self.generate(pre_compile_data, 1.0)
elif task == 'greedy_until':
self.greedy_until(
pre_compile_data, pre_compile_data,
self.config.greedy_until_max_length
)
elif task == 'chat':
self.process_chat('a', 'a', 1.0)
else:
raise ValueError(f'Invalid precompile task: {task}!')
uvicorn.run(self.app, host=self.config.host, port=self.config.port)
class LMClient(object):
""" A simple client for the LM server. """
@staticmethod
def get_default_config(updates=None):
config = ConfigDict()
config.url = 'http://localhost:5007'
config.batch_size = 1
config.wait_for_ready = True
config.dummy = False
if updates is not None:
config.update(ConfigDict(updates).copy_and_resolve_references())
return config
def __init__(self, config=None):
self.config = self.get_default_config(config)
if self.config.wait_for_ready:
self.wait_for_ready()
def wait_for_ready(self):
if self.config.dummy:
return
while True:
try:
requests.get(urllib.parse.urljoin(self.config.url, 'ready'))
return
except (Timeout, ConnectionError) as e:
time.sleep(10)
@staticmethod
def batched(iterator, batch_size):
batch = []
for example in iterator:
batch.append(example)
if len(batch) == batch_size:
yield batch
batch = []
if len(batch) > 0:
yield batch
def loglikelihood(self, prefix, text):
prefix, text = list(prefix), list(text)
if self.config.dummy:
return [-1.0 for _ in text], [False for _ in text]
log_likelihood = []
is_greedy = []
batched_iterator = list(zip(
self.batched(prefix, self.config.batch_size),
self.batched(text, self.config.batch_size)
))
for batch_prefix, batch_text in tqdm(batched_iterator, ncols=0):
response = requests.post(
urllib.parse.urljoin(self.config.url, 'loglikelihood'),
json={'prefix_text': batch_prefix, 'text': batch_text}
).json()
log_likelihood.extend(response['log_likelihood'])
is_greedy.extend(response['is_greedy'])
return log_likelihood, is_greedy
def loglikelihood_rolling(self, text):
text = list(text)
if self.config.dummy:
return [-1.0 for _ in text], [False for _ in text]
log_likelihood = []
is_greedy = []
batched_iterator = list(self.batched(text, self.config.batch_size))
for batch_text in tqdm(batched_iterator, ncols=0):
response = requests.post(
urllib.parse.urljoin(self.config.url, 'loglikelihood-rolling'),
json={'text': batch_text}
).json()
log_likelihood.extend(response['log_likelihood'])
is_greedy.extend(response['is_greedy'])
return log_likelihood, is_greedy
def greedy_until(self, prefix, until):
prefix, until = list(prefix), list(until)
if self.config.dummy:
results = []
for u in until:
if isinstance(u, str):
results.append('dummy text ' + u)
else:
results.append('dummy text ' + u[0])
return results
batched_iterator = list(zip(
self.batched(prefix, self.config.batch_size),
self.batched(until, self.config.batch_size),
))
output_text = []
for batch_prefix, batch_until in tqdm(batched_iterator, ncols=0):
response = requests.post(
urllib.parse.urljoin(self.config.url, 'greedy-until'),
json={'prefix_text': batch_prefix, 'until': batch_until}
).json()
output_text.extend(response['output_text'])
return output_text
def generate(self, prefix, temperature=None):
prefix = list(prefix)
if self.config.dummy:
return ['' for _ in prefix]
output_text = []
batched_iterator = list(self.batched(prefix, self.config.batch_size))
for batch_prefix in tqdm(batched_iterator, ncols=0):
response = requests.post(
urllib.parse.urljoin(self.config.url, 'generate'),
json={
'prefix_text': batch_prefix,
'temperature': temperature,
}
).json()
output_text.extend(response['output_text'])
return output_text
def chat(self, prompt, context, temperature=None):
if self.config.dummy:
return ''
response = requests.post(
urllib.parse.urljoin(self.config.url, 'chat'),
json={
'prompt': prompt,
'context': context,
'temperature': temperature,
}
).json()
return response['response'], response['context']
|