Guide
CFGGuide
Bases: Guide
Guide to generate text that is in the language of a context-free grammar.
Source code in outlines/fsm/guide.py
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 |
|
copy()
get_next_instruction(state)
Generate an instruction for the next step.
Upon initialization, the CFG incremental parser is used to determine the first regex and construct the first FSM to generate the first terminal.
This FSM is used for proposals until either:
- The FSM is exhausted, and its only remaining option is the EOS token, in which case we feed the generated terminal to the CFG incremental parser and allow it to propose the next regex corresponding to the next set of valid terminals.
- The current FSM can be exhausted, but the EOS token is not the only remaining option. In this case we allow proposal of current terminal extensions, store the current FSM and its state, then also use the CFG parser to propose a new regex corresponding to terminating the current terminal and starting the next one. The model can then sample from either of these sets to determine whether to extend the current terminal or terminate it and start the next one.
The CFG incremental parser is allowed to propose the EOS token from any accepting state, and once it is generated, the FSM will continue to always generate the EOS token.
Parameters
state The current state of the FSM.
Returns
A list that contains the tokens to mask.
Source code in outlines/fsm/guide.py
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 |
|
get_next_state(state, token_id)
Update the state of the guide.
Transitions the underlying regex FSM to its next state. If at max tokens or EOS token, transition permanently to the final state. Update stored partial generations for subsequent incremental parsing.
Parameters
state The current state of the FSM. token_id The id of the token that was just generated.
Returns
The new state of the FSM.
Source code in outlines/fsm/guide.py
Generate
dataclass
Generate instruction
Attributes
tokens
The tokens that lead to a valid completion if generated. A value
of None
indicates that all tokens are allowed.
Source code in outlines/fsm/guide.py
Guide
Bases: Protocol
Base definition of a generation guide.
A generation guide defines the behavior of a finite-state machine that guides
a text generation procedure. Unlike the DFAs built from regular expressions
guides can also emit a Write
instructions which tells the model that it can
append a sequence of tokens (or token word) instead of generating it.
Source code in outlines/fsm/guide.py
RegexGuide
Bases: Guide
Guide to generate text in the language of a regular expression.
Source code in outlines/fsm/guide.py
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 |
|
get_next_instruction(state)
Return the next instruction for guided generation.
The initialization of the guide builds an index which maps FSM states to a map from authorized tokens to the state in which the guide needs to move if said token is generated. Therefore the authorized tokens at the current state are the keys of the map returned by the value of the index for current state.
If the current state is not contained in the end this means that we are in a final state of the guide. We only authorize EOS tokens in the final state.
Parameters
state The current state of the guide.
Returns
A Generate
instance that contains the model and the allowed token ids.
Source code in outlines/fsm/guide.py
get_next_state(state, token_id)
Update the state of the guide.
We use the index to determine to which state the guide should transition given the token that was just generated.
Parameters
state The current state of the guide. token_id The id of the token that was just generated.
Returns
The new state of the guide.
Source code in outlines/fsm/guide.py
StopAtEOSGuide
Bases: Guide
Guide to generate tokens until the EOS token has been generated.
Source code in outlines/fsm/guide.py
__init__(tokenizer)
Initialize the generation guide.
model The logit generator used to generate the next token.
Write
dataclass
Write instruction.
Attributes
tokens The sequence of tokens to be added to the current sequence by the generation process.
Source code in outlines/fsm/guide.py
create_states_mapping(regex_string, tokenizer, regex_parser=interegular.parse_pattern, frozen_tokens=[])
Create the variables related to the mapping between states and tokens The parameters of the function are used for caching purpose.
Parameters
regex_string: (str
):
The regular expression string to generate a states mapping for.
tokenizer: (Tokenizer
):
The model's tokenizer.
regex_parser: (Callable[[str], interegular.Pattern]
, optional):
A function that parses a regex string into an interegular
Pattern object.
frozen_tokens: (List[str]
, optional):
A list of tokens that should be kept as-is when expanding the token-level FSM
into a byte-level FSM. Defaults to an empty list.
Returns
states_to_token_maps: (Dict[int, Dict[int, int]]
):
A mapping from states to a mapping from token ids originating from that state
to the next state to transition to given that token. The structure is as follows:
(origin_state -> (token_id -> next_state))
empty_token_ids: (Set[int]
):
A set of token ids that correspond to empty strings.
final_states: (set
):
A set of final states in the FSM.