File size: 21,814 Bytes
5374a2d |
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 |
import copy
import asyncio
from typing import Callable, Dict, Union, Awaitable
from pydantic import Field
import dspy
from ...optimizers.engine.registry import ParamRegistry # Replace with your own path
from typing import List
# import warnings
from ...core.logging import logger
from ...prompts.template import PromptTemplate
from dspy.utils.saving import get_dependency_versions
from pathlib import Path
import cloudpickle
import ujson
class PromptTuningModule(dspy.Module):
"""
A prompt tuning module that manages interactions between predictors,
parameter registry, and program functions.
This module coordinates prompt optimization through:
1. Maintaining a set of predictors for different tasks
2. Synchronizing optimized parameters back to the program
3. Executing the program with updated parameters
Parameters
----------
program : Union[Callable[..., dict], Callable[..., Awaitable[dict]]]
The main program function to execute. Can be either synchronous or asynchronous.
Must return a dictionary containing execution results.
signature_dict : Dict[str, dspy.Signature]
A mapping of task names to their corresponding DSPy signatures.
Each signature defines the input/output structure for a specific task.
registry : ParamRegistry
A registry that maintains tunable parameters shared between
predictors and the program.
"""
@classmethod
def from_registry(
cls,
program: Union[Callable[..., dict], Callable[..., Awaitable[dict]]],
registry: ParamRegistry,
) -> "PromptTuningModule":
"""
Factory method to create a PromptTuningModule from a registry and program.
This method:
1. Creates signatures for each field in the registry
2. Initializes a PromptTuningModule with the program and signatures
3. Sets up predictors for each signature
Parameters
----------
program : Union[Callable[..., dict], Callable[..., Awaitable[dict]]]
The main program function to execute
registry : ParamRegistry
Registry containing tunable parameters
Returns
-------
PromptTuningModule
A configured PromptTuningModule instance
Examples
--------
>>> registry = ParamRegistry()
>>> registry.register("task1", "What is {topic}?")
>>> registry.register("task2", PromptTemplate(system="You are helpful.", user="{query}"))
>>> def my_program(**kwargs) -> dict:
... return {"result": "done"}
>>> module = PromptTuningModule.from_registry(my_program, registry)
"""
from .signature_utils import signature_from_registry
# Create signatures for each field in the registry
signature_dict, signature_name2register_name = signature_from_registry(
registry=registry,
)
# Create and return the module instance
return cls(program=program, signature_dict=signature_dict, registry=registry, signature_name2register_name=signature_name2register_name)
def __init__(
self,
program: Union[Callable[..., dict], Callable[..., Awaitable[dict]]],
signature_dict: Dict[str, dspy.Signature],
registry: ParamRegistry,
signature_name2register_name: Dict[str, str],
):
"""
Initialize a PromptTuningModule instance.
Parameters
----------
program : Union[Callable[..., dict], Callable[..., Awaitable[dict]]]
The main program function to execute
signature_dict : Dict[str, dspy.Signature]
Mapping of task names to signatures
registry : ParamRegistry
Parameter registry
signature_name2register_name : Dict[str, str]
Mapping of signature names to register names
"""
super().__init__()
self.program = program
self.predicts = []
seen = set()
for name, signature in signature_dict.items():
if name in seen:
raise ValueError(f"Duplicate name {name} in signature_dict")
seen.add(name)
self.predicts.append(dspy.Predict(signature, name=name))
self.registry = registry
self.signature_name2register_name = signature_name2register_name
def reset(self):
"""
Reset the module to its initial state.
"""
self.registry.reset()
for predict in self.predicts:
signature = predict.signature
signature_name = signature.__name__
register_name = self.signature_name2register_name[signature_name]
register_element = self.registry.get(register_name)
if isinstance(register_element, PromptTemplate):
predict.signature.instructions = register_element.instruction
predict.demos = register_element.demonstrations
elif isinstance(register_element, str):
predict.signature.instructions = register_element
predict.demos = []
else:
logger.warning(f"Unsupported register element type: {type(register_element)}")
# raise ValueError(f"Unsupported register element type: {type(register_element)}")
return self
def escape_braces(self, text):
"""
Escape all braces in the text.
Parameters
----------
text : str
Text that needs escaping
Returns
-------
str
Escaped text
"""
def helper(s, start=0):
result = ''
i = start
while i < len(s):
if s[i] == '{':
inner, new_i = helper(s, i + 1)
result += '{{' + inner + '}}'
i = new_i
elif s[i] == '}':
return result, i + 1
else:
result += s[i]
i += 1
return result, i
escaped, _ = helper(text)
return escaped
def _validate_prompt(self, prompt: str, input_names: List[str], verbose: bool = True) -> str:
"""
Validate if the generated prompt is valid. Currently only checks if required inputs are wrapped in braces.
Parameters
----------
prompt : str
The prompt to validate
input_names : List[str]
List of required input names
verbose : bool, optional
Whether to show detailed information, defaults to True
Returns
-------
str
Validated and potentially modified prompt
"""
modified_messages = []
required_inputs = input_names
missing_required_inputs = [name for name in required_inputs if f"{{{name}}}" not in prompt]
if missing_required_inputs:
input_values = "\n\n".join([f"{name}: {{{name}}}" for name in missing_required_inputs])
prompt += f"\n\nThe followings are some required input values: \n{input_values}"
modified_messages.append(f"added missing inputs: {', '.join(missing_required_inputs)}")
prompt = self.escape_braces(prompt)
for name in input_names:
prompt = prompt.replace(f"{{{{{name}}}}}", f"{{{name}}}")
prompt = prompt.replace(r"{{{{", r"{{").replace(r"}}}}", r"}}")
# if verbose and modified_messages:
# logger.warning("Prompt modified: " + " | ".join(modified_messages))
return prompt
def get_field_type(self, field: Field) -> str:
"""
Get the type of the field.
Parameters
----------
field : Field
The field to get type from
Returns
-------
str
The field type
"""
return field.json_schema_extra.get('__dspy_field_type') if field.json_schema_extra.get('__dspy_field_type') else None
def is_prompt_template(self, register_name: str) -> bool:
"""
Check if the register name is a prompt template.
Parameters
----------
register_name : str
The register name to check
Returns
-------
bool
Whether it is a prompt template
"""
return self.registry.get(register_name) is not None and isinstance(self.registry.get(register_name), PromptTemplate)
def get_demos(self, demos: list) -> List[dict]:
result = []
for demo in demos:
if isinstance(demo, dspy.Example):
demo = demo.toDict()
result.append(demo)
return result
def _inject_demos_to_string(self, instruction: str, demos: List[dict], input_names: List[str], output_names: List[str]) -> str:
"""
Inject demos to the instruction.
"""
if not demos:
return instruction
def _escape_braces(text: str) -> str:
return text.replace("{", "{{").replace("}", "}}")
def format_demo(demo: dict) -> str:
demo_str = "Inputs:\n"
inputs = {name: demo.get(name, "Not provided") for name in input_names}
demo_str += "\n".join([f"{name}:\n{_escape_braces(str(value))}" for name, value in inputs.items()])
demo_str += "\n\nOutputs:\n"
outputs = {name: demo.get(name, "Not provided") for name in output_names}
demo_str += "\n".join([f"{name}:\n{_escape_braces(str(value))}" for name, value in outputs.items()])
return demo_str
demos_string = "\n\n".join([f"Example {i+1}:\n{format_demo(demo)}" for i, demo in enumerate(demos)])
prompt = f"{instruction}\n\nThe following are some examples:\n{demos_string}"
return prompt
def sync_predict_inputs_to_program(self):
"""
Synchronize current input values from all predictors back to the registry.
This method ensures that any optimized parameters in the predictors' configurations
are properly reflected in the registry, which in turn affects program execution.
Synchronization process:
1. Iterate through all predictors
2. For each predictor, check its signature's input fields
3. If a field has a value in the predictor's config, update the registry
Note: Values in predictor configs take precedence as they may contain
optimized values from recent tuning iterations.
"""
for predict in self.predicts:
signature = predict.signature
instruction = signature.instructions
demos = predict.demos
input_names = [name for name, field in predict.signature.fields.items() if self.get_field_type(field) == 'input']
output_names = [name for name, field in predict.signature.fields.items() if self.get_field_type(field) == 'output']
signature_name = signature.__name__
register_name = self.signature_name2register_name[signature_name]
if self.is_prompt_template(register_name):
prompt_template: PromptTemplate = self.registry.get(register_name)
prompt_template.instruction = instruction
prompt_template.demonstrations = self.get_demos(demos)
self.registry.set(register_name, prompt_template)
else:
instruction = self._validate_prompt(instruction, input_names)
# self.registry.set(register_name, instruction)
prompt = self._inject_demos_to_string(instruction, self.get_demos(demos), input_names, output_names)
self.registry.set(register_name, prompt)
def constrcut_trace(self, execution_data: dict) -> dict:
"""
Construct the trace of the execution.
Parameters
----------
execution_data : dict
Execution data
Returns
-------
dict
Trace information
"""
trace: List[dict] = []
for predict in self.predicts:
input_names = [name for name, field in predict.signature.fields.items() if self.get_field_type(field) == 'input']
output_names = [name for name, field in predict.signature.fields.items() if self.get_field_type(field) == 'output']
input_dict = {}
output_dict = {}
# Check if input_names and output_names exist in execution data
for name in input_names:
if name not in execution_data:
logger.warning(f"Input {name} not found in execution data")
for name in output_names:
if name not in execution_data:
logger.warning(f"Output {name} not found in execution data")
# Add input_names and output_names from execution data to trace
for name in input_names:
if name in execution_data:
input_dict[name] = execution_data[name]
for name in output_names:
if name in execution_data:
output_dict[name] = execution_data[name]
trace_tuple = (predict, input_dict, output_dict)
trace.append(trace_tuple)
return trace
def forward(self, **kwargs) -> dict:
"""
Execute the program with synchronized parameters and optional inputs.
This method:
1. Synchronizes optimized prompts back to the program via registry
2. Executes the program (handles both sync and async functions)
3. Validates and returns the program's output
Parameters
----------
**kwargs : dict
Optional keyword arguments to pass to the program function
Returns
-------
dict
The program's execution results
Raises
------
ValueError
If the program doesn't return a dictionary
"""
# 1. Sync optimized prompts back to program
self.sync_predict_inputs_to_program()
# 2. Execute the program (handle both sync/async)
if asyncio.iscoroutinefunction(self.program):
output, execution_data = asyncio.run(self.program(**kwargs)) if kwargs else asyncio.run(self.program())
else:
output, execution_data = self.program(**kwargs) if kwargs else self.program()
trace = self.constrcut_trace(execution_data)
# Use context manager to set trace
if dspy.settings.trace is not None:
dspy_trace = dspy.settings.trace
dspy_trace.extend(trace)
return output
def deepcopy(self):
"""
Deep copy the module.
This is a tweak to the default Python deepcopy that only deep copies `self.parameters()`,
and for other attributes, we just do a shallow copy.
Returns
-------
PromptTuningModule
A deep copy of the module
"""
try:
# If the instance itself is copyable, we can just deep copy it
new_instance = copy.deepcopy(self)
setattr(new_instance, "program", self.program)
return new_instance
except Exception:
pass
# Create an empty instance
new_instance = self.__class__.__new__(self.__class__)
# Set attributes of the copied instance
for attr, value in self.__dict__.items():
if isinstance(value, dspy.Module):
setattr(new_instance, attr, value.deepcopy())
else:
try:
# Try to deep copy the attribute
setattr(new_instance, attr, copy.deepcopy(value))
except Exception:
try:
# Fall back to shallow copy if deep copy fails
setattr(new_instance, attr, copy.copy(value))
except Exception:
# If even shallow copy fails, just copy the reference
setattr(new_instance, attr, value)
# set the new instance's program to the original program
setattr(new_instance, "program", self.program)
return new_instance
def save(self, path, save_program=False):
"""Save the module.
Save the module to a directory or a file. There are two modes:
- `save_program=False`: Save only the state of the module to a json or pickle file, based on the value of
the file extension.
- `save_program=True`: Save the whole module to a directory via cloudpickle, which contains both the state and
architecture of the model.
We also save the dependency versions, so that the loaded model can check if there is a version mismatch on
critical dependencies or DSPy version.
Args:
path (str): Path to the saved state file, which should be a .json or .pkl file when `save_program=False`,
and a directory when `save_program=True`.
save_program (bool): If True, save the whole module to a directory via cloudpickle, otherwise only save
the state.
"""
metadata = {}
metadata["dependency_versions"] = get_dependency_versions()
path = Path(path)
if not path.is_dir():
# file
if not path.parent.exists():
path.parent.mkdir(parents=True)
else:
# directory
if not path.exists():
# Create the directory (and any parent directories)
if not path.exists():
# Create the directory (and any parent directories)
path.mkdir(parents=True)
if hasattr(self.program, "save"):
self.program.save(str(path))
return
if save_program:
if path.suffix:
raise ValueError(
f"`path` must point to a directory without a suffix when `save_program=True`, but received: {path}"
)
if path.exists() and not path.is_dir():
raise NotADirectoryError(f"The path '{path}' exists but is not a directory.")
try:
with open(path / "program.pkl", "wb") as f:
cloudpickle.dump(self, f)
except Exception as e:
raise RuntimeError(
f"Saving failed with error: {e}. Please remove the non-picklable attributes from your DSPy program, "
"or consider using state-only saving by setting `save_program=False`."
)
with open(path / "metadata.json", "w") as f:
ujson.dump(metadata, f, indent=4)
return
state = self.dump_state()
state["metadata"] = metadata
if path.suffix == ".json":
try:
with open(path, "w") as f:
f.write(ujson.dumps(state, indent=4))
except Exception as e:
raise RuntimeError(
f"Failed to save state to {path} with error: {e}. Your DSPy program may contain non "
"json-serializable objects, please consider saving the state in .pkl by using `path` ending "
"with `.pkl`, or saving the whole program by setting `save_program=True`."
)
elif path.suffix == ".pkl":
with open(path, "wb") as f:
cloudpickle.dump(state, f)
else:
raise ValueError(f"`path` must end with `.json` or `.pkl` when `save_program=False`, but received: {path}")
def load(self, path):
"""Load the saved module. You may also want to check out dspy.load, if you want to
load an entire program, not just the state for an existing program.
Args:
path (str): Path to the saved state file, which should be a .json or a .pkl file
"""
path = Path(path)
if hasattr(self.program, "load"):
self.program.load(str(path))
# todo: sync the program's parameters to the signature
return
if path.suffix == ".json":
with open(path) as f:
state = ujson.loads(f.read())
elif path.suffix == ".pkl":
with open(path, "rb") as f:
state = cloudpickle.load(f)
else:
raise ValueError(f"`path` must end with `.json` or `.pkl`, but received: {path}")
dependency_versions = get_dependency_versions()
saved_dependency_versions = state["metadata"]["dependency_versions"]
for key, saved_version in saved_dependency_versions.items():
if dependency_versions[key] != saved_version:
logger.warning(
f"There is a mismatch of {key} version between saved model and current environment. "
f"You saved with `{key}=={saved_version}`, but now you have "
f"`{key}=={dependency_versions[key]}`. This might cause errors or performance downgrade "
"on the loaded model, please consider loading the model in the same environment as the "
"saving environment."
)
self.load_state(state)
self.sync_predict_inputs_to_program() # sync the signature values to the program parameters
|