File size: 19,138 Bytes
b43d8da | 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 | """Custom trainer + dataset adapter (docs/modules/training.md §2.2, §3.2.3).
Two public types:
- :class:`EpisodeDatasetAdapter` — stateless iterable feeding
``GRPOTrainer.train_dataset``. Each ``__iter__`` tick yields
``{"prompt": str, "_meta": {...}}`` where ``_meta`` carries the
``GoalSpec``, the monotonically-derived ``episode_seed``, the curriculum
``stage``, and the ``language_weights``. One call to
``task_generator.generate`` per step; one call to
``tokenizer.apply_chat_template(messages, tokenize=False,
add_generation_prompt=True)`` to render the prompt.
- :class:`DriftCallGRPOTrainer` — ``GRPOTrainer`` subclass whose
``_generate_and_score_completions`` override runs G multi-turn episodes
via a caller-provided ``RolloutGroupFn`` and plumbs the resulting
frozen ``Episode`` tuple into ``reward_fn`` (step_13) before handing the
G reward scalars + padded completions back to the inherited GRPO
advantage / KL / optimizer step path. **The inherited code path is
untouched** (training.md §3.2.3).
``trl`` and ``torch`` are imported lazily. Pure-Python fallbacks for
``_generate_and_score_completions`` are provided so the class shape
can be verified on CPU-only CI.
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Literal, Protocol, cast
if TYPE_CHECKING: # pragma: no cover - typing only
from collections.abc import Callable, Iterator
from cells.step_13_grpo_config import BETA_KL
PINNED_SYSTEM_PROMPT: str = (
"You are a concierge assistant. Use the provided tools. "
"Respond in the caller's language. Submit with calibrated confidence."
)
LanguageCode = Literal["hi", "ta", "kn", "en", "hinglish"]
class EpisodeSampler(Protocol):
"""Draws a ``GoalSpec`` for one prompt slot (training.md §2.2)."""
def __call__(self, step: int) -> Any: ...
class EnvFactory(Protocol):
"""Returns a fresh ``DriftCallEnv`` per rollout (training.md §3.2)."""
def __call__(self) -> Any: ...
class RolloutGroupFn(Protocol):
"""Runs G multi-turn rollouts sharing one goal.
Returns a tuple ``(episodes, completions)`` of length G each.
"""
def __call__(
self,
*,
model: Any,
tokenizer: Any,
goal: Any,
episode_seed: int,
num_generations: int,
env_factory: EnvFactory,
) -> tuple[tuple[Any, ...], tuple[str, ...]]: ...
@dataclass(frozen=True)
class AdapterRecord:
"""Frozen view of one :class:`EpisodeDatasetAdapter` yield.
Tests consume this view rather than dict-typing ``_meta`` inline.
"""
prompt: str
goal: Any
episode_seed: int
stage: Literal[1, 2, 3]
language_weights: dict[LanguageCode, float]
def render_initial_prompt(tokenizer: Any, goal: Any) -> str:
"""Render the turn-0 chat template (training.md §3.2.1).
Messages: pinned system prompt + ``goal.seed_utterance`` as the user
turn. ``add_generation_prompt=True`` tells Gemma to emit an assistant
turn. Tool schemas live in later turns so only these two messages
appear at ``step == 0``.
"""
seed_utterance = getattr(goal, "seed_utterance", "")
messages: list[dict[str, str]] = [
{"role": "system", "content": PINNED_SYSTEM_PROMPT},
{"role": "user", "content": seed_utterance},
]
result = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
)
return str(result)
class EpisodeDatasetAdapter:
"""Stateless streaming dataset (training.md §2.2).
Constructor signature matches training.md §2.2: a ``task_gen`` callable
accepting ``(seed, stage, language_weights)``, an ``env_factory``
producing fresh envs, the curriculum ``stage``, a ``stage_base_seed``
used to derive per-step ``episode_seed``, the per-language sampling
``language_weights``, and the ``tokenizer`` used to render prompts.
Iteration is infinite — exactly one record per GRPO training step.
Step counter is local to ``__iter__`` so resume simply restarts from
whatever step TRL's ``resume_from_checkpoint`` restores.
"""
def __init__(
self,
*,
task_gen: Callable[..., Any],
env_factory: EnvFactory,
stage: Literal[1, 2, 3],
stage_base_seed: int,
language_weights: dict[LanguageCode, float],
tokenizer: Any,
) -> None:
self.task_gen = task_gen
self.env_factory = env_factory
self.stage: Literal[1, 2, 3] = stage
self.stage_base_seed = stage_base_seed
self.language_weights = dict(language_weights)
self.tokenizer = tokenizer
def _build_record(self, step: int) -> dict[str, Any]:
episode_seed = self.stage_base_seed + step
goal = self.task_gen(
seed=episode_seed,
stage=self.stage,
language_weights=self.language_weights,
)
prompt = render_initial_prompt(self.tokenizer, goal)
return {
"prompt": prompt,
"_meta": {
"goal": goal,
"episode_seed": episode_seed,
"stage": self.stage,
"language_weights": dict(self.language_weights),
},
}
def __iter__(self) -> Iterator[dict[str, Any]]:
step = 0
while True:
yield self._build_record(step)
step += 1
def __len__(self) -> int:
"""Length sentinel for TRL 0.24+ ``RepeatSampler``.
The dataset is logically infinite (one record per GRPO step), but
TRL 0.24's ``RepeatSampler`` calls ``len(data_source)`` to size the
sampler. Returning a large finite number lets training proceed; the
actual step count is bounded by ``GRPOConfig.max_steps``.
"""
return 1_000_000
def __getitem__(self, idx: int) -> dict[str, Any]:
"""Map-style indexing for TRL 0.24+ DataLoader.
TRL 0.24 treats the train_dataset as a Map-style dataset and looks
records up by integer index. We honour the contract by deriving the
record purely from ``idx`` — the adapter is stateless so any index
produces a deterministic ``(prompt, _meta)`` pair for that step.
"""
return self._build_record(int(idx))
def peek(self, step: int) -> AdapterRecord:
"""Materialize the record at ``step`` without advancing iteration.
Used by tests (§1.2 U14–U18) to assert record shape at arbitrary
steps without consuming a generator.
"""
rec = self._build_record(step)
meta = rec["_meta"]
return AdapterRecord(
prompt=rec["prompt"],
goal=meta["goal"],
episode_seed=meta["episode_seed"],
stage=meta["stage"],
language_weights=meta["language_weights"],
)
def _import_grpo_trainer() -> type[Any]:
"""Lazy import of ``trl.GRPOTrainer``; isolated for mocking in tests."""
from trl import GRPOTrainer
return cast("type[Any]", GRPOTrainer)
def _make_driftcall_init(
base_cls: type[Any],
) -> Callable[..., None]:
"""Build an ``__init__`` bound to ``base_cls``; avoids super() recursion
when the returned class is itself further subclassed.
DriftCall-specific kwargs added on top of ``GRPOTrainer.__init__``:
- ``rollout_group_fn``, ``env_factory``, ``reward_fn_driftcall`` — the
multi-turn rollout override surface (see class docstring).
- ``enable_adaptive_kl`` (default ``True``) — auto-attach an
:class:`AdaptiveKLCallback` so β retargets to the measured KL each
logging tick (training.md §3.3.1). Set ``False`` to disable.
- ``adaptive_kl_target`` — override the default ``target_kl=BETA_KL``.
- ``adaptive_kl_kp`` — override the proportional gain.
- ``adaptive_kl_beta_min`` / ``adaptive_kl_beta_max`` — override clamp
bounds.
"""
def _init(
self: Any,
*args: Any,
rollout_group_fn: RolloutGroupFn,
env_factory: EnvFactory,
reward_fn_driftcall: Callable[..., list[float]],
enable_adaptive_kl: bool = True,
adaptive_kl_target: float | None = None,
adaptive_kl_kp: float = DEFAULT_KP,
adaptive_kl_beta_min: float = DEFAULT_BETA_MIN,
adaptive_kl_beta_max: float = DEFAULT_BETA_MAX,
**kwargs: Any,
) -> None:
# TRL 0.24 made ``reward_funcs`` a required arg on GRPOTrainer.
# Our custom ``_generate_and_score_completions`` short-circuits the
# base reward path entirely (calls ``reward_fn_driftcall`` directly),
# so the parent's ``reward_funcs`` value is never invoked. Pass a
# placeholder identity reward to satisfy the signature on TRL>=0.24.
if "reward_funcs" not in kwargs:
def _placeholder_reward(
completions: Any = None,
**_unused: Any,
) -> list[float]:
n = len(completions) if completions is not None else 0
return [0.0] * n
kwargs["reward_funcs"] = [_placeholder_reward]
base_cls.__init__(self, *args, **kwargs)
self.rollout_group_fn = rollout_group_fn
self.env_factory = env_factory
self.reward_fn_driftcall = reward_fn_driftcall
if enable_adaptive_kl:
target = (
adaptive_kl_target if adaptive_kl_target is not None else BETA_KL
)
callback = AdaptiveKLCallback(
target_kl=target,
kp=adaptive_kl_kp,
beta_min=adaptive_kl_beta_min,
beta_max=adaptive_kl_beta_max,
)
self.adaptive_kl_callback = callback
add_callback = getattr(base_cls, "add_callback", None)
if callable(add_callback):
# Production path (TRL ≥ 0.23): register through the TRL
# callback handler so ``on_log`` fires alongside default
# loggers with the correct ``args``/``state``/``control``.
self.add_callback(callback)
else:
# Fallback: minimal bases in tests lack ``add_callback``.
# Keep a private list so callers can still invoke the hook.
if not hasattr(self, "_driftcall_callbacks"):
self._driftcall_callbacks = []
self._driftcall_callbacks.append(callback)
else:
self.adaptive_kl_callback = None
return _init
def _driftcall_generate_and_score_completions(
self: Any, inputs: list[dict[str, Any]]
) -> dict[str, Any]:
"""Run the multi-turn rollout, then call ``reward_fn``.
Expects ``inputs`` to carry one row per prompt slot with the
``_meta`` dict produced by :class:`EpisodeDatasetAdapter`.
Returns a dict with keys ``episodes``, ``completions``, ``rewards``,
``prompts`` — each length G (num_generations).
"""
if not inputs:
raise ValueError("inputs must be a non-empty list")
row = inputs[0]
meta = row["_meta"]
prompt = row["prompt"]
goal = meta["goal"]
episode_seed = meta["episode_seed"]
num_generations = int(getattr(self.args, "num_generations", 8))
episodes, completions = self.rollout_group_fn(
model=self.model,
tokenizer=self.processing_class,
goal=goal,
episode_seed=episode_seed,
num_generations=num_generations,
env_factory=self.env_factory,
)
if len(episodes) != num_generations or len(completions) != num_generations:
raise ValueError(
f"rollout_group_fn produced {len(episodes)} episodes and "
f"{len(completions)} completions; expected {num_generations} each"
)
prompts = [prompt] * num_generations
metas = [dict(meta) for _ in range(num_generations)]
rewards = self.reward_fn_driftcall(
prompts=prompts,
completions=list(completions),
_meta=metas,
episodes=list(episodes),
)
return {
"episodes": episodes,
"completions": completions,
"rewards": rewards,
"prompts": prompts,
}
def make_driftcall_grpo_trainer_cls(base_cls: type[Any] | None = None) -> type[Any]:
"""Build the :class:`DriftCallGRPOTrainer` class bound to ``base_cls``.
Default ``base_cls`` is ``trl.GRPOTrainer`` (imported lazily). Tests
pass a stub base class so they can exercise the override path without
TRL installed.
GRPOTrainer subclass with multi-turn rollout override
(training.md §3.2.3). Construction adds three DriftCall-specific
kwargs over the standard ``GRPOTrainer.__init__``:
- ``rollout_group_fn``: :class:`RolloutGroupFn` running G multi-turn
episodes and returning ``(episodes, completions)``.
- ``env_factory``: :class:`EnvFactory` producing a fresh
``DriftCallEnv`` per rollout.
- ``reward_fn_driftcall``: the step_13 ``reward_fn`` — called
directly with the frozen ``Episode`` tuple after rollout.
``_generate_and_score_completions`` replaces the TRL default.
Advantage + KL + optimizer step paths are inherited unchanged.
"""
resolved_base: type[Any] = (
base_cls if base_cls is not None else _import_grpo_trainer()
)
return type(
"DriftCallGRPOTrainer",
(resolved_base,),
{
"__init__": _make_driftcall_init(resolved_base),
"_generate_and_score_completions": _driftcall_generate_and_score_completions,
"__doc__": "GRPOTrainer subclass with multi-turn rollout override.",
},
)
def driftcall_grpo_trainer_methods() -> tuple[str, ...]:
"""Return the method names the subclass overrides (introspection helper).
Used by the shape test (U in §1.x) to verify the override surface.
"""
return ("__init__", "_generate_and_score_completions")
# ---------------------------------------------------------------------------
# Adaptive KL controller (training.md §3.3 — retarget β from measured KL)
# ---------------------------------------------------------------------------
DEFAULT_BETA_MIN: float = 0.001
DEFAULT_BETA_MAX: float = 1.0
DEFAULT_KP: float = 2.0
def _trainer_callback_base() -> type:
"""Return ``transformers.TrainerCallback`` if importable, else ``object``.
Importing transformers lazily keeps step_14 importable on CPU-only CI
runners that don't have transformers installed.
"""
try:
from transformers.trainer_callback import TrainerCallback
return TrainerCallback
except Exception:
return object
class AdaptiveKLCallback(_trainer_callback_base()): # type: ignore[misc]
"""Retarget β each step based on the ratio of measured KL to ``target_kl``.
Proportional controller with symmetric log-space update:
err = (kl - target_kl) / target_kl
new_beta = beta * exp(kp * err)
new_beta = clamp(new_beta, beta_min, beta_max)
When ``kl`` matches ``target_kl``, ``err == 0`` and β is left unchanged.
Safe on missing / NaN / non-numeric KL signals (no-op, no exception).
Inherits from :class:`transformers.trainer_callback.TrainerCallback` when
available (production path) so all the no-op callback events
(``on_train_begin``, ``on_step_begin``, etc.) come for free; falls back
to ``object`` on CPU-only CI when transformers is not installed.
"""
def __init__(
self,
target_kl: float = BETA_KL,
*,
kp: float = DEFAULT_KP,
beta_min: float = DEFAULT_BETA_MIN,
beta_max: float = DEFAULT_BETA_MAX,
) -> None:
if target_kl <= 0.0:
raise ValueError(f"target_kl must be > 0; got {target_kl}")
if beta_min <= 0.0 or beta_max <= 0.0:
raise ValueError(
f"beta bounds must be > 0; got min={beta_min}, max={beta_max}"
)
if beta_min > beta_max:
raise ValueError(
f"beta_min ({beta_min}) must be <= beta_max ({beta_max})"
)
self.target_kl = float(target_kl)
self.kp = float(kp)
self.beta_min = float(beta_min)
self.beta_max = float(beta_max)
def _coerce_kl(self, raw: Any) -> float | None:
"""Return a finite float or ``None`` — propagates no-op on bad input."""
try:
value = float(raw)
except (TypeError, ValueError):
return None
if math.isnan(value) or math.isinf(value):
return None
return value
def _next_beta(self, beta: float, kl: float) -> tuple[float, bool, bool]:
"""Return ``(new_beta, clamped_to_min, clamped_to_max)``."""
err = (kl - self.target_kl) / self.target_kl
# Clamp the exponent so extreme KL spikes don't overflow math.exp;
# the result is clamped anyway and exp(±50) easily saturates either bound.
exponent = max(-50.0, min(50.0, self.kp * err))
scaled = beta * math.exp(exponent)
if scaled <= self.beta_min:
return self.beta_min, True, False
if scaled >= self.beta_max:
return self.beta_max, False, True
return scaled, False, False
def on_log(
self,
args: Any,
state: Any,
control: Any,
*,
logs: dict[str, Any] | None = None,
**_kwargs: Any,
) -> Any:
"""TRL hook — called with every ``trainer.log(...)`` dict.
On a well-formed KL signal: mutates ``args.beta`` with the new
coefficient and writes five diagnostic fields back into ``logs``
so TRL's default reporter forwards them to wandb / CSV / etc.:
- ``train/beta_adaptive`` current KL coefficient
- ``train/kl_measured`` sanitised KL input
- ``train/kl_target`` constant — aids chart-by-reference
- ``train/beta_clamped_to_min`` 0/1 — fires on collapse
- ``train/beta_clamped_to_max`` 0/1 — fires on runaway divergence
"""
if logs is None:
return control
if "kl" not in logs:
return control
kl = self._coerce_kl(logs["kl"])
if kl is None:
return control
beta = float(getattr(args, "beta", BETA_KL))
new_beta, clamped_lo, clamped_hi = self._next_beta(beta, kl)
args.beta = new_beta
logs["train/beta_adaptive"] = new_beta
logs["train/kl_measured"] = kl
logs["train/kl_target"] = self.target_kl
logs["train/beta_clamped_to_min"] = 1 if clamped_lo else 0
logs["train/beta_clamped_to_max"] = 1 if clamped_hi else 0
return control
__all__ = [
"AdapterRecord",
"AdaptiveKLCallback",
"DEFAULT_BETA_MAX",
"DEFAULT_BETA_MIN",
"DEFAULT_KP",
"EnvFactory",
"EpisodeDatasetAdapter",
"EpisodeSampler",
"LanguageCode",
"PINNED_SYSTEM_PROMPT",
"RolloutGroupFn",
"driftcall_grpo_trainer_methods",
"make_driftcall_grpo_trainer_cls",
"render_initial_prompt",
]
|