File size: 45,520 Bytes
1c8c60e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 | """
TODO: Add coefficients for losses (depend on total number of tokens or batch)
TODO: adapt reinforce step for torch.compile
TODO: add lr schedulers support
"""
import logging
import os
import pickle
import sys
from abc import ABC, abstractmethod
from typing import Callable, Literal, Union
import numpy as np
import torch
import torch.nn.functional as F
from accelerate import Accelerator
from pandas._libs.tslibs.offsets import CBMonthBegin
from peft import LoraConfig
from torch.nn.utils.rnn import pad_sequence
from transformers import AutoModelForCausalLM, AutoTokenizer
from mllm.markov_games.rollout_tree import *
from mllm.markov_games.rollout_tree import RolloutTreeRootNode
from mllm.training.annealing_methods import sigmoid_annealing
from mllm.training.credit_methods import (
get_discounted_returns,
get_generalized_advantage_estimates,
get_rloo_credits,
whiten_advantages,
whiten_advantages_time_step_wise,
)
from mllm.training.tally_metrics import Tally
from mllm.training.tally_rollout import RolloutTally, RolloutTallyItem
from mllm.training.tally_tokenwise import ContextualizedTokenwiseTally
from mllm.training.tokenize_chats import *
from mllm.training.tokenize_chats import process_training_chat
from mllm.training.training_data_utils import *
from mllm.training.training_data_utils import (
TrainingBatch,
TrajectoryBatch,
get_tokenwise_credits,
)
from mllm.utils.resource_context import resource_logger_context
logger = logging.getLogger(__name__)
logger.addHandler(logging.StreamHandler(sys.stdout))
@dataclass
class TrainerAnnealingState:
annealing_step_counter: int = 0
class BaseTrainer(ABC):
"""
Trainer
"""
def __init__(
self,
policy: AutoModelForCausalLM,
policy_optimizer: torch.optim.Optimizer,
critic: Union[AutoModelForCausalLM, None],
critic_optimizer: Union[torch.optim.Optimizer, None],
tokenizer: AutoTokenizer,
lr_scheduler: torch.optim.lr_scheduler.LRScheduler,
critic_lr_scheduler: Union[torch.optim.lr_scheduler.LRScheduler, None],
######################################################################
entropy_coeff: float,
entropy_topk: int,
entropy_mask_regex: Union[str, None],
kl_coeff: float,
gradient_clipping: Union[float, None],
restrict_tokens: Union[list[str], None],
mini_batch_size: int,
use_gradient_checkpointing: bool,
temperature: float,
device: str,
whiten_advantages: bool,
whiten_advantages_time_step_wise: bool,
use_gae: bool,
use_gae_lambda_annealing: bool,
gae_lambda_annealing_limit: float,
gae_lambda_annealing_method: Literal["sigmoid_annealing"],
gae_lambda_annealing_method_params: dict,
pg_loss_normalization: Literal["batch", "nb_tokens"],
use_rloo: bool,
skip_discounted_state_visitation: bool,
discount_factor: float,
enable_tokenwise_logging: bool,
save_path: str,
reward_normalizing_constant: float = 1.0,
critic_loss_type: Literal["mse", "huber"] = "huber",
exploration_prompts_to_remove: list[str] = [],
filter_higher_refprob_tokens_kl: bool = False,
truncated_importance_sampling_ratio_cap: float = 0.0,
importance_sampling_strategy: Literal[
"per_token", "per_sequence"
] = "per_token",
):
"""
Initialize the REINFORCE trainer with reward shaping for multi-agent or single-agent training.
Args:
model (AutoModelForCausalLM): The main policy model.
tokenizer (AutoTokenizer): Tokenizer for the model.
optimizer (torch.optim.Optimizer): Optimizer for the policy model.
lr_scheduler (torch.optim.lr_scheduler.LRScheduler): Learning rate scheduler for the policy model.
critic (AutoModelForCausalLM or None): Critic model for value estimation (optional).
critic_optimizer (torch.optim.Optimizer or None): Optimizer for the critic model (optional).
critic_lr_scheduler (torch.optim.lr_scheduler.LRScheduler or None): LR scheduler for the critic (optional).
config (RtConfig): Configuration object for training.
"""
self.tokenizer = tokenizer
# self.tokenizer.padding_side = "left" # needed for flash attention
if self.tokenizer.pad_token_id is None:
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
self.lr_scheduler = lr_scheduler
self.accelerator = Accelerator()
(
self.policy,
self.policy_optimizer,
self.critic,
self.critic_optimizer,
) = self.accelerator.prepare(policy, policy_optimizer, critic, critic_optimizer)
self.critic_lr_scheduler = critic_lr_scheduler
self.tally = Tally()
if use_gradient_checkpointing == True:
self.policy.gradient_checkpointing_enable(dict(use_reentrant=False))
if critic is not None:
self.critic.gradient_checkpointing_enable(dict(use_reentrant=False))
self.save_path = save_path
# Load trainer state if it exists
self.trainer_annealing_state_path = os.path.join(
self.save_path, "trainer_annealing_state.pkl"
)
if os.path.exists(self.trainer_annealing_state_path):
logger.info(
f"Loading trainer state from {self.trainer_annealing_state_path}"
)
self.trainer_annealing_state = pickle.load(
open(self.trainer_annealing_state_path, "rb")
)
else:
self.trainer_annealing_state = TrainerAnnealingState()
# Load policy optimizer state if it exists
self.policy_optimizer_path = os.path.join(
self.save_path, "policy_optimizer_state.pt"
)
if os.path.exists(self.policy_optimizer_path):
logger.info(
f"Loading policy optimizer state from {self.policy_optimizer_path}"
)
self.policy_optimizer.load_state_dict(
torch.load(self.policy_optimizer_path)
)
# Load critic optimizer state if it exists
self.critic_optimizer_path = os.path.join(
self.save_path, "critic_optimizer_state.pt"
)
if (
os.path.exists(self.critic_optimizer_path)
and self.critic_optimizer is not None
):
logger.info(
f"Loading critic optimizer state from {self.critic_optimizer_path}"
)
self.critic_optimizer.load_state_dict(
torch.load(self.critic_optimizer_path)
)
self.device = self.accelerator.device
self.entropy_coeff = entropy_coeff
self.entropy_topk = entropy_topk
self.entropy_mask_regex = entropy_mask_regex
self.kl_coeff = kl_coeff
self.gradient_clipping = gradient_clipping
self.restrict_tokens = restrict_tokens
self.mini_batch_size = mini_batch_size
self.use_gradient_checkpointing = use_gradient_checkpointing
self.temperature = temperature
self.use_gae = use_gae
self.whiten_advantages = whiten_advantages
self.whiten_advantages_time_step_wise = whiten_advantages_time_step_wise
self.use_rloo = use_rloo
self.skip_discounted_state_visitation = skip_discounted_state_visitation
self.use_gae_lambda_annealing = use_gae_lambda_annealing
self.gae_lambda_annealing_limit = gae_lambda_annealing_limit
if use_gae_lambda_annealing:
self.gae_lambda_annealing_method: Callable[
[int], float
] = lambda step: eval(gae_lambda_annealing_method)(
step=step, **gae_lambda_annealing_method_params
)
self.discount_factor = discount_factor
self.enable_tokenwise_logging = enable_tokenwise_logging
self.reward_normalizing_constant = reward_normalizing_constant
self.pg_loss_normalization = pg_loss_normalization
self.critic_loss_type = critic_loss_type
self.exploration_prompts_to_remove = exploration_prompts_to_remove
# Common containers used by all trainers
self.training_data: dict = {}
self.debug_path_list: list[str] = []
self.policy_gradient_data = None
self.tally = Tally()
self.rollout_tally = RolloutTally()
self.tokenwise_tally: Union[ContextualizedTokenwiseTally, None] = None
self.filter_higher_refprob_tokens_kl = filter_higher_refprob_tokens_kl
self.truncated_importance_sampling_ratio_cap = (
truncated_importance_sampling_ratio_cap
)
self.importance_sampling_strategy = importance_sampling_strategy
def mask_non_restricted_token_logits(self, logits: torch.Tensor) -> torch.Tensor:
"""
Masks logits so that only allowed tokens (as specified in config.restrict_tokens)
and the EOS token are active.
All other logits are set to -inf, effectively removing them from the softmax.
Args:
logits (torch.Tensor): The logits tensor of shape (B, S, V).
Returns:
torch.Tensor: The masked logits tensor.
"""
# TODO: verify. Not sure what we do here is differentiable
# also, we recompute for nothing
if self.restrict_tokens is not None:
allowed_token_ids = []
for token in self.restrict_tokens:
token_ids = self.tokenizer(token, add_special_tokens=False)["input_ids"]
allowed_token_ids.append(token_ids[0])
allowed_token_ids.append(
self.tokenizer.eos_token_id
) # This token should always be active
allowed_token_ids = torch.tensor(allowed_token_ids, device=logits.device)
# Mask log_probs and probs to only allowed tokens
mask = torch.zeros_like(logits).bool() # (B, S, V)
mask[..., allowed_token_ids] = True
logits = torch.where(
mask,
logits,
torch.tensor(-float("inf"), device=logits.device),
)
return logits
# def get_gradient_magnitude(self, loss_term: torch.Tensor) -> float:
# """
# Computes the L2 norm of the gradients of the given loss term with respect to the model parameters.
# Args:
# loss_term (torch.Tensor): The loss tensor to compute gradients for.
# Returns:
# float: The L2 norm of the gradients, or 0.0 if no gradients are present.
# """
# with torch.no_grad():
# grads = torch.autograd.grad(
# loss_term,
# [p for p in self.policy.parameters() if p.requires_grad],
# retain_graph=True,
# allow_unused=True,
# )
# grads = [g for g in grads if g is not None]
# if not grads:
# return torch.tensor(0.0, device=loss_term.device)
# return torch.norm(torch.stack([g.norm(2) for g in grads])).item()
def apply_reinforce_step(
self,
training_batch: TrainingBatch,
) -> None:
"""
Applies a single REINFORCE policy gradient step using the provided batch of rollouts.
Handles batching, loss computation (including entropy and KL regularization), gradient accumulation, and optimizer step.
Optionally logs various metrics and statistics.
Args:
paths (list[str]): List of game complete file paths for each rollout.
contexts (list[torch.Tensor]): List of context tensors for each rollout.
credits (list[torch.Tensor]): List of credit tensors (rewards/advantages) for each rollout.
action_masks (list[torch.Tensor]): List of action mask tensors for each rollout.
"""
with resource_logger_context(logger, "Apply reinforce step"):
self.policy.train()
mb_size = self.mini_batch_size
nb_rollouts = len(training_batch)
# Initialize running mean logs
running_mean_logs = {
"rl_objective": 0.0,
"policy_gradient_loss": 0.0,
"policy_gradient_norm": 0.0,
"log_probs": 0.0,
"credits": 0.0,
"entropy": 0.0,
"engine_log_probs_diff_clampfrac": 0.0,
"tis_imp_ratio": 0.0,
"ref_log_probs_diff_clampfrac": 0.0,
"higher_refprob_frac": 0.0,
"tis_imp_ratio_clampfrac": 0.0,
}
if self.entropy_coeff != 0.0:
running_mean_logs["entropy"] = 0.0
if self.kl_coeff != 0.0:
running_mean_logs["kl_divergence"] = 0.0
# Get total number of tokens generated
total_tokens_generated = 0
for att_mask in training_batch.batch_action_mask:
total_tokens_generated += att_mask.sum()
# Obtain loss normalization
if self.pg_loss_normalization == "nb_tokens":
normalization_factor = total_tokens_generated
elif self.pg_loss_normalization == "batch":
normalization_factor = np.ceil(nb_rollouts / mb_size).astype(int)
else:
raise ValueError(
f"Invalid pg_loss_normalization: {self.pg_loss_normalization}"
)
# Gradient accumulation for each mini-batch
for mb in range(0, nb_rollouts, mb_size):
logger.info(f"Processing mini-batch {mb} of {nb_rollouts}")
loss = 0.0
training_mb = training_batch[mb : mb + mb_size]
training_mb = training_mb.get_padded_tensors()
training_mb.to(self.device)
(
tokens_mb,
action_mask_mb,
entropy_mask_mb,
credits_mb,
engine_log_probs_mb,
timesteps_mb,
) = (
training_mb.batch_input_ids,
training_mb.batch_action_mask,
training_mb.batch_entropy_mask,
training_mb.batch_credits,
training_mb.batch_engine_log_probs,
training_mb.batch_timesteps,
)
# Next token prediction
contexts_mb = tokens_mb[:, :-1]
shifted_contexts_mb = tokens_mb[:, 1:]
action_mask_mb = action_mask_mb[:, 1:]
entropy_mask_mb = entropy_mask_mb[:, 1:]
credits_mb = credits_mb[:, 1:]
engine_log_probs_mb = engine_log_probs_mb[:, 1:]
timesteps_mb = timesteps_mb[:, 1:]
if self.enable_tokenwise_logging:
self.tokenwise_tally.set_action_mask(action_mask=action_mask_mb)
self.tokenwise_tally.set_range(range=(mb, mb + mb_size))
self.tokenwise_tally.add_contexts(contexts=contexts_mb)
self.tokenwise_tally.add_data(
metric_id="next_token",
metrics=shifted_contexts_mb,
to_tids=True,
)
self.tokenwise_tally.add_data(
metric_id="entropy_mask",
metrics=entropy_mask_mb,
)
if self.enable_tokenwise_logging:
self.tokenwise_tally.add_data(
metric_id="next_token_credit", metrics=credits_mb
)
# Forward pass + cast to FP-32 for higher prec.
# TODO: create attention mask if not relying on default (assume causal llm)
logits = self.policy(input_ids=contexts_mb)[0] # (B, S, V)
# Mask non-restricted tokens
if self.restrict_tokens is not None:
logits = self.mask_non_restricted_token_logits(logits)
logits /= self.temperature # (B, S, V)
# Compute new log probabilities
log_probs = F.log_softmax(logits, dim=-1) # (B, S, V)
# Get log probabilities of actions taken during rollouts
action_log_probs = log_probs.gather(
dim=-1, index=shifted_contexts_mb.unsqueeze(-1)
).squeeze(
-1
) # (B, S)
if self.pg_loss_normalization == "batch":
den_running_mean = action_mask_mb.sum() * normalization_factor
else:
den_running_mean = normalization_factor
running_mean_logs["log_probs"] += (
action_log_probs * action_mask_mb
).sum().item() / den_running_mean
running_mean_logs["credits"] += (
credits_mb * action_mask_mb
).sum().item() / den_running_mean
if self.enable_tokenwise_logging:
self.tokenwise_tally.add_data(
metric_id="next_token_log_prob",
metrics=action_log_probs,
)
self.tokenwise_tally.add_data(
metric_id="engine_next_token_log_prob",
metrics=engine_log_probs_mb,
)
self.tokenwise_tally.add_data(
metric_id="next_token_prob",
metrics=torch.exp(action_log_probs),
)
top_k_indices = torch.topk(logits, k=5, dim=-1).indices
self.tokenwise_tally.add_data(
metric_id=f"top_{5}_tids",
metrics=top_k_indices,
to_tids=True,
)
self.tokenwise_tally.add_data(
metric_id=f"top_{5}_probs",
metrics=torch.exp(log_probs).gather(
dim=-1, index=top_k_indices
),
)
rewarded_action_log_probs = (
action_mask_mb * credits_mb * action_log_probs
)
# (B, S)
INVALID_LOGPROB = 1.0
CLAMP_VALUE = 40.0
masked_action_log_probs = torch.masked_fill(
action_log_probs, ~action_mask_mb, INVALID_LOGPROB
)
masked_engine_log_probs = torch.masked_fill(
engine_log_probs_mb, ~action_mask_mb, INVALID_LOGPROB
)
with torch.no_grad():
action_engine_log_probs_diff = (
masked_action_log_probs - masked_engine_log_probs
).clamp(-CLAMP_VALUE, CLAMP_VALUE)
running_mean_logs["engine_log_probs_diff_clampfrac"] += (
action_engine_log_probs_diff.abs()
.eq(CLAMP_VALUE)
.float()
.sum()
.item()
/ den_running_mean
)
if self.importance_sampling_strategy == "per_sequence":
tis_imp_ratio = torch.zeros_like(action_engine_log_probs_diff)
for mb_idx in range(action_engine_log_probs_diff.shape[0]):
valid_token_mask = action_mask_mb[mb_idx]
timestep_ids = timesteps_mb[mb_idx][valid_token_mask]
timestep_logprob_diffs = action_engine_log_probs_diff[mb_idx][
valid_token_mask
]
max_timestep = int(timestep_ids.max().item()) + 1
timestep_sums = torch.zeros(
max_timestep,
device=action_engine_log_probs_diff.device,
dtype=action_engine_log_probs_diff.dtype,
)
timestep_sums.scatter_add_(
0, timestep_ids, timestep_logprob_diffs
)
timestep_ratios = torch.exp(timestep_sums)
tis_imp_ratio[
mb_idx, valid_token_mask
] = timestep_ratios.gather(0, timestep_ids)
else:
tis_imp_ratio = torch.exp(action_engine_log_probs_diff)
running_mean_logs["tis_imp_ratio"] += (
tis_imp_ratio * action_mask_mb
).sum().item() / den_running_mean
if self.truncated_importance_sampling_ratio_cap > 0.0:
tis_imp_ratio = torch.clamp(
tis_imp_ratio, max=self.truncated_importance_sampling_ratio_cap
)
running_mean_logs["tis_imp_ratio_clampfrac"] += (
tis_imp_ratio.eq(self.truncated_importance_sampling_ratio_cap)
.float()
.sum()
.item()
) / den_running_mean
rewarded_action_log_probs = (
rewarded_action_log_probs * tis_imp_ratio
)
if self.enable_tokenwise_logging:
self.tokenwise_tally.add_data(
metric_id="next_token_clogπ",
metrics=rewarded_action_log_probs,
)
# Add value term to loss
if self.pg_loss_normalization == "batch":
nb_act_tokens = action_mask_mb.sum()
mb_value = -rewarded_action_log_probs.sum() / nb_act_tokens
else:
mb_value = -rewarded_action_log_probs.sum()
loss += mb_value
running_mean_logs["rl_objective"] += mb_value.item() / den_running_mean
# -------------------------------------------------
# Entropy Regularization
# -------------------------------------------------
# Only apply entropy on distribution defined over most probable tokens
if self.entropy_topk is not None:
top_k_indices = torch.topk(
logits, k=self.entropy_topk, dim=-1
).indices
entropy_logits = logits.gather(dim=-1, index=top_k_indices)
else:
entropy_logits = logits
token_entropy_terms = -F.softmax(
entropy_logits, dim=-1
) * F.log_softmax(
entropy_logits, dim=-1
) # (B, S, T)
token_entropy_terms *= (
action_mask_mb[:, :, None] * entropy_mask_mb[:, :, None]
) # only get loss on specific action tokens
mb_entropy = token_entropy_terms.sum(dim=-1)
if self.enable_tokenwise_logging:
self.tokenwise_tally.add_data(
metric_id="entropy",
metrics=mb_entropy,
)
if self.pg_loss_normalization == "batch":
nb_act_tokens = action_mask_mb.sum()
mb_entropy = -mb_entropy.sum() / nb_act_tokens
else:
mb_entropy = -mb_entropy.sum()
running_mean_logs["entropy"] += -mb_entropy.item() / den_running_mean
if self.entropy_coeff != 0.0:
mb_entropy *= self.entropy_coeff
loss += mb_entropy
# -------------------------------------------------
# KL-DIVERGENCE
# -------------------------------------------------
if self.kl_coeff != 0.0:
ref_model_logits = self.policy.get_base_model_logits(contexts_mb)
ref_model_logits = ref_model_logits / self.temperature
# (B, S, V)
ref_model_logits = self.mask_non_restricted_token_logits(
logits=ref_model_logits
)
# (B, S, V)
ref_model_log_probs = F.log_softmax(ref_model_logits, dim=-1)
# (B, S, V)
ref_model_action_log_probs = ref_model_log_probs.gather(
dim=-1, index=shifted_contexts_mb.unsqueeze(-1)
).squeeze(
-1
) # (B,S)
# Approximating KL Divergence (see refs in docstring)
# Ref 1: http://joschu.net/blog/kl-approx.html
# Ref 2: https://github.dev/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py#L1332
masked_ref_model_action_log_probs = torch.masked_fill(
ref_model_action_log_probs, ~action_mask_mb, INVALID_LOGPROB
)
action_log_probs_diff = (
masked_ref_model_action_log_probs - masked_action_log_probs
).clamp(-CLAMP_VALUE, CLAMP_VALUE)
running_mean_logs["ref_log_probs_diff_clampfrac"] += (
action_log_probs_diff.abs().eq(CLAMP_VALUE).float().sum().item()
/ den_running_mean
)
if self.filter_higher_refprob_tokens_kl:
higher_refprob_tokens_mask = action_log_probs_diff > 0.0
running_mean_logs["higher_refprob_frac"] += (
higher_refprob_tokens_mask.sum().item() / den_running_mean
)
action_log_probs_diff = action_log_probs_diff * (
~higher_refprob_tokens_mask
)
kl_div = torch.expm1(action_log_probs_diff) - action_log_probs_diff
kl_div *= action_mask_mb # We only care about KLD of action tokens
if self.truncated_importance_sampling_ratio_cap > 0.0:
kl_div = kl_div * tis_imp_ratio
kl_div *= self.kl_coeff
if self.enable_tokenwise_logging:
self.tokenwise_tally.add_data(
metric_id="ref_model_next_token_log_prob",
metrics=ref_model_action_log_probs,
)
self.tokenwise_tally.add_data(
metric_id="kl_divergence",
metrics=kl_div,
)
if self.pg_loss_normalization == "batch":
nb_act_tokens = action_mask_mb.sum()
mb_kl = kl_div.sum() / nb_act_tokens
else:
mb_kl = kl_div.sum()
running_mean_logs["kl_divergence"] += (
mb_kl.item() / den_running_mean
)
loss += mb_kl
# Accumulate gradient
running_mean_logs["policy_gradient_loss"] += (
loss.item() / den_running_mean
)
loss /= normalization_factor
self.accelerator.backward(loss)
# ensure gpu memory is freed
del training_mb
del log_probs
del logits
del loss
del action_log_probs
del rewarded_action_log_probs
logger.info(
f"Accumulated the policy gradient loss for {total_tokens_generated} tokens."
)
# Clip gradients and take step
if self.gradient_clipping is not None:
grad_norm = self.accelerator.clip_grad_norm_(
self.policy.parameters(), self.gradient_clipping
)
running_mean_logs["policy_gradient_norm"] += grad_norm.item()
# Take step
self.policy_optimizer.step()
self.policy_optimizer.zero_grad()
# Store logs
for key, value in running_mean_logs.items():
self.tally.add_metric(path=key, metric=value)
# Clear
# TODO: verify
self.accelerator.clear(self.policy, self.policy_optimizer)
import gc
gc.collect()
torch.cuda.empty_cache()
return running_mean_logs
def get_advantages_with_critic_gradient_accumulation(
self, trajectories: TrajectoryBatch, critic_loss_scaling_factor: float = 2.0
) -> torch.FloatTensor:
"""
TOWRITE
Uses GAE if enabled, otherwise uses Monte Carlo returns.
Optionally trains the critic if GAE is used.
Returns:
advantages: NestedFloatTensors
"""
mb_size = self.mini_batch_size
batch_size = trajectories.rollout_ids.shape[0]
agent_id = trajectories.agent_ids[0]
batch_rewards = trajectories.batch_rewards
######################################
# use critic for advantage estimation
######################################
if self.use_gae:
if "buffer" in agent_id:
self.critic.eval()
training = False
else:
self.critic.train()
training = True
advantages = []
# critic_loss_scaling_factor comes learning single critic for two agents
normalization_factor = (
np.ceil(batch_size / mb_size).astype(int) * critic_loss_scaling_factor
)
# For each minibatch
for mb in range(0, batch_size, mb_size):
trajectory_mb = trajectories[mb : mb + mb_size]
trajectory_mb.to(self.device)
rewards_mb = trajectory_mb.batch_rewards
(
tokens_mb,
state_ends_mask_mb,
timestep_counts,
) = trajectory_mb.get_padded_tensors_for_critic()
# critic causal attention up to end flags
if training:
vals_estimate_full = self.critic(tokens_mb)
else:
with torch.no_grad():
vals_estimate_full = self.critic(tokens_mb)
# if vals_estimate_full.dim() == 3:
# vals_estimate_full = vals_estimate_full.squeeze(-1)
# Select only positions where states end, per sample → list of (jT,)
B = tokens_mb.shape[0]
vals_list = [
vals_estimate_full[b][state_ends_mask_mb[b]] for b in range(B)
]
# Pad to (B, max_jT) = (B, S)
vals_estimate_mb = pad_sequence(
vals_list, batch_first=True, padding_value=0.0
)
dtype = vals_estimate_mb.dtype
rewards_mb = pad_sequence(
rewards_mb, batch_first=True, padding_value=0.0
).to(
dtype=dtype
) # (B, S)
self.rollout_tally.add_metric(
path=["batch_rewards"],
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectory_mb.crn_ids,
rollout_ids=trajectory_mb.rollout_ids,
agent_ids=trajectory_mb.agent_ids,
metric_matrix=rewards_mb,
),
)
if self.reward_normalizing_constant != 1.0:
rewards_mb /= self.reward_normalizing_constant
det_vals_estimate_mb = vals_estimate_mb.detach() # (B, max_jT)
self.rollout_tally.add_metric(
path=["mb_value_estimates_critic"],
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectory_mb.crn_ids,
rollout_ids=trajectory_mb.rollout_ids,
agent_ids=trajectory_mb.agent_ids,
metric_matrix=det_vals_estimate_mb,
),
)
# Append a 0 value to the end of the value estimates
if det_vals_estimate_mb.shape[1] == rewards_mb.shape[1]:
Bsize = det_vals_estimate_mb.shape[0]
device = det_vals_estimate_mb.device
dtype = det_vals_estimate_mb.dtype
det_vals_estimate_mb = torch.cat(
[
det_vals_estimate_mb,
torch.zeros((Bsize, 1), device=device, dtype=dtype),
],
dim=1,
) # (B, max_jT+1)
else:
raise ValueError(
"Incompatible shapes for value estimates and rewards."
)
# Get annealed lambda
if self.use_gae_lambda_annealing:
annealing_constant = self.gae_lambda_annealing_method(
step=self.trainer_annealing_state.annealing_step_counter
)
annealed_lambda = (
self.gae_lambda_annealing_limit * annealing_constant
)
self.tally.add_metric(
path="annealed_lambda", metric=annealed_lambda
)
else:
annealed_lambda = self.gae_lambda_annealing_limit
# Get GAE advantages
gae_advantages = get_generalized_advantage_estimates(
rewards=rewards_mb,
value_estimates=det_vals_estimate_mb,
discount_factor=self.discount_factor,
lambda_coef=annealed_lambda,
) # (B, max_jT)
self.rollout_tally.add_metric(
path=["mb_gae_advantages"],
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectory_mb.crn_ids,
rollout_ids=trajectory_mb.rollout_ids,
agent_ids=trajectory_mb.agent_ids,
metric_matrix=gae_advantages,
),
)
if training:
targets = (
gae_advantages.to(dtype=dtype) + det_vals_estimate_mb[:, :-1]
) # (B, max_jT) # A(s, a, b) + V(s) = Q(s, a, b)
self.rollout_tally.add_metric(
path=["mb_targets_critic"],
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectory_mb.crn_ids,
rollout_ids=trajectory_mb.rollout_ids,
agent_ids=trajectory_mb.agent_ids,
metric_matrix=targets,
),
)
if self.critic_loss_type == "mse":
loss = F.mse_loss(
input=vals_estimate_mb,
target=targets,
)
elif self.critic_loss_type == "huber":
loss = F.huber_loss(
input=vals_estimate_mb,
target=targets,
)
self.tally.add_metric(path=["mb_critic_loss"], metric=loss.item())
# Accumulate gradient
loss /= normalization_factor
self.accelerator.backward(loss)
del loss
del targets
del vals_estimate_mb
del trajectory_mb
del vals_estimate_full
# Get jagged back using timestep_counts
advantages.extend(
[gae_advantages[i, : timestep_counts[i]] for i in range(B)]
)
######################################
# use exclusively Monte Carlo returns & rloo for advantage estimation
######################################
else:
lengths = [len(c) for c in batch_rewards]
padded_rewards = pad_sequence(
batch_rewards, batch_first=True, padding_value=0.0
)
self.rollout_tally.add_metric(
path=["mb_rewards"],
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectories.crn_ids,
rollout_ids=trajectories.rollout_ids,
agent_ids=trajectories.agent_ids,
metric_matrix=padded_rewards,
),
)
if self.reward_normalizing_constant != 1.0:
padded_rewards /= self.reward_normalizing_constant
padded_advantages = get_discounted_returns(
rewards=padded_rewards,
discount_factor=self.discount_factor,
) # no baseline for now
if self.use_rloo:
is_grouped_by_rng = (
trajectories.crn_ids.unique().shape[0]
!= trajectories.crn_ids.shape[0]
)
if is_grouped_by_rng:
for crn_id in trajectories.crn_ids.unique():
rng_mask = trajectories.crn_ids == crn_id
rng_advantages = padded_advantages[rng_mask]
rng_advantages, _ = get_rloo_credits(credits=rng_advantages)
padded_advantages[rng_mask] = rng_advantages
else:
padded_advantages, _ = get_rloo_credits(credits=padded_advantages)
self.rollout_tally.add_metric(
path=["mb_rloo_advantages"],
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectories.crn_ids,
rollout_ids=trajectories.rollout_ids,
agent_ids=trajectories.agent_ids,
metric_matrix=padded_advantages,
),
)
advantages = [
padded_advantages[i, : lengths[i]]
for i in range(padded_advantages.shape[0])
]
if self.whiten_advantages_time_step_wise or self.whiten_advantages:
lengths = [len(c) for c in advantages]
padded_advantages = pad_sequence(
advantages, batch_first=True, padding_value=0.0
)
if self.whiten_advantages_time_step_wise:
whitened_padded_advantages = whiten_advantages_time_step_wise(
padded_advantages
)
path = ["mb_whitened_advantages_time_step_wise"]
elif self.whiten_advantages:
whitened_padded_advantages = whiten_advantages(padded_advantages)
path = ["mb_whitened_advantages"]
self.rollout_tally.add_metric(
path=path,
rollout_tally_item=RolloutTallyItem(
crn_ids=trajectories.crn_ids,
rollout_ids=trajectories.rollout_ids,
agent_ids=trajectories.agent_ids,
metric_matrix=whitened_padded_advantages,
),
)
advantages = [
whitened_padded_advantages[i, : lengths[i]]
for i in range(whitened_padded_advantages.shape[0])
]
self.trainer_annealing_state.annealing_step_counter += 1
return advantages
@abstractmethod
def set_agent_trajectory_data(
self, agent_id: str, roots: list[RolloutTreeRootNode]
) -> None:
"""
TOWRITE
"""
pass
def set_trajectory_data(
self, roots: list[RolloutTreeRootNode], agent_ids: list[str]
) -> None:
"""
TOWRITE
"""
for agent_id in agent_ids:
self.set_agent_trajectory_data(agent_id, roots)
@abstractmethod
def share_advantage_data(self) -> list[AdvantagePacket]:
pass
@abstractmethod
def receive_advantage_data(self, advantage_packets: list[AdvantagePacket]) -> None:
pass
def set_policy_gradient_data(self, agent_ids: list[str]) -> None:
"""
Already set earlier # TODO: make it separate and clean
"""
self.policy_gradient_data = None
# for agent_id, trajectory_batch in self.training_data.items():
# if "buffer" in agent_id:
# continue
for agent_id in agent_ids:
assert "buffer" not in agent_id, "Buffer agents do not train policy"
trajectory_batch = self.training_data[agent_id]
tokenwise_batch_credits = get_tokenwise_credits(
batch_timesteps=trajectory_batch.batch_timesteps,
batch_credits=trajectory_batch.batch_credits,
)
policy_gradient_data = TrainingBatch(
rollout_ids=trajectory_batch.rollout_ids,
batch_input_ids=trajectory_batch.batch_input_ids,
batch_action_mask=trajectory_batch.batch_action_mask,
batch_entropy_mask=trajectory_batch.batch_entropy_mask,
batch_credits=tokenwise_batch_credits,
batch_engine_log_probs=trajectory_batch.batch_engine_log_probs,
batch_timesteps=trajectory_batch.batch_timesteps,
)
if self.policy_gradient_data is None:
self.policy_gradient_data = policy_gradient_data
else:
self.policy_gradient_data.append(policy_gradient_data)
self.training_data = {}
self.tokenwise_tally = ContextualizedTokenwiseTally(
tokenizer=self.tokenizer,
paths=self.debug_path_list,
)
def train(self) -> None:
"""
TOWRITE
"""
assert self.policy_gradient_data is not None, "Policy gradient data is not set"
if self.critic_optimizer is not None:
if self.gradient_clipping is not None:
grad_norm = self.accelerator.clip_grad_norm_(
self.critic.parameters(), self.gradient_clipping
)
self.tally.add_metric(
path="gradient_norm_critic", metric=grad_norm.item()
)
# Take step
self.critic_optimizer.step()
self.critic_optimizer.zero_grad()
self.accelerator.clear(self.critic, self.critic_optimizer)
import gc
gc.collect()
torch.cuda.empty_cache()
running_mean_logs = self.apply_reinforce_step(
training_batch=self.policy_gradient_data
)
return running_mean_logs
def export_training_tally(self, identifier: str, folder: str) -> None:
"""
Saves and resets the collected training metrics using the tally object.
"""
os.makedirs(folder, exist_ok=True)
self.tally.save(identifier=identifier, folder=folder)
self.tokenwise_tally.save(
path=os.path.join(folder, f"{identifier}_tokenwise.csv")
)
self.rollout_tally.save(identifier=identifier, folder=folder)
self.tally.reset()
self.tokenwise_tally = None
self.rollout_tally.reset()
self.debug_path_list = []
def export_optimizer_states(self) -> None:
"""
Saves the optimizer states for both the main model and critic (if it exists).
"""
try:
os.makedirs(self.save_path, exist_ok=True)
torch.save(self.policy_optimizer.state_dict(), self.policy_optimizer_path)
logger.info(f"Saved main optimizer state to {self.policy_optimizer_path}")
if self.critic_optimizer is not None:
torch.save(
self.critic_optimizer.state_dict(), self.critic_optimizer_path
)
logger.info(
f"Saved critic optimizer state to {self.critic_optimizer_path}"
)
except Exception as e:
logger.error(f"Error saving optimizer states: {str(e)}")
raise
def export_trainer_annealing_state(self) -> None:
"""
Saves the trainer state.
"""
with open(self.trainer_annealing_state_path, "wb") as f:
pickle.dump(self.trainer_annealing_state, f)
logger.info(f"Saved trainer state to {self.trainer_annealing_state_path}")
def export_trainer_states(self) -> None:
"""
Saves the trainer states.
"""
self.export_optimizer_states()
self.export_trainer_annealing_state()
|