# Copyright 2025 The EasyDeL Author @erfanzar (Erfan Zare Chavoshi).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import typing as tp
from dataclasses import field
from eformer.pytree import auto_pytree
from easydel.utils import Registry
from easydel.utils.compiling_utils import hash_fn
from ..training_configurations import TrainingArguments
[docs]@Registry.register("trainer-arguments", "grpo")
@auto_pytree
class GRPOConfig(TrainingArguments):
"""Configuration class for Group Relative Policy Optimization training.
GRPO is an efficient RLHF algorithm that optimizes policies using group-based
relative comparisons of rewards. It provides better training stability compared
to standard PPO by normalizing rewards within groups of samples.
This configuration extends TrainingArguments with GRPO-specific parameters
for controlling the policy optimization process, reward computation, and
generation sampling strategies.
Key concepts:
- Group-based normalization: Rewards are normalized within groups to reduce variance
- KL regularization: Prevents the policy from deviating too far from reference
- Reference model syncing: Optionally updates reference model during training
"""
trainer_prefix: str | None = field(
default="grpotrainer",
metadata={"help": "default prefix name for trainer."},
)
remove_unused_columns: bool | None = field(
default=False,
metadata={"help": "Whether to remove unused columns from the dataset."},
)
max_prompt_length: int = field(
default=512,
metadata={"help": "The maximum length of the prompt."},
)
max_completion_length: int = field(
default=256,
metadata={"help": "The maximum length of the completion."},
)
dataset_num_proc: int | None = field(
default=None,
metadata={"help": "The number of processes to use for dataset processing."},
)
learning_rate: float = field(
default=1e-6,
metadata={"help": "The learning rate."},
)
beta: float = field(
default=0.04,
metadata={"help": "The beta parameter for GRPO."},
)
epsilon: float = field(
default=0.2,
metadata={"help": "Lower clipping bound for importance sampling weights."},
)
epsilon_high: float | None = field(
default=None,
metadata={"help": "Upper clipping bound for importance sampling weights. If None, defaults to `epsilon`."},
)
delta: float | None = field(
default=None,
metadata={
"help": "Optional two-sided clipping bound. If set, importance weights are additionally clipped to `delta`."
},
)
sync_ref_model: bool = field(
default=False,
metadata={"help": "Whether to periodically sync the reference model with the policy model."},
)
ref_model_mixup_alpha: float = field(
default=0.9,
metadata={"help": "The alpha parameter for mixing the reference model with the policy model."},
)
ref_model_sync_steps: int = field(
default=64,
metadata={"help": "The number of steps between syncing the reference model."},
)
num_iterations: int = field(
default=1,
metadata={"help": "How many optimizer updates to perform per generated batch."},
)
loss_type: str = field(
default="dapo",
metadata={"help": "Loss variant to use. One of ['grpo', 'bnpo', 'dr_grpo', 'dapo', 'cispo']."},
)
importance_sampling_level: str = field(
default="token",
metadata={"help": "Importance sampling applied per 'token' or aggregated per 'sequence'."},
)
reward_weights: list[float] | None = field(
default=None,
metadata={
"help": "Optional weights for each reward function. Must match the number of reward functions if set."
},
)
scale_rewards: str | bool = field(
default="group",
metadata={
"help": "Reward scaling strategy: 'group', 'batch', 'none', or the booleans True/False for group/none."
},
)
tools: list[dict | tp.Callable] | None = field(
default=None,
metadata={"help": "Additional tools for training."},
)
skip_apply_chat_template: bool = field(
default=False,
metadata={"help": "whenever to skip extracting prompt from dataset."},
)
num_return_sequences: int = field(
default=4,
metadata={
"help": "The number of sequences to return for each input prompt. Used during sampling to "
"generate multiple completions per prompt."
},
)
num_generations: int | None = field(
default=None,
metadata={"help": "Alias for num_return_sequences to keep parity with TRL's interface."},
)
temperature: float = field(
default=1.0,
metadata={"help": "Sampling temperature used during generation."},
)
top_p: float = field(
default=1.0,
metadata={"help": "Top-p nucleus sampling parameter."},
)
top_k: int | None = field(
default=None,
metadata={"help": "Top-k sampling parameter. None disables top-k."},
)
min_p: float | None = field(
default=None,
metadata={"help": "Minimum token probability threshold (see HF top-p-min sampling)."},
)
repetition_penalty: float = field(
default=1.0,
metadata={"help": "Repetition penalty applied during generation."},
)
generation_kwargs: dict | None = field(
default=None,
metadata={"help": "Additional generation kwargs forwarded to the generation config."},
)
chat_template_kwargs: dict | None = field(
default=None,
metadata={"help": "Extra kwargs forwarded to chat template application during generation."},
)
mask_truncated_completions: bool = field(
default=False,
metadata={"help": "If True, drop completions that do not terminate with EOS from the loss calculation."},
)
top_entropy_quantile: float = field(
default=1.0,
metadata={"help": "Keep only the top quantile of tokens by entropy in the loss (1.0 disables filtering)."},
)
def __post_init__(self):
"""Post initialization to set dependent parameters."""
self.max_sequence_length = self.max_prompt_length + self.max_completion_length
if self.num_generations is None:
self.num_generations = self.num_return_sequences
else:
self.num_return_sequences = self.num_generations
if self.epsilon_high is None:
self.epsilon_high = self.epsilon
if self.scale_rewards is True:
self.scale_rewards = "group"
elif self.scale_rewards is False:
self.scale_rewards = "none"
if hasattr(super(), "__post_init__"):
super().__post_init__()
__hash__ = hash_fn