# Copyright (c) OpenMMLab. All rights reserved.
from typing import Any, cast
import torch
import torch.distributed as dist
import torch.nn.functional as F
from xtuner.v1.utils import get_logger
from ..utils import gather_logprobs
from .base_loss import (
BaseRLLossConfig,
BaseRLLossContext,
BaseRLLossKwargs,
compute_kl_loss_weight,
)
from .loss_fn import get_policy_loss_fn, kl_penalty
logger = get_logger()
[文档]class GRPOLossConfig(BaseRLLossConfig):
"""Configuration for GRPO loss computation in XTuner RL.
``GRPOLossConfig`` implements the loss configuration used by Group Relative
Policy Optimization. It consumes advantages computed for each rollout group
and supports optional KL regularization through ``BaseRLLossConfig``.
Args:
policy_loss_cfg (dict[str, Any]): Configuration parameters for the main policy loss.
Contains algorithm-specific parameters for policy optimization.
use_kl_loss (bool): Whether to include KL divergence penalty in the loss.
When True, requires a reference model for KL computation. Defaults to False.
kl_loss_coef (float): Coefficient for weighting the KL divergence penalty.
Controls the strength of regularization against the reference policy. Defaults to 0.001.
kl_loss_type (Literal["kl", "k1", "abs", "mse", "k2", "low_var_kl", "k3"] | None):
Type of KL penalty computation method. Different types provide various
regularization behaviors and numerical stability properties. Defaults to None.
rollout_is (RolloutImportanceSampling): Rollout importance sampling
configuration. Defaults to ``RolloutImportanceSampling()``.
**Examples:**
Example GRPO loss configuration::
config = GRPOLossConfig(
policy_loss_cfg={
"loss_type": "vanilla",
"cliprange_low": 0.2,
"cliprange_high": 0.2,
},
use_kl_loss=True,
kl_loss_coef=0.001,
kl_loss_type="low_var_kl",
)
"""
@property
def loss_ctx_cls(self) -> type["GRPOLossContext"]:
return GRPOLossContext
@property
def _loss_kwargs_cls(self) -> type["GRPOLossKwargs"]:
return GRPOLossKwargs
[文档]class GRPOLossKwargs(BaseRLLossKwargs):
"""Keyword arguments for GRPO loss computation.
Args:
shifted_labels (torch.Tensor): The shifted labels for the input sequences.
old_logprobs (torch.Tensor): Log probabilities from the old policy.
advantages (torch.Tensor): Advantage estimates for the actions taken.
policy_loss_weight (torch.Tensor): Weights for each token in the policy loss computation.
ref_logprobs (torch.Tensor | None): Reference log probabilities for KL penalty, if used.
kl_loss_weight (torch.Tensor | None): Weights for each token in the KL loss computation, if used.
"""
pass
[文档]class GRPOLossContext(BaseRLLossContext):
"""GRPO loss context for reinforcement learning.
Args:
loss_cfg (GRPOLossConfig): Configuration for GRPO loss computation.
loss_kwargs (GRPOLossKwargs): Keyword arguments required for loss calculation.
"""
loss_cfg: GRPOLossConfig
loss_kwargs: GRPOLossKwargs
def __init__(self, loss_cfg: GRPOLossConfig, loss_kwargs: GRPOLossKwargs):
super().__init__(loss_cfg, loss_kwargs)
self.policy_loss_fn = get_policy_loss_fn(self.loss_cfg.policy_loss_cfg.get("loss_type", "vanilla"))
@staticmethod
def build_batches(loss_ctx_list: list["GRPOLossContext"]) -> list["GRPOLossContext"]: # type: ignore[override]
assert len(loss_ctx_list) > 0, "loss_ctx_list can not be empty"
loss_cfg = loss_ctx_list[0].loss_cfg
shifted_labels_list = [loss_ctx.loss_kwargs.shifted_labels for loss_ctx in loss_ctx_list]
# Compute the denominator used in the global calibration of the loss
rank_grad_tokens = sum((labels != loss_cfg.ignore_idx).sum() for labels in shifted_labels_list)
rank_grad_tokens = cast(torch.Tensor, rank_grad_tokens)
global_grad_tokens = rank_grad_tokens
if dist.is_initialized():
dist.all_reduce(global_grad_tokens, op=dist.ReduceOp.SUM)
if global_grad_tokens == 0:
logger.warning(
"Global gradient tokens is 0, which may lead to division by zero in loss weight calculation."
)
global_grad_tokens.add_(1) # Avoid division by zero
for loss_ctx in loss_ctx_list:
loss_kwargs = loss_ctx.loss_kwargs
shifted_labels = loss_kwargs.shifted_labels
assert loss_kwargs.old_logprobs is not None, "old_logprobs can not be None"
# compute loss weight
policy_loss_weight = torch.ones_like(shifted_labels, dtype=torch.float32) / global_grad_tokens
policy_loss_weight[shifted_labels == loss_cfg.ignore_idx] = 0.0
if loss_kwargs.is_weights is not None:
policy_loss_weight = policy_loss_weight * loss_kwargs.is_weights
if loss_cfg.use_kl_loss:
assert loss_kwargs.ref_logprobs is not None, "ref_logprobs can not be None"
kl_loss_weight = compute_kl_loss_weight(
shifted_labels, global_grad_tokens, loss_cfg.kl_loss_coef, loss_cfg.ignore_idx
)
else:
kl_loss_weight = None
loss_kwargs.policy_loss_weight = policy_loss_weight
loss_kwargs.kl_loss_weight = kl_loss_weight
loss_kwargs.global_grad_tokens = global_grad_tokens
return loss_ctx_list
def loss_fn(
self,
hidden_states: torch.Tensor,
head_weight: torch.Tensor,
head_bias: torch.Tensor | None,
loss_kwargs: GRPOLossKwargs,
) -> tuple[torch.Tensor, tuple[torch.Tensor | None, dict[str, Any]]]:
"""Step 2.a and 2.b in the loss calculation in
xtuner/v1/loss/base_loss_ctx.py."""
# We do linear forward here to simplify the implementation of chunk loss (saving memory).
logits = F.linear(hidden_states, head_weight, head_bias)
logits = logits.float()
shifted_labels = loss_kwargs.shifted_labels
old_logprobs = loss_kwargs.old_logprobs
advantages = loss_kwargs.advantages
policy_loss_weight = loss_kwargs.policy_loss_weight
logprobs = gather_logprobs(logits, shifted_labels)
loss = self.policy_loss_fn(
logprobs,
old_logprobs,
advantages,
policy_loss_weight,
self.loss_cfg.policy_loss_cfg,
)
assert old_logprobs is not None
valid_mask = shifted_labels != self.loss_cfg.ignore_idx
valid_float = valid_mask.float()
cliprange_low = self.loss_cfg.policy_loss_cfg.get("cliprange_low")
cliprange_high = self.loss_cfg.policy_loss_cfg.get("cliprange_high")
log_ratio = logprobs.detach() - old_logprobs.detach()
log_ratio_safe = torch.clamp(log_ratio, min=-20.0, max=20.0)
ratio = torch.exp(log_ratio_safe)
kl1 = -log_ratio
kl3 = ratio - 1.0 - log_ratio_safe
ratio_abs_dev = (ratio - 1.0).abs()
ratio_max = ratio.masked_fill(~valid_mask, 0.0).max()
ratio_min = ratio.masked_fill(~valid_mask, float("inf")).min()
extra_info = {
"max_ratio": ratio_max,
"reduced_train_policy_ratio_abs_dev_sum": (ratio_abs_dev * valid_float).sum(),
"reduced_train_policy_kl1_sum": (kl1 * valid_float).sum(),
"reduced_train_policy_kl3_sum": (kl3 * valid_float).sum(),
"reduced_train_policy_valid_count": valid_float.sum(),
"reduced_train_policy_ratio_max": ratio_max,
"reduced_train_policy_ratio_min": ratio_min,
}
if cliprange_low is not None and cliprange_high is not None:
clip_low_mask = ratio < 1 - cliprange_low
clip_high_mask = ratio > 1 + cliprange_high
extra_info["reduced_train_policy_clip_low_count"] = (clip_low_mask & valid_mask).float().sum()
extra_info["reduced_train_policy_clip_high_count"] = (clip_high_mask & valid_mask).float().sum()
if self.loss_cfg.use_kl_loss:
ref_logprobs = loss_kwargs.ref_logprobs
kl_loss_weight = loss_kwargs.kl_loss_weight
assert ref_logprobs is not None and kl_loss_weight is not None, (
"loss_kwargs.ref_logprobs and loss_kwargs.kl_loss_weight can not be None when use_kl_loss=True"
)
kl_loss = kl_penalty(logprobs, ref_logprobs, kl_loss_weight, self.loss_cfg.kl_loss_type)
loss = loss + kl_loss
return loss, (logits, extra_info)
