File size: 8,671 Bytes
8c9ba62 | 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 | """KL penalty and loss.
Ref:
https://github.com/volcengine/verl/blob/main/verl/trainer/ppo/core_algos.py
https://github.com/volcengine/verl/blob/main/verl/trainer/ppo/ray_trainer.py
https://github.com/OpenRLHF/OpenRLHF/blob/main/openrlhf/models/utils.py
"""
from abc import ABC, abstractmethod
from typing import Any, Dict, Optional, Tuple
import torch
from trinity.algorithm.utils import aggregate_loss, masked_mean
class KLFn(ABC):
"""
KL penalty and loss.
"""
def __init__(
self,
adaptive: bool = False,
kl_coef: float = 0.001,
target_kl: Optional[float] = None,
horizon: Optional[float] = None,
) -> None:
self.kl_coef = kl_coef
self.adaptive = adaptive
self.target_kl = target_kl
self.horizon = horizon
if adaptive and (target_kl is None or horizon is None):
raise ValueError("Target KL and horizon must be provided for adaptive KL.")
def update_kl_coef(self, current_kl: float, batch_size: int) -> None:
"""Update kl coefficient."""
if self.adaptive:
target_kl = self.target_kl
proportional_error = torch.clip(current_kl / target_kl - 1, -0.2, 0.2).item() # type: ignore
multiplier = 1 + proportional_error * batch_size / self.horizon
self.kl_coef *= multiplier
def apply_kl_penalty_to_reward(self, experiences: Any) -> Tuple[Any, Dict]:
"""Apply KL penalty to reward. Only support DataProto input for now."""
responses = experiences.batch["responses"]
response_length = responses.size(1)
token_level_scores = experiences.batch["token_level_scores"]
batch_size = experiences.batch.batch_size[0]
attention_mask = experiences.batch["attention_mask"]
response_mask = experiences.batch["response_mask"]
assert response_mask.shape == attention_mask[:, -response_length:].shape
logprob = experiences.batch["old_log_probs"]
ref_logprob = experiences.batch["ref_log_prob"]
if "ref_log_prob" in experiences.batch.keys():
kl = self.calculate_kl(logprob, ref_logprob)
kl = kl * response_mask
kl_coef = self.kl_coef
experiences.batch["token_level_rewards"] = token_level_scores - kl_coef * kl
else:
kl_coef = 0.0
kl = torch.zeros_like(response_mask, dtype=torch.float32)
experiences.batch["token_level_rewards"] = token_level_scores
current_kl = masked_mean(kl, mask=response_mask, axis=-1).mean(dim=0).item()
self.update_kl_coef(current_kl=current_kl, batch_size=batch_size)
metrics = {
"kl": current_kl,
"kl_coef": kl_coef,
}
return experiences, metrics
def calculate_kl_loss(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
response_mask: torch.Tensor,
loss_agg_mode: str,
old_logprob: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Dict]:
"""Compute KL loss.
Args:
logprob: Log probabilities from current policy
ref_logprob: Log probabilities from reference policy
response_mask: Mask for valid response tokens
loss_agg_mode: Loss aggregation mode
old_logprob: Log probabilities from old policy (for importance sampling)
"""
kl = self.calculate_kl(logprob, ref_logprob, old_logprob)
kl_loss = aggregate_loss(kl, response_mask, loss_agg_mode=loss_agg_mode)
metrics = {
"kl_loss": kl_loss.detach().item(),
"kl_coef": self.kl_coef,
}
return kl_loss * self.kl_coef, metrics
@abstractmethod
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Compute KL divergence between logprob and ref_logprob.
Args:
logprob: Log probabilities from current policy
ref_logprob: Log probabilities from reference policy
old_logprob: Log probabilities from old policy (for importance sampling)
"""
@classmethod
def default_args(cls):
"""Get the default initialization arguments."""
return {"adaptive": False, "kl_coef": 0.001}
class DummyKLFn(KLFn):
"""
Dummy KL function.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
return torch.zeros_like(logprob)
def apply_kl_penalty_to_reward(self, experiences: Any) -> Tuple[Any, Dict]:
experiences.batch["token_level_rewards"] = experiences.batch["token_level_scores"]
return experiences, {}
def calculate_kl_loss(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
response_mask: torch.Tensor,
loss_agg_mode: str,
old_logprob: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Dict]:
# return a zero tensor
return torch.tensor(0.0), {}
class K1Fn(KLFn):
"""
KL K1 function.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
return logprob - ref_logprob
class K2Fn(KLFn):
"""
KL K2 function.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
return (logprob - ref_logprob).square() * 0.5
class K3Fn(KLFn):
"""
KL K3 function.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
logr = ref_logprob - logprob
return logr.exp() - 1 - logr
class LowVarKLFn(KLFn):
"""
Low Variance KL function.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
kl = ref_logprob - logprob
kl = torch.clamp(kl, min=-20, max=20)
ratio = torch.exp(kl)
kld = (ratio - kl - 1).contiguous()
return torch.clamp(kld, min=-10, max=10)
class AbsFn(KLFn):
"""
KL Abs function.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
return torch.abs(logprob - ref_logprob)
class CorrectedK3Fn(KLFn):
"""
Corrected K3 function with importance sampling.
This method applies importance sampling correction to the standard K3 KL divergence.
The corrected KL is computed as:
KL_corrected = (π_θ / π_old) * KL_standard(π_ref || π_θ)
where:
- π_θ: current policy
- π_old: old policy (from rollout)
- π_ref: reference policy
- KL_standard: exp(log(π_ref/π_θ)) - log(π_ref/π_θ) - 1
If old_logprob is not provided, it falls back to standard K3.
"""
def calculate_kl(
self,
logprob: torch.Tensor,
ref_logprob: torch.Tensor,
old_logprob: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Compute corrected K3 KL divergence with importance sampling.
Args:
logprob: Log probabilities from current policy (log π_θ)
ref_logprob: Log probabilities from reference policy (log π_ref)
old_logprob: Log probabilities from old policy (log π_old), optional
Returns:
KL divergence tensor with same shape as input
"""
# Standard K3 KL term: exp(log_ratio) - log_ratio - 1
# where log_ratio = log(π_ref / π_θ) = ref_logprob - logprob
logr = ref_logprob - logprob
kl_term = logr.exp() - 1 - logr
if old_logprob is None:
# Fall back to standard K3 if old_logprob is not provided
return kl_term
# Compute importance sampling ratio: π_θ / π_old
log_ratio_is = logprob - old_logprob
ratio_is = log_ratio_is.exp()
# Clamp ratio for numerical stability, range [0, 2]
ratio_is = torch.clamp(ratio_is, min=0.0, max=2.0)
# Corrected KL with importance sampling
corrected_kl = ratio_is * kl_term
return corrected_kl
|