mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-06-26 07:22:12 +00:00
7b9b86441f
* feat: modify forward fn of critic and reward model * feat: modify calc_action_log_probs * to: add wandb in sft and rm trainer * feat: update train_sft * feat: update train_rm * style: modify type annotation and add warning * feat: pass tokenizer to ppo trainer * to: modify trainer base and maker base * feat: add wandb in ppo trainer * feat: pass tokenizer to generate * test: update generate fn tests * test: update train tests * fix: remove action_mask * feat: remove unused code * fix: fix wrong ignore_index * fix: fix mock tokenizer * chore: update requirements * revert: modify make_experience * fix: fix inference * fix: add padding side * style: modify _on_learn_batch_end * test: use mock tokenizer * fix: use bf16 to avoid overflow * fix: fix workflow * [chat] fix gemini strategy * [chat] fix * sync: update colossalai strategy * fix: fix args and model dtype * fix: fix checkpoint test * fix: fix requirements * fix: fix missing import and wrong arg * fix: temporarily skip gemini test in stage 3 * style: apply pre-commit * fix: temporarily skip gemini test in stage 1&2 --------- Co-authored-by: Mingyan Jiang <1829166702@qq.com>
98 lines
2.7 KiB
Python
98 lines
2.7 KiB
Python
from typing import Optional
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from .utils import masked_mean
|
|
|
|
|
|
class GPTLMLoss(nn.Module):
|
|
"""
|
|
GPT Language Model Loss
|
|
"""
|
|
|
|
def __init__(self):
|
|
super().__init__()
|
|
# NOTE: default ignore_index is -100, which is equal to IGNORE_INDEX in sft_dataset.py
|
|
self.loss = nn.CrossEntropyLoss()
|
|
|
|
def forward(self, logits: torch.Tensor, labels: torch.Tensor) -> torch.Tensor:
|
|
shift_logits = logits[..., :-1, :].contiguous()
|
|
shift_labels = labels[..., 1:].contiguous()
|
|
# Flatten the tokens
|
|
return self.loss(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
|
|
|
|
|
class PolicyLoss(nn.Module):
|
|
"""
|
|
Policy Loss for PPO
|
|
"""
|
|
|
|
def __init__(self, clip_eps: float = 0.2) -> None:
|
|
super().__init__()
|
|
self.clip_eps = clip_eps
|
|
|
|
def forward(
|
|
self,
|
|
log_probs: torch.Tensor,
|
|
old_log_probs: torch.Tensor,
|
|
advantages: torch.Tensor,
|
|
action_mask: Optional[torch.Tensor] = None,
|
|
) -> torch.Tensor:
|
|
ratio = (log_probs - old_log_probs).exp()
|
|
surr1 = ratio * advantages
|
|
surr2 = ratio.clamp(1 - self.clip_eps, 1 + self.clip_eps) * advantages
|
|
loss = -torch.min(surr1, surr2)
|
|
if action_mask is not None:
|
|
loss = masked_mean(loss, action_mask)
|
|
loss = loss.mean()
|
|
return loss
|
|
|
|
|
|
class ValueLoss(nn.Module):
|
|
"""
|
|
Value Loss for PPO
|
|
"""
|
|
|
|
def __init__(self, clip_eps: float = 0.4) -> None:
|
|
super().__init__()
|
|
self.clip_eps = clip_eps
|
|
|
|
def forward(
|
|
self,
|
|
values: torch.Tensor,
|
|
old_values: torch.Tensor,
|
|
reward: torch.Tensor,
|
|
action_mask: Optional[torch.Tensor] = None,
|
|
) -> torch.Tensor:
|
|
values_clipped = old_values + (values - old_values).clamp(-self.clip_eps, self.clip_eps)
|
|
surr1 = (values_clipped - reward) ** 2
|
|
surr2 = (values - reward) ** 2
|
|
loss = torch.max(surr1, surr2)
|
|
loss = loss.mean()
|
|
return 0.5 * loss
|
|
|
|
|
|
class LogSigLoss(nn.Module):
|
|
"""
|
|
Pairwise Loss for Reward Model
|
|
Details: https://arxiv.org/abs/2203.02155
|
|
"""
|
|
|
|
def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor:
|
|
probs = torch.sigmoid(chosen_reward - reject_reward)
|
|
log_probs = torch.log(probs)
|
|
loss = -log_probs.mean()
|
|
return loss
|
|
|
|
|
|
class LogExpLoss(nn.Module):
|
|
"""
|
|
Pairwise Loss for Reward Model
|
|
Details: https://arxiv.org/abs/2204.05862
|
|
"""
|
|
|
|
def forward(self, chosen_reward: torch.Tensor, reject_reward: torch.Tensor) -> torch.Tensor:
|
|
loss = torch.log(1 + torch.exp(reject_reward - chosen_reward)).mean()
|
|
return loss
|