mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-06-26 15:32:22 +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>
47 lines
1.6 KiB
Python
47 lines
1.6 KiB
Python
from typing import Optional
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from ..lora import LoRAModule
|
|
|
|
|
|
class RewardModel(LoRAModule):
|
|
"""
|
|
Reward model base class.
|
|
|
|
Args:
|
|
model (nn.Module): Reward model.
|
|
value_head (nn.Module): Value head to get reward score.
|
|
lora_rank (int): LoRA rank.
|
|
lora_train_bias (str): LoRA bias training mode.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
model: nn.Module,
|
|
value_head: Optional[nn.Module] = None,
|
|
lora_rank: int = 0,
|
|
lora_train_bias: str = "none",
|
|
) -> None:
|
|
super().__init__(lora_rank=lora_rank, lora_train_bias=lora_train_bias)
|
|
self.model = model
|
|
self.convert_to_lora()
|
|
|
|
if value_head is not None:
|
|
if value_head.out_features != 1:
|
|
raise ValueError("The value head of reward model's output dim should be 1!")
|
|
self.value_head = value_head
|
|
else:
|
|
self.value_head = nn.Linear(model.config.n_embd, 1)
|
|
|
|
def forward(self, sequences: torch.LongTensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
|
outputs = self.model(sequences, attention_mask=attention_mask)
|
|
last_hidden_states = outputs["last_hidden_state"]
|
|
sequence_lengths = torch.max(attention_mask * torch.arange(sequences.size(1), device=sequences.device), dim=1)[
|
|
0
|
|
]
|
|
sequence_hidden_states = last_hidden_states[torch.arange(last_hidden_states.size(0)), sequence_lengths]
|
|
values = self.value_head(sequence_hidden_states).squeeze(1) # ensure shape is (B, )
|
|
return values
|