mirror of
https://github.com/hpcaitech/ColossalAI.git
synced 2025-07-18 17:31:53 +00:00
d20c8ffd97
* add grpo, support rlvr * add grpo, support rlvr * tested deepseek r1 pipeline * add ci * verify grpo r1 * verify grpo r1 * update readme, remove unused code * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * remove path * clean code * fix circular import * fix ci OOM * fix ci OOM * skip kto tp, fix qwen generation --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
51 lines
1.3 KiB
Python
51 lines
1.3 KiB
Python
"""
|
|
reward model
|
|
"""
|
|
|
|
from typing import Callable, List, Optional
|
|
|
|
import torch
|
|
|
|
|
|
class RLVRRewardModel:
|
|
"""
|
|
RLVRReward model class. Support varifiable reward.
|
|
|
|
Args:
|
|
reward_fn_list List: list of reward functions
|
|
**kwargs: all other kwargs as in reward functions
|
|
"""
|
|
|
|
def __init__(self, reward_fn_list: List[Callable], **kwargs) -> None:
|
|
self.reward_fn_list = reward_fn_list
|
|
self.kwargs = kwargs
|
|
|
|
def __call__(
|
|
self,
|
|
input_ids: torch.LongTensor,
|
|
attention_mask: Optional[torch.Tensor] = None,
|
|
response_start: List = None,
|
|
response_end: List = None,
|
|
gt_answer: List = None,
|
|
) -> torch.Tensor:
|
|
# apply varifiable reward
|
|
bs = input_ids.size(0)
|
|
rewards = torch.zeros(bs, device=input_ids.device)
|
|
for i in range(bs):
|
|
for reward_fn in self.reward_fn_list:
|
|
rewards[i] += reward_fn(
|
|
input_ids[i],
|
|
attention_mask[i],
|
|
response_start=response_start[i],
|
|
response_end=response_end[i],
|
|
gt_answer=gt_answer[i],
|
|
**self.kwargs,
|
|
)
|
|
return rewards
|
|
|
|
def to(self, device):
|
|
return self
|
|
|
|
def eval(self):
|
|
return self
|