Add GRPO and Support RLVR for PPO (#6186)

* 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>

applications/ColossalChat/coati/trainer/ppo.py

@@ -3,13 +3,13 @@ PPO trainer
 """
 
 import os
-from typing import Dict, List, Optional
+from typing import Dict, List, Optional, Union
 
 import torch
 import wandb
 from coati.experience_buffer import NaiveExperienceBuffer
 from coati.experience_maker import Experience, NaiveExperienceMaker
-from coati.models import Critic, RewardModel
+from coati.models import Critic, RewardModel, RLVRRewardModel
 from coati.models.loss import GPTLMLoss, PolicyLoss, ValueLoss
 from coati.models.utils import calc_action_log_probs
 from coati.trainer.callbacks import Callback
@@ -84,7 +84,7 @@ class PPOTrainer(OLTrainer):
         critic_booster: Booster,
         actor: PreTrainedModel,
         critic: Critic,
-        reward_model: RewardModel,
+        reward_model: Union[RewardModel, RLVRRewardModel],
         initial_model: PreTrainedModel,
         actor_optim: Optimizer,
         critic_optim: Optimizer,
@@ -210,6 +210,7 @@ class PPOTrainer(OLTrainer):
         return self.experience_maker.make_experience(
             input_ids=prompts["input_ids"].to(get_current_device()),
             attention_mask=prompts["attention_mask"].to(get_current_device()),
+            gt_answer=prompts["gt_answer"],
             **self.generate_kwargs,
         )