filter flags

verl/trainer/ppo/ray_trainer.py

@@ -313,6 +313,10 @@ class RayPPOTrainer(object):
         self._validate_config()
         self._create_dataloader()
 
+        # effective dapo
+        self.filter_flags = {}
+
+
     def _validate_config(self):
         config = self.config
         # number of GPUs total
@@ -811,6 +815,11 @@ class RayPPOTrainer(object):
         with open(local_latest_checkpointed_iteration, 'w') as f:
             f.write(str(self.global_steps))
 
+        # save filter flags
+        local_latest_filter_flags_path = os.path.join(local_global_step_folder, 'latest_filter_flags.json')
+        with open(local_latest_filter_flags_path, 'w', encoding='utf-8') as f:
+            json.dump(self.filter_flags, f)
+
     def _load_checkpoint(self):
         if self.config.trainer.resume_mode == 'disable':
             return 0
@@ -864,6 +873,15 @@ class RayPPOTrainer(object):
         else:
             print(f"Warning: No dataloader state found at {dataloader_local_path}, will start from scratch")
 
+        # load filter flags
+        local_latest_filter_flags_path = os.path.join(global_step_folder, 'latest_filter_flags.json')
+        try:
+            with open(local_latest_filter_flags_path, 'r', encoding='utf-8') as f:
+                self.filter_flags = json.load(f)
+        except Exception as e:
+            print(f'Failed to load filter flags from {local_latest_filter_flags_path}.')
+            self.filter_flags = {}
+
     def _balance_batch(self, batch: DataProto, metrics, logging_prefix='global_seqlen'):
         """Reorder the data on single controller such that each dp rank gets similar total tokens"""
         attention_mask = batch.batch['attention_mask']
@@ -925,6 +943,26 @@ class RayPPOTrainer(object):
                 metrics = {}
 
                 new_batch: DataProto = DataProto.from_single_dict(batch_dict)
+
+                if 'problem_id' in new_batch.non_tensor_batch.keys():
+                    filter_key = 'problem_id'
+                elif 'question' in new_batch.batch.keys():
+                    filter_key = 'question'
+                else:
+                    filter_key = 'raw_prompt_ids' 
+
+                select_idx = []
+                for id, item in enumerate(new_batch.non_tensor_batch[filter_key]):
+                    if item in self.filter_flags and self.filter_flags[item] == 1:
+                        continue
+                    select_idx.append(id)
+                
+                if len(select_idx) == 0:
+                    # 换下一批样本继续生成吧
+                    continue
+
+                new_batch = new_batch.select_idxs(select_idx)
+
                 num_gen_batches += 1
                 # pop those keys for generation
                 if 'multi_modal_inputs' in new_batch.non_tensor_batch.keys():
@@ -1037,6 +1075,18 @@ class RayPPOTrainer(object):
                         for idx, traj_from_prompt_uid in enumerate(new_batch.non_tensor_batch['uid']):
                             if traj_from_prompt_uid in kept_prompt_uids:
                                 kept_traj_idxs.append(idx)
+                                    
+                        # filter uids
+                        filter_prompt_uids = [uid for uid, std in prompt_uid2metric_std.items() if std == 0 and prompt_uid2metric_vals[uid] != 0]
+                        filter_traj_idxs = []
+                        for idx, traj_from_prompt_uid in enumerate(new_batch.non_tensor_batch['uid']):
+                            if traj_from_prompt_uid in filter_prompt_uids:
+                                filter_traj_idxs.append(idx)
+                        
+                        for idx in filter_traj_idxs:
+                            print(new_batch.non_tensor_batch[filter_key][idx])
+                            self.filter_flags[new_batch.non_tensor_batch[filter_key][idx]] = 1
+
 
                         new_batch = new_batch[kept_traj_idxs]
                         if batch is None: