[feat] reduce RL memory

fix: resolve CUDA OOM in train_grpo.py on GPUs with <=8GB VRAM

trainer/train_agent.py

@@ -269,6 +269,8 @@ def rl_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_model
         old_per_token_logps = torch.tensor([old_logps + [0.0] * ((max_len - 1) - len(old_logps)) for _, _, _, old_logps in packed_samples], device=args.device, dtype=torch.float32)
         full_mask = (input_ids != tokenizer.pad_token_id).long()
 
+        rewards = calculate_rewards(prompts, completions, gt_batch, tools_batch, args.num_generations, reward_model, device=args.device, turn_outputs_batch=turn_outputs_batch, unfinished_batch=unfinished_batch)
+
         model_unwrapped = model.module if isinstance(model, DistributedDataParallel) else model
         with autocast_ctx:
             res = model_unwrapped(input_ids, attention_mask=full_mask)
@@ -288,7 +290,6 @@ def rl_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_model
         completion_mask = completion_mask * (pos <= eos_idx.unsqueeze(1)).float()
         token_counts = completion_mask.sum(dim=1)
         valid_rows = token_counts > 0
-        rewards = calculate_rewards(prompts, completions, gt_batch, tools_batch, args.num_generations, reward_model, device=args.device, turn_outputs_batch=turn_outputs_batch, unfinished_batch=unfinished_batch)
 
         if args.debug_mode and is_main_process() and step % args.debug_interval == 0:
             for i in range(len(messages_batch)):

trainer/train_grpo.py

@@ -86,11 +86,13 @@ def grpo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_mod
         outputs = rollout_result.output_ids
         completion_ids = rollout_result.completion_ids
         completions = rollout_result.completions
-        old_per_token_logps = rollout_result.per_token_logps.to(args.device)
+        old_per_token_logps = rollout_result.per_token_logps.to(args.device).detach()
         prompt_lens = rollout_result.prompt_lens.to(args.device)
         full_mask = (outputs != tokenizer.pad_token_id).long()
         logp_pos = prompt_lens.unsqueeze(1) - 1 + torch.arange(completion_ids.size(1), device=args.device).unsqueeze(0)
 
+        rewards = calculate_rewards(prompts, completions, reward_model).to(args.device)  # [B*num_gen]
+
         model_unwrapped = model.module if isinstance(model, DistributedDataParallel) else model
         with autocast_ctx:
             res = model_unwrapped(outputs, attention_mask=full_mask)
@@ -99,7 +101,6 @@ def grpo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_mod
 
         with torch.no_grad():
             ref_per_token_logps = F.log_softmax(ref_model(outputs, attention_mask=full_mask).logits[:, :-1, :], dim=-1).gather(2, outputs[:, 1:].unsqueeze(-1)).squeeze(-1).gather(1, logp_pos)
-        rewards = calculate_rewards(prompts, completions, reward_model).to(args.device)  # [B*num_gen]
 
         if args.debug_mode and is_main_process() and step % args.debug_interval == 0:
             for i in range(len(prompts)):