Merge 1a18e651f6 into a3c7b01cc0

trainer/rollout_engine.py

@@ -84,7 +84,7 @@ class TorchRolloutEngine(RolloutEngine):
             ).clone()  # [B*num_gen, P+R]
             prompt_len = prompt_ids.size(1)
             completion_ids = output_ids[:, prompt_len:]  # [B*num_gen, R]
-            full_mask = (output_ids != self.tokenizer.pad_token_id).long()
+            full_mask = torch.cat([attention_mask.repeat_interleave(num_generations, dim=0), attention_mask.new_ones(output_ids.size(0), completion_ids.size(1))], dim=1)
             per_token_logps = compute_per_token_logps(self.policy_model, output_ids, completion_ids.size(1), attention_mask=full_mask)
         completions = self.tokenizer.batch_decode(completion_ids, skip_special_tokens=True)
         return RolloutResult(output_ids, completion_ids, per_token_logps, completions,

trainer/train_agent.py

@@ -106,34 +106,33 @@ def rollout_single(rollout_engine, tokenizer, messages, tools, max_turns=3, max_
     open_thinking = random.random() < thinking_ratio
     for turn in range(max_turns):
         context = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, tools=tools, open_thinking=open_thinking)
-        inputs = tokenizer(context, return_tensors="pt", add_special_tokens=False).to(device)
-        context_ids = inputs["input_ids"][0].tolist()
         if prompt_ids is None:
-            prompt_ids = context_ids
+            prompt_ids = tokenizer(context, add_special_tokens=False)["input_ids"]
+        input_ids = torch.tensor([prompt_ids + response_ids], device=device)
         rollout_result = rollout_engine.rollout(
-            prompt_ids=inputs["input_ids"],
-            attention_mask=inputs["attention_mask"],
+            prompt_ids=input_ids,
+            attention_mask=torch.ones_like(input_ids),
             num_generations=1,
             max_new_tokens=max_new_tokens,
             temperature=0.8,
         )
-        new_ids = rollout_result.completion_ids[0].tolist()
-        new_logps = rollout_result.per_token_logps[0].tolist()
-        if len(new_ids) != len(new_logps): Logger(f"rollout token/logprob length mismatch: {len(new_ids)} vs {len(new_logps)}")
-        pairs = [(t, lp) for t, lp in zip(new_ids, new_logps) if t != tokenizer.pad_token_id and t != tokenizer.eos_token_id]
-        new_ids = [t for t, _ in pairs]
-        new_logps = [lp for _, lp in pairs]
+        valid_len = int(rollout_result.completion_mask[0].sum().item())
+        new_ids = rollout_result.completion_ids[0, :valid_len].tolist()
+        new_logps = rollout_result.per_token_logps[0, :valid_len].tolist()
+        if len(new_ids) != len(new_logps):
+            raise RuntimeError(f"rollout token/logprob length mismatch: {len(new_ids)} vs {len(new_logps)}")
         new_text = rollout_result.completions[0]
         all_outputs.append(new_text)
         response_ids.extend(new_ids)
-        response_mask.extend([1] * len(new_ids))
+        response_mask.extend([int(t != tokenizer.eos_token_id) for t in new_ids])
         response_old_logps.extend(new_logps)
         final_context = context + new_text
         calls = parse_tool_calls(new_text)
         if not calls:
             break
         unfinished = turn == max_turns - 1
-        messages.append({"role": "assistant", "content": new_text})
+        assistant_message = {"role": "assistant", "content": new_text}
+        messages.append(assistant_message)
         for call in calls:
             name, raw = call.get("name", ""), call.get("arguments", {})
             if isinstance(raw, str):
@@ -143,10 +142,17 @@ def rollout_single(rollout_engine, tokenizer, messages, tools, max_turns=3, max_
             result_str = (json.dumps(result, ensure_ascii=False) if result else '{"error": "tool not found"}')[:2048]  # 防止天文数字撑爆tokenizer
             messages.append({"role": "tool", "content": result_str})
 
+        marker = f"<|agent_observation_{id(messages)}_{len(response_ids)}|>"
+        assistant_message["content"] += marker
+        marked_context = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=not unfinished, tools=tools, open_thinking=open_thinking)
+        assistant_message["content"] = new_text
+        _, found, observation = marked_context.partition(marker)
+        if not found:
+            raise RuntimeError("chat template did not preserve the assistant content boundary")
+        obs_delta = tokenizer(observation, add_special_tokens=False)["input_ids"]
+        if new_ids and new_ids[-1] == tokenizer.eos_token_id and obs_delta[:1] == [tokenizer.eos_token_id]:
+            obs_delta = obs_delta[1:]
         observe_context = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=not unfinished, tools=tools, open_thinking=open_thinking)
-        observe_ids = tokenizer(observe_context, return_tensors="pt", add_special_tokens=False)["input_ids"][0].tolist()
-        current_len = len(prompt_ids) + len(response_ids)
-        obs_delta = observe_ids[current_len:]
         response_ids.extend(obs_delta)
         response_mask.extend([0] * len(obs_delta))
         response_old_logps.extend([0.0] * len(obs_delta))
@@ -268,7 +274,7 @@ def rl_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_model
         prompt_lens = torch.tensor([prompt_len for _, _, prompt_len, _ in packed_samples], device=args.device)
         full_response_masks = torch.tensor([mask + [0] * (max_len - len(mask)) for _, mask, _, _ in packed_samples], device=args.device, dtype=torch.float32)
         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()
+        full_mask = (torch.arange(max_len, device=args.device).unsqueeze(0) < seq_lens.unsqueeze(1)).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)
 

trainer/train_dpo.py

@@ -50,6 +50,19 @@ def dpo_loss(ref_log_probs, policy_log_probs, mask, beta):
     return loss.mean()
 
 
+def sft_ce_loss(logits, labels, mask):
+    """SFT正则项：对chosen回答部分计算标准语言建模交叉熵（按掩码位置取平均），
+    与SFT阶段的损失口径一致，用于防止偏好优化中遗忘语言能力。"""
+    vocab_size = logits.size(-1)
+    ce = F.cross_entropy(
+        logits.reshape(-1, vocab_size),
+        labels.reshape(-1),
+        reduction='none'
+    )
+    mask_flat = mask.reshape(-1).float()
+    return (ce * mask_flat).sum() / mask_flat.sum().clamp(min=1.0)
+
+
 def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=None, beta=0.1):
     start_time = time.time()
     last_step = start_step
@@ -81,7 +94,13 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=
             policy_log_probs = logits_to_log_probs(logits, y)
             
             dpo_loss_val = dpo_loss(ref_log_probs, policy_log_probs, mask, beta=beta)
-            loss = dpo_loss_val + outputs.aux_loss
+            # SFT/CE正则项：只对chosen回答部分计算交叉熵，权重由 --ce_alpha 控制
+            chosen_num = x.size(0) // 2
+            if args.ce_alpha > 0:
+                ce_loss_val = sft_ce_loss(logits[:chosen_num], y[:chosen_num], mask[:chosen_num])
+            else:
+                ce_loss_val = logits.new_zeros(())
+            loss = dpo_loss_val + args.ce_alpha * ce_loss_val + outputs.aux_loss
             loss = loss / args.accumulation_steps
 
         scaler.scale(loss).backward()
@@ -97,13 +116,14 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=
             spend_time = time.time() - start_time
             current_loss = loss.item() * args.accumulation_steps
             current_dpo_loss = dpo_loss_val.item()
+            current_ce_loss = ce_loss_val.item()
             current_aux_loss = outputs.aux_loss.item()
             current_lr = optimizer.param_groups[-1]['lr']
             eta_min = spend_time / max(step - start_step, 1) * (iters - step) // 60
             
-            Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, dpo_loss: {current_dpo_loss:.4f}, aux_loss: {current_aux_loss:.4f}, learning_rate: {current_lr:.8f}, epoch_time: {eta_min:.3f}min')
+            Logger(f'Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), loss: {current_loss:.4f}, dpo_loss: {current_dpo_loss:.4f}, ce_loss: {current_ce_loss:.4f}, aux_loss: {current_aux_loss:.4f}, learning_rate: {current_lr:.8f}, epoch_time: {eta_min:.3f}min')
             
-            if wandb: wandb.log({"loss": current_loss, "dpo_loss": current_dpo_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})
+            if wandb: wandb.log({"loss": current_loss, "dpo_loss": current_dpo_loss, "ce_loss": current_ce_loss, "aux_loss": current_aux_loss, "learning_rate": current_lr, "epoch_time": eta_min})
 
         if (step % args.save_interval == 0 or step == iters) and is_main_process():
             model.eval()
@@ -118,7 +138,7 @@ def train_epoch(epoch, loader, iters, ref_model, lm_config, start_step=0, wandb=
             del state_dict
 
         del x_chosen, x_rejected, y_chosen, y_rejected, mask_chosen, mask_rejected, x, y, mask
-        del ref_outputs, ref_logits, ref_log_probs, outputs, logits, policy_log_probs, loss
+        del ref_outputs, ref_logits, ref_log_probs, outputs, logits, policy_log_probs, ce_loss_val, loss
 
     if last_step > start_step and last_step % args.accumulation_steps != 0:
         scaler.unscale_(optimizer)
@@ -150,6 +170,7 @@ if __name__ == "__main__":
     parser.add_argument('--from_weight', default='full_sft', type=str, help="基于哪个权重训练")
     parser.add_argument('--from_resume', default=0, type=int, choices=[0, 1], help="是否自动检测&续训（0=否，1=是）")
     parser.add_argument('--beta', default=0.15, type=float, help="DPO中的beta参数")
+    parser.add_argument('--ce_alpha', default=0.1, type=float, help="SFT/CE正则项权重（0表示关闭，建议0.05~0.5；防止DPO训练中遗忘语言能力）")
     parser.add_argument("--use_wandb", action="store_true", help="是否使用wandb")
     parser.add_argument("--wandb_project", type=str, default="MiniMind-DPO", help="wandb项目名")
     parser.add_argument("--use_compile", default=0, type=int, choices=[0, 1], help="是否使用torch.compile加速（0=否，1=是）")
@@ -225,4 +246,4 @@ if __name__ == "__main__":
     # ========== 9. 清理分布进程 ==========
     if dist.is_initialized():
         dist.barrier()
-        dist.destroy_process_group()
+        dist.destroy_process_group()

trainer/train_grpo.py

@@ -91,6 +91,7 @@ def grpo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, reward_mod
         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)
+        full_mask.scatter_(1, logp_pos + 1, rollout_result.completion_mask.to(args.device, dtype=full_mask.dtype))
 
         rewards = calculate_rewards(prompts, completions, reward_model).to(args.device)  # [B*num_gen]
 

trainer/train_ppo.py

@@ -120,6 +120,7 @@ def ppo_train_epoch(epoch, loader, iters, rollout_engine, ref_model, actor_sched
         resp_idx = torch.arange(resp_labels.size(1), device=gen_out.device).unsqueeze(0)
         logp_pos = prompt_lens.unsqueeze(1) - 1 + resp_idx
         resp_pad_mask = rollout_result.completion_mask.to(args.device).bool()
+        full_mask.scatter_(1, logp_pos + 1, resp_pad_mask.to(full_mask.dtype))
         resp_lengths = resp_pad_mask.sum(dim=1); valid_resp = resp_lengths > 0; eos_mask = resp_labels.eq(tokenizer.eos_token_id) & resp_pad_mask
         has_eos = eos_mask.any(dim=1); eos_pos = torch.argmax(eos_mask.int(), dim=1)
         resp_lengths = torch.where(has_eos, eos_pos + 1, resp_lengths).long().clamp(min=1)