import os import sys __package__ = "trainer" sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..'))) import argparse import re import warnings import torch import torch.distributed as dist import torch.nn.functional as F from transformers import AutoTokenizer from contextlib import nullcontext from torch import optim from torch.nn.parallel import DistributedDataParallel from torch.utils.data import DataLoader, DistributedSampler from torch.nn.utils import clip_grad_norm_ from torch.optim.lr_scheduler import CosineAnnealingLR from transformers import AutoModel from model.model_minimind import MiniMindConfig, MiniMindForCausalLM from dataset.lm_dataset import RLAIFDataset from trainer.trainer_utils import Logger, is_main_process, lm_checkpoint, init_distributed_mode, setup_seed, SkipBatchSampler, init_model warnings.filterwarnings('ignore') def calculate_rewards(prompts, responses, reward_model, reward_tokenizer, args): """ 计算复杂奖励，包含格式奖励、模型打分以及 DAPO 特有的长度惩罚。 输入的 prompts/responses 已经是展开后的（batch_size * group_size）条目。 """ def reasoning_model_reward(rewards): # 1. 结构化格式奖励 pattern = r"^\n.*?\n\n\n.*?\n$" pattern2 = r"^\n.*?\n\n\n\n.*?\n$" matches_pattern = [re.match(pattern, r, re.S) for r in responses] matches_pattern2 = [re.match(pattern2, r, re.S) for r in responses] format_rewards = [0.5 if (m1 or m2) else 0.0 for m1, m2 in zip(matches_pattern, matches_pattern2)] rewards += torch.tensor(format_rewards, device=args.device) # 2. 标签存在奖励 def mark_num(text): reward = 0 for tag in ["", "", "", ""]: if text.count(tag) == 1: reward += 0.25 return reward mark_rewards = [mark_num(r) for r in responses] rewards += torch.tensor(mark_rewards, device=args.device) return rewards rewards = torch.zeros(len(responses), device=args.device) if args.reasoning == 1: rewards = reasoning_model_reward(rewards) # 3. 超长惩罚 (Overlong Punishment - DAPO 特性) len_penalty = [] for resp in responses: curr_len = len(resp) if curr_len > args.length_threshold: penalty = -args.length_penalty * ((curr_len - args.length_threshold) / 100) len_penalty.append(max(penalty, -1.0)) else: len_penalty.append(0.0) rewards += torch.tensor(len_penalty, device=args.device) # 4. 奖励模型评分 with torch.no_grad(): reward_model_scores = [] for prompt, response in zip(prompts, responses): pattern = r"<\|im_start\|>(system|user|assistant)\s+(.*?)<\|im_end\|>" matches = re.findall(pattern, prompt, re.DOTALL) messages = [{"role": role, "content": content.strip()} for role, content in matches] tmp_chat = messages + [{"role": "assistant", "content": response}] score = reward_model.get_score(reward_tokenizer, tmp_chat) scale = 3.0 score = max(min(score, scale), -scale) if args.reasoning == 1: answer_match = re.search(r'(.*?)', response, re.DOTALL) if answer_match: answer_content = answer_match.group(1).strip() tmp_chat = messages + [{"role": "assistant", "content": answer_content}] answer_score = reward_model.get_score(reward_tokenizer, tmp_chat) answer_score = max(min(answer_score, scale), -scale) score = score * 0.4 + answer_score * 0.6 reward_model_scores.append(score) reward_model_scores = torch.tensor(reward_model_scores, device=args.device) rewards += reward_model_scores return rewards def dapo_train_epoch(epoch, loader, iters, old_actor_model, ref_model, actor_scheduler, reward_model, reward_tokenizer, start_step=0, wandb=None): actor_model.train() for step, batch in enumerate(loader, start=start_step + 1): prompts = batch["prompt"] batch_size = len(prompts) G = args.group_size # 每个 prompt 重复 G 次，展开为 (batch_size * G) 条输入 prompts_expanded = [p for p in prompts for _ in range(G)] enc = tokenizer( prompts_expanded, return_tensors="pt", padding=True, truncation=True, max_length=args.max_seq_len, padding_side="left" ).to(args.device) prompt_length = enc.input_ids.shape[1] with torch.no_grad(): model_for_gen = actor_model.module if isinstance(actor_model, DistributedDataParallel) else actor_model gen_out = model_for_gen.generate( input_ids=enc.input_ids, attention_mask=enc.attention_mask, max_new_tokens=args.max_gen_len, do_sample=True, temperature=0.8, pad_token_id=tokenizer.pad_token_id, eos_token_id=tokenizer.eos_token_id, ) # gen_out: (batch_size * G, seq_len) responses_text = [ tokenizer.decode(gen_out[i, prompt_length:], skip_special_tokens=True) for i in range(len(prompts_expanded)) ] rewards_flat = calculate_rewards(prompts_expanded, responses_text, reward_model, reward_tokenizer, args) # rewards_flat: (batch_size * G,) # ------------------------------------------------------------------ # advantage = (r - mean_group) / std_group # ------------------------------------------------------------------ rewards_grouped = rewards_flat.view(batch_size, G) mean_r = rewards_grouped.mean(dim=1, keepdim=True) std_r = rewards_grouped.std(dim=1, keepdim=True).clamp(min=1e-8) advantages = ((rewards_grouped - mean_r) / std_r).view(-1) # (batch_size * G,) # ------------------------------------------------------------------ # Dynamic Sampling # ------------------------------------------------------------------ if args.use_dynamic_sampling == 1: if (std_r < 1e-6).all().item(): if is_main_process(): Logger(f"Step {step}: All groups have zero reward variance, skipping...") continue # ------------------------------------------------------------------ # Actor forward # ------------------------------------------------------------------ full_mask = (gen_out != tokenizer.pad_token_id).long() with autocast_ctx: res = actor_model(input_ids=gen_out, attention_mask=full_mask) logits = res.logits aux_loss = res.aux_loss if lm_config.use_moe else torch.tensor(0.0, device=args.device) labels = gen_out[:, 1:].clone() logp_tokens = F.log_softmax(logits[:, :-1], dim=-1).gather(2, labels.unsqueeze(-1)).squeeze(-1) seq_len = gen_out.size(1) - 1 resp_mask = torch.arange(seq_len, device=gen_out.device).unsqueeze(0) >= prompt_length - 1 final_mask = resp_mask & (~labels.eq(tokenizer.pad_token_id)) # Token-level 平均 log-prob（防止长回复主导梯度） token_count = final_mask.sum(dim=1).clamp(min=1).float() actor_logp = (logp_tokens * final_mask).sum(dim=1) / token_count with torch.no_grad(): old_logits = old_actor_model(input_ids=gen_out, attention_mask=full_mask).logits old_logp_tokens = F.log_softmax(old_logits[:, :-1], dim=-1).gather(2, labels.unsqueeze(-1)).squeeze(-1) old_logp = (old_logp_tokens * final_mask).sum(dim=1) / token_count ref_logits = ref_model(input_ids=gen_out, attention_mask=full_mask).logits ref_logp_tokens = F.log_softmax(ref_logits[:, :-1], dim=-1).gather(2, labels.unsqueeze(-1)).squeeze(-1) ref_logp = (ref_logp_tokens * final_mask).sum(dim=1) / token_count kl = (actor_logp - old_logp).mean() kl_ref = (actor_logp - ref_logp).mean() ratio = torch.exp(actor_logp - old_logp) # ------------------------------------------------------------------ # Decoupled Clip # Adv > 0：只限上界，允许正向大步更新prompts_expanded = [p for p in prompts for _ in range(G)] # Adv ≤ 0：只限下界，防止负向更新反弹 # ------------------------------------------------------------------ surr1 = ratio * advantages clipped_ratio = torch.where( advantages > 0, torch.clamp(ratio, max=1.0 + args.clip_epsilon_high), torch.clamp(ratio, min=1.0 - args.clip_epsilon_low), ) surr2 = clipped_ratio * advantages policy_loss = -torch.min(surr1, surr2).mean() loss = (policy_loss + args.kl_coef * kl_ref + aux_loss) / args.accumulation_steps loss.backward() if (step + 1) % args.accumulation_steps == 0: clip_grad_norm_(actor_model.parameters(), args.grad_clip) actor_optimizer.step() actor_scheduler.step() actor_optimizer.zero_grad() if is_main_process(): response_ids = gen_out[:, enc.input_ids.shape[1]:] is_eos = (response_ids == tokenizer.eos_token_id) eos_indices = torch.argmax(is_eos.int(), dim=1) lengths = torch.where( is_eos.any(dim=1), eos_indices + 1, torch.tensor(response_ids.shape[1], device=is_eos.device) ) avg_len = lengths.float().mean() reward_std = rewards_flat.std().item() if wandb is not None: wandb.log({ "actor_loss": policy_loss.item(), "reward": rewards_flat.mean().item(), "reward_std": reward_std, "kl": kl.item(), "kl_ref": kl_ref.item(), "avg_response_len": avg_len.item(), "actor_lr": actor_optimizer.param_groups[0]['lr'], }) Logger( f"Epoch:[{epoch + 1}/{args.epochs}]({step}/{iters}), " f"Loss: {policy_loss.item():.4f}, Reward: {rewards_flat.mean().item():.4f}, " f"KL: {kl.item():.4f}, Len: {avg_len.item():.2f}" ) if (step + 1) % args.update_old_actor_freq == 0: raw_actor = actor_model.module if isinstance(actor_model, DistributedDataParallel) else actor_model raw_actor = getattr(raw_actor, '_orig_mod', raw_actor) old_actor_model.load_state_dict({k: v.detach().cpu() for k, v in raw_actor.state_dict().items()}) old_actor_model.to(args.device) if (step % args.save_interval == 0 or step == iters - 1) and is_main_process(): actor_model.eval() moe_suffix = '_moe' if lm_config.use_moe else '' ckp = f'{args.save_dir}/{args.save_weight}_{lm_config.hidden_size}{moe_suffix}.pth' raw_actor = actor_model.module if isinstance(actor_model, DistributedDataParallel) else actor_model raw_actor = getattr(raw_actor, '_orig_mod', raw_actor) torch.save({k: v.half().cpu() for k, v in raw_actor.state_dict().items()}, ckp) lm_checkpoint( lm_config, weight=args.save_weight, model=actor_model, optimizer=actor_optimizer, epoch=epoch, step=step, wandb=wandb, save_dir='../checkpoints', scheduler=actor_scheduler, ) actor_model.train() del enc, gen_out, responses_text, rewards_flat, rewards_grouped del advantages, full_mask, logits, labels, logp_tokens, final_mask, token_count del actor_logp, old_logits, old_logp, ref_logits, ref_logp del kl, kl_ref, ratio, surr1, surr2, clipped_ratio, policy_loss, loss if __name__ == "__main__": parser = argparse.ArgumentParser( description="MiniMind DAPO (Decoupled Clip and Dynamic Sampling Policy Optimization)" ) parser.add_argument("--save_dir", type=str, default="../out", help="模型保存目录") parser.add_argument('--save_weight', type=str, default='dapo', help="保存权重的前缀名") parser.add_argument("--epochs", type=int, default=1, help="训练轮数") parser.add_argument("--batch_size", type=int, default=2, help="prompt 级别的 batch size（实际前向 = batch_size * group_size）") parser.add_argument("--learning_rate", type=float, default=8e-8, help="Actor 学习率") parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu", help="训练设备") parser.add_argument("--dtype", type=str, default="bfloat16", help="混合精度类型") parser.add_argument("--num_workers", type=int, default=8, help="数据加载线程数") parser.add_argument("--accumulation_steps", type=int, default=1, help="梯度累积步数") parser.add_argument("--grad_clip", type=float, default=1.0, help="梯度裁剪阈值") parser.add_argument("--log_interval", type=int, default=1, help="日志打印间隔") parser.add_argument("--save_interval", type=int, default=10, help="模型保存间隔") parser.add_argument('--hidden_size', type=int, default=512, help="隐藏层维度") parser.add_argument('--num_hidden_layers', type=int, default=8, help="隐藏层数量") parser.add_argument('--use_moe', type=int, default=0, choices=[0, 1], help="是否使用 MoE 架构（0=否，1=是）") parser.add_argument('--max_seq_len', type=int, default=66, help="Prompt 最大长度") parser.add_argument("--max_gen_len", type=int, default=1024, help="生成的最大长度") parser.add_argument("--data_path", type=str, default="../dataset/rlaif-mini.jsonl", help="RLAIF 数据路径") # DAPO 核心参数 parser.add_argument("--clip_epsilon_low", type=float, default=0.1, help="Decoupled Clip 下界") parser.add_argument("--clip_epsilon_high", type=float, default=0.2, help="Decoupled Clip 上界") parser.add_argument("--group_size", type=int, default=8, help="每个 prompt 的采样回复数") parser.add_argument("--use_dynamic_sampling", type=int, default=1, choices=[0, 1], help="是否开启动态采样过滤") parser.add_argument("--length_penalty", type=float, default=0.1, help="长度惩罚系数") parser.add_argument("--length_threshold", type=int, default=1200, help="触发长度惩罚的字符阈值") parser.add_argument("--kl_coef", type=float, default=0.02, help="KL 散度惩罚系数（相对 ref_model）") parser.add_argument("--reasoning", type=int, default=1, choices=[0, 1], help="推理模型类型（0=普通模型，1=推理模型）") parser.add_argument("--update_old_actor_freq", type=int, default=4, help="更新 old_actor_model 的频率（步数）") parser.add_argument("--reward_model_path", type=str, default="../../internlm2-1_8b-reward", help="Reward 模型路径") parser.add_argument('--from_resume', type=int, default=0, choices=[0, 1], help="是否自动检测并续训（0=否，1=是）") parser.add_argument("--use_wandb", action="store_true", help="是否使用 wandb/swanlab") parser.add_argument("--wandb_project", type=str, default="MiniMind-DAPO", help="wandb 项目名") parser.add_argument("--use_compile", type=int, default=0, choices=[0, 1], help="是否使用 torch.compile 加速（0=否，1=是）") args = parser.parse_args() # ========== 1. 初始化环境和随机种子 ========== local_rank = init_distributed_mode() if dist.is_initialized(): args.device = f"cuda:{local_rank}" setup_seed(42 + (dist.get_rank() if dist.is_initialized() else 0)) # ========== 2. 配置目录、模型参数、检查 ckp ========== os.makedirs(args.save_dir, exist_ok=True) lm_config = MiniMindConfig( hidden_size=args.hidden_size, num_hidden_layers=args.num_hidden_layers, use_moe=bool(args.use_moe) ) ckp_data = lm_checkpoint(lm_config, weight=args.save_weight, save_dir='../checkpoints') \ if args.from_resume == 1 else None # ========== 3. 设置混合精度 ========== device_type = "cuda" if "cuda" in args.device else "cpu" dtype = torch.bfloat16 if args.dtype == "bfloat16" else torch.float16 autocast_ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast(dtype=dtype) # ========== 4. 配 wandb (SwanLab) ========== wandb = None if args.use_wandb and is_main_process(): import swanlab as wandb wandb_id = ckp_data.get('wandb_id') if ckp_data else None resume = 'must' if wandb_id else None wandb.init( project=args.wandb_project, name=f"MiniMind-DAPO-{args.learning_rate}", id=wandb_id, resume=resume, ) # ========== 5. 初始化模型 ========== base_weight = "reason" if args.reasoning == 1 else "full_sft" actor_model, tokenizer = init_model(lm_config, base_weight, device=args.device) if args.use_compile == 1: actor_model = torch.compile(actor_model) old_actor_model, _ = init_model(lm_config, base_weight, device=args.device) old_actor_model = old_actor_model.eval().requires_grad_(False) ref_model, _ = init_model(lm_config, base_weight, device=args.device) ref_model = ref_model.eval().requires_grad_(False) reward_model = AutoModel.from_pretrained( args.reward_model_path, torch_dtype=torch.float16, trust_remote_code=True, ) reward_model = reward_model.to(args.device).eval().requires_grad_(False) reward_tokenizer = AutoTokenizer.from_pretrained(args.reward_model_path, trust_remote_code=True) # ========== 6. 数据、优化器、调度器 ========== train_ds = RLAIFDataset(args.data_path, tokenizer, max_length=(args.max_seq_len + args.max_gen_len)) train_sampler = DistributedSampler(train_ds) if dist.is_initialized() else None actor_optimizer = optim.AdamW(actor_model.parameters(), lr=args.learning_rate) iters = len(DataLoader(train_ds, batch_size=args.batch_size, sampler=train_sampler)) total_steps = (iters // args.accumulation_steps) * args.epochs actor_scheduler = CosineAnnealingLR(actor_optimizer, T_max=total_steps, eta_min=args.learning_rate / 10) if ckp_data: actor_model.load_state_dict(ckp_data['model']) actor_optimizer.load_state_dict(ckp_data['optimizer']) actor_scheduler.load_state_dict(ckp_data['scheduler']) start_epoch = ckp_data['epoch'] start_step = ckp_data.get('step', 0) else: start_epoch, start_step = 0, 0 if dist.is_initialized(): actor_model = DistributedDataParallel(actor_model, device_ids=[local_rank]) old_actor_model.to(args.device) # ========== 7. 打印配置摘要 ========== if is_main_process(): Logger("=" * 60) Logger("DAPO Training Config (Critic-free):") Logger(f" group_size = {args.group_size}") Logger(f" clip_epsilon_low = {args.clip_epsilon_low}") Logger(f" clip_epsilon_high = {args.clip_epsilon_high}") Logger(f" use_dynamic_sampling = {args.use_dynamic_sampling}") Logger(f" kl_coef = {args.kl_coef}") Logger(f" effective_batch = {args.batch_size} prompts × {args.group_size} = {args.batch_size * args.group_size}") Logger("=" * 60) # ========== 8. 开始训练 ========== for epoch in range(start_epoch, args.epochs): train_sampler and train_sampler.set_epoch(epoch) setup_seed(42 + epoch) indices = torch.randperm(len(train_ds)).tolist() skip = start_step if (epoch == start_epoch and start_step > 0) else 0 loader = DataLoader( train_ds, batch_sampler=SkipBatchSampler(train_sampler or indices, args.batch_size, skip), num_workers=args.num_workers, pin_memory=True, ) dapo_train_epoch( epoch, loader, len(loader), old_actor_model, ref_model, actor_scheduler, reward_model, reward_tokenizer, skip, wandb, ) if dist.is_initialized(): dist.destroy_process_group()