""" DeepSpeed ZeRO-2 SearchShortQA GRPO 入口。 推荐在 trainer/ 目录启动： deepspeed --num_gpus=7 train_search_grpo_deepspeed.py --model_profile SearchLM-300M """ import os import sys __package__ = "trainer" sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))) import argparse import math import warnings from contextlib import nullcontext import torch import torch.distributed as dist import torch.nn.functional as F from torch.utils.data import DataLoader, DistributedSampler from dataset.search_shortqa_dataset import SearchShortQARLDataset from trainer.rollout_engine import create_rollout_engine from trainer.search_shortqa_reward import average_metrics, score_search_shortqa_response from trainer.trainer_utils import ( Logger, SkipBatchSampler, add_deepspeed_args, add_model_profile_args, apply_model_profile, build_lm_config, assert_no_training_output_collision, init_model, is_main_process, get_cuda_peak_memory_gb, load_deepspeed_checkpoint, load_deepspeed_config, log_training_setup, reduce_metrics, resolve_attention_type, save_deepspeed_checkpoint, save_model_weights, setup_seed, ) warnings.filterwarnings("ignore") def init_deepspeed_runtime(args, deepspeed): deepspeed.init_distributed(dist_backend="nccl") local_rank = int(os.environ.get("LOCAL_RANK", args.local_rank if args.local_rank >= 0 else 0)) if torch.cuda.is_available(): torch.cuda.set_device(local_rank) args.device = f"cuda:{local_rank}" else: args.device = "cpu" args.training_backend = "DeepSpeed-ZeRO" setup_seed(42 + (dist.get_rank() if dist.is_initialized() else local_rank)) return local_rank def collate_fn(batch): return { "prompt": [x["prompt"] for x in batch], "question": [x["question"] for x in batch], "answers": [x["answers"] for x in batch], "gold_doc_ids": [x["gold_doc_ids"] for x in batch], "contexts": [x["contexts"] for x in batch], } def calculate_search_rewards(args, completions, batch): rewards = [] metrics = [] for i, response in enumerate(completions): sample_idx = i // args.num_generations reward, parts = score_search_shortqa_response( response=response, answers=batch["answers"][sample_idx], gold_doc_ids=batch["gold_doc_ids"][sample_idx], contexts=batch["contexts"][sample_idx], ) rewards.append(reward) metrics.append(parts) return torch.tensor(rewards, dtype=torch.float32, device=args.device), average_metrics(metrics) def train_epoch(args, lm_config, model_engine, tokenizer, loader, iters, rollout_engine, ref_model, autocast_ctx, start_step=0, wandb=None): for step, batch in enumerate(loader, start=start_step + 1): prompts = batch["prompt"] prompt_inputs = tokenizer( prompts, return_tensors="pt", padding=True, return_token_type_ids=False, padding_side="left", add_special_tokens=False, ).to(args.device) if args.max_seq_len: prompt_inputs["input_ids"] = prompt_inputs["input_ids"][:, -args.max_seq_len:] prompt_inputs["attention_mask"] = prompt_inputs["attention_mask"][:, -args.max_seq_len:] rollout_result = rollout_engine.rollout( prompt_ids=prompt_inputs["input_ids"], attention_mask=prompt_inputs["attention_mask"], num_generations=args.num_generations, max_new_tokens=args.max_gen_len, temperature=args.temperature, ) 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).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, reward_stats = calculate_search_rewards(args, completions, batch) res = model_engine(outputs, attention_mask=full_mask) aux_loss = res.aux_loss if getattr(lm_config, "use_moe", False) and res.aux_loss is not None else torch.tensor(0.0, device=args.device) per_token_logps = ( F.log_softmax(res.logits[:, :-1, :], dim=-1) .gather(2, outputs[:, 1:].unsqueeze(-1)) .squeeze(-1) .gather(1, logp_pos) ) with torch.no_grad(), autocast_ctx: 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) ) grouped_rewards = rewards.view(-1, args.num_generations) mean_r = grouped_rewards.mean(dim=1).repeat_interleave(args.num_generations) std_r = grouped_rewards.std(dim=1, unbiased=False).repeat_interleave(args.num_generations) advantages = (rewards - mean_r) / (std_r + 1e-4) completion_pad_mask = rollout_result.completion_mask.to(args.device).bool() is_eos = (completion_ids == tokenizer.eos_token_id) & completion_pad_mask eos_idx = torch.full((is_eos.size(0),), is_eos.size(1) - 1, dtype=torch.long, device=args.device) eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)] completion_mask = ( (torch.arange(is_eos.size(1), device=args.device).expand(is_eos.size(0), -1) <= eos_idx.unsqueeze(1)) & completion_pad_mask ).int() kl_div = ref_per_token_logps - per_token_logps per_token_kl = torch.exp(kl_div) - kl_div - 1 ratio = torch.exp(per_token_logps - old_per_token_logps) clipped_ratio = torch.clamp(ratio, 1 - args.epsilon, 1 + args.epsilon) per_token_loss1 = ratio * advantages.unsqueeze(1) per_token_loss2 = clipped_ratio * advantages.unsqueeze(1) per_token_loss = -(torch.min(per_token_loss1, per_token_loss2) - args.beta * per_token_kl) policy_loss = ((per_token_loss * completion_mask).sum(dim=1) / completion_mask.sum(dim=1).clamp(min=1)).mean() loss = policy_loss + aux_loss model_engine.backward(loss) model_engine.step() if step % args.log_interval == 0 or step == iters: reward_val = rewards.mean().item() avg_len_val = completion_mask.sum(dim=1).float().mean().item() kl_ref_val = ((ref_per_token_logps - per_token_logps) * completion_mask).sum().item() / max(completion_mask.sum().item(), 1) lr_list = model_engine.get_lr() current_lr = lr_list[0] if lr_list else args.learning_rate metrics = reduce_metrics({ "reward": reward_val, "avg_len": avg_len_val, "kl_ref": kl_ref_val, "policy_loss": policy_loss.item(), "aux_loss": aux_loss.item(), **reward_stats, }) peak_memory_gb = get_cuda_peak_memory_gb() memory_text = f", peak_memory: {peak_memory_gb:.2f} GB" if peak_memory_gb is not None else "" Logger( f"Epoch:[{args.current_epoch + 1}/{args.epochs}]({step}/{iters}), " f"Reward: {metrics['reward']:.4f}, F1: {metrics.get('answer_f1', 0):.4f}, " f"EM: {metrics.get('exact_match', 0):.4f}, Cite: {metrics.get('citation_score', 0):.4f}, " f"Format: {metrics.get('format_score', 0):.4f}, KL_ref: {metrics['kl_ref']:.4f}, " f"Avg Len: {metrics['avg_len']:.2f}, LR: {current_lr:.8f}" f"{memory_text}" ) if wandb and is_main_process(): log_data = { "reward": metrics["reward"], "kl_ref": metrics["kl_ref"], "policy_loss": metrics["policy_loss"], "aux_loss": metrics["aux_loss"], "avg_response_len": metrics["avg_len"], "learning_rate": current_lr, } log_data.update({key: value for key, value in metrics.items() if key not in log_data}) wandb.log(log_data) if step % args.save_interval == 0 or step == iters: model_engine.eval() save_deepspeed_checkpoint( model_engine, lm_config, weight=args.save_weight, epoch=args.current_epoch, step=step, wandb=wandb, save_dir="../checkpoints", ) if is_main_process(): ckp = save_model_weights(lm_config, model_engine, save_dir=args.save_dir, weight=args.save_weight) Logger(f"Lightweight model saved: {ckp}") model_engine.train() rollout_engine.update_policy(model_engine) del prompt_inputs, outputs, completion_ids, per_token_logps, ref_per_token_logps del completions, rewards, grouped_rewards, mean_r, std_r, advantages, completion_mask, completion_pad_mask, prompt_lens del res, loss, policy_loss, aux_loss def parse_args(): parser = argparse.ArgumentParser(description="SearchShortQA GRPO with DeepSpeed ZeRO") parser.add_argument("--save_dir", type=str, default="../out") parser.add_argument("--save_weight", default="search_grpo", type=str) parser.add_argument("--epochs", type=int, default=1) parser.add_argument("--batch_size", type=int, default=2) parser.add_argument("--learning_rate", type=float, default=3e-7) parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu") parser.add_argument("--dtype", type=str, default="float16", choices=["float16", "bfloat16", "float32"]) parser.add_argument("--num_workers", type=int, default=8) parser.add_argument("--accumulation_steps", type=int, default=1) parser.add_argument("--grad_clip", type=float, default=1.0) parser.add_argument("--log_interval", type=int, default=1) parser.add_argument("--save_interval", type=int, default=50) parser.add_argument("--hidden_size", default=768, type=int) parser.add_argument("--num_hidden_layers", default=8, type=int) parser.add_argument("--num_attention_heads", default=8, type=int) parser.add_argument("--num_key_value_heads", default=4, type=int) parser.add_argument("--intermediate_size", default=None, type=int) parser.add_argument("--use_moe", default=0, type=int, choices=[0, 1]) parser.add_argument("--attention_type", default="gqa", choices=["gqa", "mha", "mqa", "mla"]) parser.add_argument("--use_mla", default=0, type=int, choices=[0, 1]) parser.add_argument("--kv_lora_rank", default=128, type=int) parser.add_argument("--q_lora_rank", default=256, type=int) parser.add_argument("--rope_dim", default=None, type=int) parser.add_argument("--max_seq_len", default=1024, type=int) parser.add_argument("--max_gen_len", type=int, default=256) parser.add_argument("--data_path", type=str, default="../dataset/search_shortqa_train.jsonl") parser.add_argument("--num_generations", type=int, default=6) parser.add_argument("--beta", type=float, default=0.05) parser.add_argument("--epsilon", type=float, default=0.2) parser.add_argument("--temperature", type=float, default=0.8) parser.add_argument("--from_weight", default="full_sft_search", type=str) parser.add_argument("--from_resume", default=0, type=int, choices=[0, 1]) parser.add_argument("--overwrite_output", default=0, type=int, choices=[0, 1]) parser.add_argument("--use_wandb", action="store_true") parser.add_argument("--wandb_project", type=str, default="SearchShortQA-GRPO-DS") parser.add_argument("--use_compile", default=0, type=int, choices=[0, 1]) parser.add_argument("--rollout_engine", type=str, default="torch", choices=["torch", "sglang"]) parser.add_argument("--sglang_base_url", type=str, default="http://localhost:8998") parser.add_argument("--sglang_model_path", type=str, default="../model") parser.add_argument("--sglang_shared_path", type=str, default="./sglang_ckpt_search_grpo") add_model_profile_args(parser) add_deepspeed_args(parser) return parser.parse_args() if __name__ == "__main__": args = parse_args() model_profile = apply_model_profile(args) try: import deepspeed except ImportError as exc: raise RuntimeError("DeepSpeed未安装。服务器环境执行: pip install deepspeed") from exc local_rank = init_deepspeed_runtime(args, deepspeed) os.makedirs(args.save_dir, exist_ok=True) lm_config = build_lm_config( hidden_size=args.hidden_size, num_hidden_layers=args.num_hidden_layers, max_seq_len=args.max_seq_len + args.max_gen_len, use_moe=bool(args.use_moe), attention_type=resolve_attention_type(args), kv_lora_rank=args.kv_lora_rank, q_lora_rank=args.q_lora_rank, rope_dim=args.rope_dim, num_attention_heads=args.num_attention_heads, num_key_value_heads=args.num_key_value_heads, intermediate_size=args.intermediate_size, ) assert_no_training_output_collision( lm_config, weight=args.save_weight, save_dir=args.save_dir, checkpoint_dir="../checkpoints", from_resume=args.from_resume, overwrite=args.overwrite_output, ) model, tokenizer = init_model(lm_config, args.from_weight, device=args.device) if args.use_compile == 1: model = torch.compile(model) Logger("torch.compile enabled before DeepSpeed initialize") train_ds = SearchShortQARLDataset(args.data_path, tokenizer, max_length=args.max_seq_len) train_sampler = DistributedSampler(train_ds) if dist.is_initialized() else None loader_for_count = DataLoader(train_ds, batch_size=args.batch_size, sampler=train_sampler, collate_fn=collate_fn) iters = len(loader_for_count) total_optimizer_steps = max(1, math.ceil(iters / args.accumulation_steps) * args.epochs) ds_config = load_deepspeed_config(args, total_optimizer_steps) model_engine, _, _, _ = deepspeed.initialize( model=model, model_parameters=model.parameters(), config=ds_config, ) ckp_data = None if args.from_resume == 1: ckp_data = load_deepspeed_checkpoint( model_engine, lm_config, weight=args.save_weight, save_dir="../checkpoints", ) ref_model, _ = init_model(lm_config, args.from_weight, device=args.device) ref_model = ref_model.eval().requires_grad_(False) if torch.cuda.is_available() and args.dtype in {"float16", "bfloat16"}: ref_dtype = torch.float16 if args.dtype == "float16" else torch.bfloat16 ref_model = ref_model.to(dtype=ref_dtype) device_type = "cuda" if "cuda" in args.device else "cpu" amp_dtype = torch.bfloat16 if args.dtype == "bfloat16" else torch.float16 autocast_ctx = nullcontext() if device_type == "cpu" or args.dtype == "float32" else torch.cuda.amp.autocast(dtype=amp_dtype) rollout_engine = create_rollout_engine( engine_type=args.rollout_engine, policy_model=model_engine, tokenizer=tokenizer, device=args.device, autocast_ctx=autocast_ctx, sglang_base_url=args.sglang_base_url, sglang_model_path=args.sglang_model_path, sglang_shared_path=args.sglang_shared_path, ) rollout_engine.update_policy(model_engine) 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 wandb.init( project=args.wandb_project, name=f"SearchShortQA-GRPO-DS-{args.attention_type}-{args.hidden_size}", id=wandb_id, resume="must" if wandb_id else None, ) log_training_setup( args, lm_config, stage="search_shortqa_grpo_deepspeed", dataset_len=len(train_ds), iters=iters, tokens_per_sample=args.max_seq_len + args.max_gen_len, extra={ "model_profile": args.model_profile if model_profile else "", "deepspeed_config": args.deepspeed_config, "zero_stage": ds_config.get("zero_optimization", {}).get("stage", "none"), "num_generations": args.num_generations, "rollout_engine": args.rollout_engine, "beta": args.beta, "total_optimizer_steps": total_optimizer_steps, }, ) start_epoch = ckp_data.get("epoch", 0) if ckp_data else 0 start_step = ckp_data.get("step", 0) if ckp_data else 0 for epoch in range(start_epoch, args.epochs): args.current_epoch = epoch train_sampler and train_sampler.set_epoch(epoch) setup_seed(42 + epoch + local_rank) indices = torch.randperm(len(train_ds)).tolist() skip = start_step if (epoch == start_epoch and start_step > 0) else 0 batch_sampler = SkipBatchSampler(train_sampler or indices, args.batch_size, skip) loader = DataLoader( train_ds, batch_sampler=batch_sampler, num_workers=args.num_workers, pin_memory=True, collate_fn=collate_fn, ) if skip > 0: Logger(f"Epoch [{epoch + 1}/{args.epochs}]: 跳过前{start_step}个step，从step {start_step + 1}开始") train_epoch(args, lm_config, model_engine, tokenizer, loader, len(loader) + skip, rollout_engine, ref_model, autocast_ctx, skip, wandb) if dist.is_initialized(): dist.destroy_process_group()