import json import pdb from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, OlmoeForCausalLM, OlmoeModel from transformers.models.auto.modeling_auto import MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES import copy from transformers.modeling_outputs import ( MoeCausalLMOutputWithPast, MoeModelOutputWithPast, ) from collections import defaultdict import numpy as np import math from torch import nn import pandas as pd from transformers.cache_utils import Cache, DynamicCache, StaticCache from dataclasses import dataclass # from transformers.models.olmoe.configuration_olmoe import OlmoeConfig # from transformers.models.olmoe.modeling_olmoe import OlmoeMLP, OlmoeAttention, OlmoeFlashAttention2, OlmoeSdpaAttention, OlmoeRMSNorm, OlmoeSparseMoeBlock, apply_rotary_pos_emb, repeat_kv, OlmoeRotaryEmbedding import os import sys import torch.distributed as dist from tqdm import tqdm from torch.utils.data import DataLoader from torch.utils.data.distributed import DistributedSampler import transformers import pickle # from transformers.models.olmoe.configuration_olmoe import OlmoeConfig from dataset import * # from utils import flash_attn_forward, flash_attn_prepare_decoder_attention_mask, get_multiround_data from peft import (get_peft_model, PeftModel) import random # from config import * from datasets import Dataset, DatasetDict, load_dataset import wandb import gc import os import argparse import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import functools from torch.optim.lr_scheduler import StepLR import torch.nn.functional as F import torch.distributed as dist import torch.multiprocessing as mp from torch.nn.parallel import DistributedDataParallel as DDP from torch.utils.data.distributed import DistributedSampler from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import ( checkpoint_wrapper, CheckpointImpl) from torch.distributed.fsdp import ( FullyShardedDataParallel as FSDP, MixedPrecision, BackwardPrefetch, ShardingStrategy, FullStateDictConfig, StateDictType, ) from torch.distributed.fsdp.wrap import ( transformer_auto_wrap_policy, enable_wrap, wrap, ) from functools import partial from torch.utils.data import DataLoader from pathlib import Path from typing import Type, List, Optional, Tuple, Union from modelforseminat_v4 import * from torch.optim.lr_scheduler import _LRScheduler class WarmupCosineScheduler(_LRScheduler): def __init__(self, optimizer, warmup_steps, total_steps, min_lr=0.0, last_epoch=-1): # self.warmup_steps = warmup_steps self.total_steps = total_steps self.min_lr = min_lr if isinstance(warmup_steps, float) and 0 < warmup_steps < 1: self.warmup_steps = int(warmup_steps * total_steps) else: self.warmup_steps = int(warmup_steps) super().__init__(optimizer, last_epoch) def get_lr(self): step = self.last_epoch + 1 lrs = [] for base_lr in self.base_lrs: if step < self.warmup_steps: # Linear warmup lr = base_lr * step / self.warmup_steps else: # Cosine decay progress = (step - self.warmup_steps) / max( 1, self.total_steps - self.warmup_steps) cosine_decay = 0.5 * (1 + math.cos(math.pi * progress)) lr = self.min_lr + (base_lr - self.min_lr) * cosine_decay lrs.append(lr) return lrs ################################# FSDP Config ##################################### def setup(): # initialize the process group local_rank = int(os.environ['LOCAL_RANK']) torch.cuda.set_device(local_rank) dist.init_process_group( backend='nccl', init_method='env://', ) def cleanup(): gc.collect() torch.cuda.empty_cache() dist.destroy_process_group() def get_fsdp_device(): # 每个进程初始化分布式环境后调用 local_rank = int(os.environ.get("LOCAL_RANK", 0)) # torchrun 自动设置 device = torch.device(f"cuda:{local_rank}") torch.cuda.set_device(device) return device def load_trained_model(model_name): DEVICE = "cuda" if torch.cuda.is_available() else "cpu" olmo_path = "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat_backup/model/OLMo-2-0425-1B" pt_path = "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat/ckp/sft-v4-1e3-len2-fc-chunklimit2-jueduipos/sft-v4-1e3-len2-fc-chunklimit2-jueduipos-steps_300.pt" config_path = "/AIRvePFS/ai4science/users/ai4science/users/zyk/seminat_backup/model/OLMo-2-0425-1B/config.json" config = AutoConfig.from_pretrained(olmo_path) model = Olmo2ForCausalLMForSemiNAT.from_pretrained(olmo_path, config=config, torch_dtype=torch.bfloat16) state_dict = torch.load(pt_path, map_location=DEVICE, weights_only=True) missing_keys, unexpected_keys = model.load_state_dict(state_dict, strict=False) print( f"Loaded with {len(missing_keys)} missing keys and {len(unexpected_keys)} unexpected keys." ) if missing_keys: print("Missing keys:", missing_keys) if unexpected_keys: print("Unexpected keys:", unexpected_keys) model = model.to(DEVICE) tokenizer = AutoTokenizer.from_pretrained(olmo_path) return model, tokenizer def setup_model(model_name): model = Olmo2ForCausalLMForSemiNAT.from_pretrained(model_name) tokenizer = AutoTokenizer.from_pretrained(model_name) # config = AutoConfig.from_pretrained(model_name) # model = Olmo2ForCausalLMForSemiNAT(config) # 注意这里不用 from_pretrained # tokenizer = AutoTokenizer.from_pretrained(model_name) return model, tokenizer def fsdp_main(args): local_rank = int(os.environ['LOCAL_RANK']) rank = int(os.environ['RANK']) world_size = int(os.environ['WORLD_SIZE']) if args.use_wandb and rank == 0: wandb.init(entity="SemiNAT", project="SemiNAT-SFT", name=args.run_name) model, tokenizer = setup_model(args.model_path) # model, tokenizer = load_trained_model(args.model_path) model.config.chunk_size_limit = args.chunk_size_limit # pdb.set_trace() if ".pkl" in args.data_path: train_dataset = pickle.load(open(args.data_path, "rb")) else: datasets = pd.read_parquet(args.data_path) train_dataset = eval(f"{args.data_type}")( tokenizer, datasets, # your data preprocessing function args.max_length, # your max input length args.data_processess_num) train_sampler = DistributedSampler(train_dataset, rank=rank, num_replicas=world_size, shuffle=True) train_dataloader = DataLoader(dataset=train_dataset, sampler=train_sampler, batch_size=args.batch_size) print(f"Size of train dataset: {len(train_dataset)}") setup() Olmo2DecoderLayerForSemiNAT_auto_wrap_policy = functools.partial( transformer_auto_wrap_policy, transformer_layer_cls={ Olmo2DecoderLayerForSemiNAT, NATDecoderForSemiNAT, }) sharding_strategy: ShardingStrategy = ShardingStrategy.FULL_SHARD #for Zero2 and FULL_SHARD for Zero3 torch.cuda.set_device(local_rank) bfSixteen = MixedPrecision( param_dtype=torch.bfloat16, reduce_dtype=torch.bfloat16, buffer_dtype=torch.bfloat16, ) # if bf16_ready: mp_policy = bfSixteen # else: # mp_policy = None # defaults to fp32 # if args.use_lora: # model = get_peft_model(model, lora_config) # model is on CPU before input to FSDP model = FSDP(model, auto_wrap_policy=Olmo2DecoderLayerForSemiNAT_auto_wrap_policy, mixed_precision=mp_policy, sharding_strategy=sharding_strategy, device_id=torch.cuda.current_device(), use_orig_params=True) optimizer = optim.AdamW( model.parameters(), lr=args.lr, betas=args.betas, weight_decay=args.weight_decay, eps=args.eps, ) # scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma) scheduler = WarmupCosineScheduler( optimizer=optimizer, # 优化器对象 warmup_steps=args.warmup_steps, # warmup 步数（或比例） total_steps=args.total_steps, # 总训练步数 min_lr=args.min_lr # 最小学习率 ) torch.autograd.set_detect_anomaly(True) loss1_list = [] loss2_list = [] loss_list = [] global_step = 0 for epoch in range(1, args.epochs + 1): # t0 = time.time() model.train() local_rank = int(os.environ['LOCAL_RANK']) # fsdp_loss = torch.zeros(2).to(local_rank) if train_sampler: train_sampler.set_epoch(epoch) if rank == 0: inner_pbar = tqdm(range(len(train_dataloader)), colour="blue", desc="r0 Training Epoch") for batch in train_dataloader: optimizer.zero_grad() loss1, loss2 = model(input_ids=batch[0], labels=batch[1], attention_mask=batch[2], slice_pos=batch[3], use_cache=False).loss loss = loss1 + loss2 # loss = loss2 loss1_list.append(loss1.item()) loss2_list.append(loss2.item()) loss_list.append(loss.item()) # pdb.set_trace() # if torch.isnan(loss): # print(f"Step {global_step}: loss is NaN, entering pdb …") # pdb.set_trace() # print(f"loss1:{loss1},loss2:{loss2}") loss.backward() # 按参数计算 for name, module in model.named_modules(): total_norm = 0.0 param_count = 0 for param in module.parameters(recurse=False): if param.grad is not None: total_norm += param.grad.data.norm(2).item()**2 param_count += 1 if param_count > 0: total_norm = total_norm**0.5 wandb.log({f"grad_norm/{name}": total_norm}, step=global_step) # 按层计算 # layer_grads = defaultdict(list) # for name, module in model.named_modules(): # # 只处理 encoder.layers.N 这类结构（按需改写匹配条件） # # pdb.set_trace() # if "model.layers." in name or "model.decoder" in name or "model.encoder" in name: # if "model.layers" in name: # # 提取层号（如 encoder.layers.0 → layer0） # parts = name.split(".") # try: # layer_idx = int(parts[3]) # except (IndexError, ValueError): # pdb.set_trace() # # continue # layer_key = f"layer{layer_idx}" # else: # layer_key = "decoder" if "model.decoder" in name else "encoder" # # 收集该模块下的参数梯度（只自身，不递归） # for param in module.parameters(recurse=False): # if param.grad is not None: # layer_grads[layer_key].append( # param.grad.detach().flatten()) # if "lm_head" in name: # layer_key = "lm_head" # layer_grads[layer_key].append( # param.grad.detach().flatten()) # # 合并每层所有子模块的梯度，并 log 总体 L2 范数 # for layer_key, grads in layer_grads.items(): # if grads: # total_grad = torch.cat(grads).norm(2).item() # wandb.log({f"grad_norm/{layer_key}": total_grad}, # step=global_step) optimizer.step() global_step += 1 if global_step % args.save_steps == 0: save_policy = FullStateDictConfig(offload_to_cpu=True, rank0_only=True) with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, save_policy): cpu_state = model.state_dict() if rank == 0: print(f"--> steps: {str(global_step)} saving model ...") if not os.path.exists(args.save_path): os.makedirs(args.save_path) save_name = f"{args.save_name}-steps_{str(global_step)}.pt" print(f"--> saving as model name {save_name}") save_path = os.path.join(args.save_path, save_name) torch.save(cpu_state, save_path) if rank == 0: inner_pbar.update(1) if args.use_wandb: wandb.log({ "length prediction loss": sum(loss1_list[-20:]) / len(loss1_list[-20:]), "nat loss": sum(loss2_list[-20:]) / len(loss2_list[-20:]), "loss": sum(loss_list[-20:]) / len(loss_list[-20:]) }) dist.all_reduce(loss, op=dist.ReduceOp.SUM) if rank == 0: inner_pbar.close() scheduler.step() # if rank == 0: # print(f"--> entering save model state") # save_policy = FullStateDictConfig(offload_to_cpu=True, rank0_only=True) # with FSDP.state_dict_type(model, StateDictType.FULL_STATE_DICT, # save_policy): # cpu_state = model.state_dict() # if rank == 0: # print(f"--> epoch: {str(epoch)} saving model ...") # if not os.path.exists(args.save_path): # os.makedirs(args.save_path) # save_name = f"{args.save_name}-epoch_{str(epoch)}.pt" # print(f"--> saving as model name {save_name}") # save_path = os.path.join(args.save_path, save_name) # torch.save(cpu_state, save_path) dist.barrier() cleanup() ################################# FSDP Config ##################################### if __name__ == "__main__": # Training settings parser = argparse.ArgumentParser() parser.add_argument('--batch-size', type=int, default=4, metavar='N', help='input batch size for training (default: 64)') parser.add_argument('--model_path', type=str) parser.add_argument('--save_path', type=str) parser.add_argument('--save_name', type=str) parser.add_argument('--data_path', type=str) parser.add_argument('--data_type', type=str) parser.add_argument('--run_name', type=str) parser.add_argument('--max_length', type=int) parser.add_argument('--chunk_size_limit', type=int) parser.add_argument('--save_steps', type=int, default=5000) parser.add_argument('--data_processess_num', type=int, default=8) parser.add_argument('--epochs', type=int, default=2, metavar='N', help='number of epochs to train (default: 3)') parser.add_argument('--lr', type=float, default=.002, metavar='LR', help='learning rate (default: .002)') parser.add_argument('--gamma', type=float, default=0.7, metavar='M', help='Learning rate step gamma (default: 0.7)') parser.add_argument('--weight_decay', type=float) parser.add_argument('--eps', type=float) parser.add_argument('--decay_norm_and_bias', type=bool) parser.add_argument( "--betas", type=float, nargs=2, # 表示需要两个 float 值 default=[0.9, 0.999] # 可选默认值 ) parser.add_argument('--decay_embeddings', type=bool) parser.add_argument('--warmup_steps', type=float) parser.add_argument('--total_steps', type=int) parser.add_argument('--min_lr', type=float) parser.add_argument('--seed', type=int, default=1, metavar='S', help='random seed (default: 1)') parser.add_argument('--use_lora', action='store_true', default=False) parser.add_argument("--use_wandb", action="store_true", help="whether to use wandb") args = parser.parse_args() torch.manual_seed(args.seed) fsdp_main(args)