update

train_embedding.py

@@ -0,0 +1,308 @@
+import os
+# 设置环境变量
+os.environ["WANDB_MODE"] = "offline"  # 或者使用 "dryrun"
+import platform
+import argparse
+import time
+import math
+import warnings
+import pandas as pd
+import torch
+import torch.distributed as dist
+from torch import optim, nn
+from torch.nn.parallel import DistributedDataParallel
+from torch.optim.lr_scheduler import CosineAnnealingLR
+from torch.utils.data import DataLoader, DistributedSampler
+from contextlib import nullcontext
+
+from transformers import AutoTokenizer
+
+# Removed: from model.model import MiniMindLM
+from model.LMConfig import LMConfig
+from model.dataset import PretrainDataset
+
+warnings.filterwarnings('ignore')
+
+
+# Define a simple model for pretraining embeddings
+class EmbeddingPretrainer(nn.Module):
+    def __init__(self, config: LMConfig):
+        super().__init__()
+        self.tok_embeddings = nn.Embedding(config.vocab_size, config.dim)
+        self.lm_head = nn.Linear(config.dim, config.vocab_size, bias=False)
+        # Tie weights (optional but common)
+        # self.tok_embeddings.weight = self.lm_head.weight
+
+    def forward(self, input_ids):
+        hidden_states = self.tok_embeddings(input_ids)
+        logits = self.lm_head(hidden_states)
+        return logits
+
+
+def Logger(content):
+    # 如果没有使用ddp或者ddp的主设备，那么就打印
+    if not ddp or dist.get_rank() == 0:
+        print(content)
+
+
+def get_lr(current_step, total_steps, lr):
+    # 更新学习率
+    # \text{get\_lr}(c, t, l) = \frac{l}{10} + 0.5 \cdot l \cdot \left(1 + \cos\left(\frac{\pi \cdot c}{t}\right)\right)
+    return lr / 10 + 0.5 * lr * (1 + math.cos(math.pi * current_step / total_steps))
+
+
+def train_epoch(epoch, wandb):
+    loss_fct = nn.CrossEntropyLoss(reduction='none', ignore_index=0) # Assuming 0 is pad_token_id
+    start_time = time.time()
+    for step, (X, Y, loss_mask) in enumerate(train_loader):
+        try:
+            # 将数据加载到设备上
+            X = X.to(args.device)
+            Y = Y.to(args.device)
+            loss_mask = loss_mask.to(args.device)
+
+            # 更新学习率
+            lr = get_lr(epoch * iter_per_epoch + step, args.epochs * iter_per_epoch, args.learning_rate)
+            for param_group in optimizer.param_groups:
+                param_group['lr'] = lr
+
+            with ctx:
+                logits = model(X) # Model returns logits directly
+                loss = loss_fct(
+                    logits.view(-1, logits.size(-1)),
+                    Y.view(-1)
+                ).view(Y.size())
+                loss = (loss * loss_mask).sum() / loss_mask.sum()
+                # Removed: loss += res.aux_loss 
+                loss = loss / args.accumulation_steps
+            
+            # Print data types for debugging
+            if step == 0 and (not ddp or dist.get_rank() == 0): # Print only for the first step of the first epoch on the main process
+                Logger("---- Data Type Check ----")
+                Logger(f"X.dtype: {X.dtype}")
+                if hasattr(model, 'module'): # DDP case
+                    Logger(f"Model parameter dtype: {next(model.module.parameters()).dtype}")
+                else: # Non-DDP case
+                    Logger(f"Model parameter dtype: {next(model.parameters()).dtype}")
+                Logger(f"logits.dtype: {logits.dtype}") # Changed from res.logits.dtype
+                Logger(f"loss.dtype: {loss.dtype}")
+                Logger("-------------------------")
+
+
+            scaler.scale(loss).backward()
+
+            if (step + 1) % args.accumulation_steps == 0:
+                scaler.unscale_(optimizer)
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
+
+                scaler.step(optimizer)
+                scaler.update()
+
+                optimizer.zero_grad(set_to_none=True)
+
+            # 打印日志
+            if step % args.log_interval == 0:
+                spend_time = time.time() - start_time
+                Logger(
+                    'Epoch:[{}/{}]({}/{}) loss:{:.3f} lr:{:.12f} epoch_Time:{}min:'.format(
+                        epoch + 1,
+                        args.epochs,
+                        step,
+                        iter_per_epoch,
+                        loss.item() * args.accumulation_steps,
+                        optimizer.param_groups[-1]['lr'],
+                        spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60))
+
+                if (wandb is not None) and (not ddp or dist.get_rank() == 0):
+                    wandb.log({"loss": loss.item() * args.accumulation_steps,
+                               "lr": optimizer.param_groups[-1]['lr'],
+                               "epoch_Time": spend_time / (step + 1) * iter_per_epoch // 60 - spend_time // 60})
+
+            # 保存模型
+            if (step + 1) % args.save_interval == 0 and (not ddp or dist.get_rank() == 0):
+                model.eval()
+                # Modified checkpoint path and content
+                ckp = f'{args.save_dir}/pretrained_embedding_dim{lm_config.dim}_vocab{lm_config.vocab_size}.pth'
+                
+                if isinstance(model, torch.nn.parallel.DistributedDataParallel):
+                    embedding_state_dict = model.module.tok_embeddings.state_dict()
+                else:
+                    embedding_state_dict = model.tok_embeddings.state_dict()
+
+                torch.save(embedding_state_dict, ckp)
+                Logger(f"Saved pretrained embedding to {ckp}")
+                model.train()
+
+        except Exception as e:
+            print(f"Error occurred: {str(e)}")
+            # Modified checkpoint path for error
+            save_path = f'{args.save_dir}/pretrained_embedding_dim{lm_config.dim}_vocab{lm_config.vocab_size}_ERROR.pth'
+            if  os.path.exists(save_path):
+                os.remove(save_path)
+            
+            if isinstance(model, torch.nn.parallel.DistributedDataParallel):
+                state_dict = model.module.tok_embeddings.state_dict()
+            else:
+                state_dict = model.tok_embeddings.state_dict()
+            torch.save(state_dict, save_path) # Save embedding state dict on error
+
+            for name, param in model.named_parameters():
+                if param.grad is not None and torch.isnan(param.grad).any():
+                    print(f"NaN gradient in parameter: {name}")
+            
+            for name, param in model.named_parameters():
+                if param.grad is not None and torch.isnan(param.grad).any():
+                    print(f"Parameter {name} values: {param.data}")
+                    print(f"Parameter {name} gradients: {param.grad}")
+            
+            raise ValueError("NaN gradient detected")
+
+
+def init_model(lm_config_params: LMConfig): # Renamed for clarity
+    # 加载tokenizer
+    tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
+    # Update vocab_size in lm_config if tokenizer has a different one
+    if tokenizer.vocab_size != lm_config_params.vocab_size:
+        Logger(f"Updating lm_config.vocab_size from {lm_config_params.vocab_size} to {tokenizer.vocab_size} based on tokenizer.")
+        lm_config_params.vocab_size = tokenizer.vocab_size
+
+    # 加载模型
+    model = EmbeddingPretrainer(lm_config_params).to(args.device) # Use EmbeddingPretrainer
+    # 打印模型参数
+    Logger(f'EmbeddingPretrainer total parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f} Million')
+    return model, tokenizer
+
+
+def init_distributed_mode():
+    if not ddp: return #如果没有启用分布式数据并行(DDP)，直接返回，不执行任何操作。
+    global ddp_local_rank, DEVICE #声明这两个变量为全局变量，以便在函数外部也能访问它们。
+
+    dist.init_process_group(backend="nccl") #初始化分布式进程组，使用NCCL后端（NVIDIA Collective Communications Library），这是NVIDIA GPU之间通信的优化库。
+    ddp_rank = int(os.environ["RANK"]) #从环境变量获取当前进程的全局编号。
+    ddp_local_rank = int(os.environ["LOCAL_RANK"]) #从环境变量获取当前进程的本地编号。
+    ddp_world_size = int(os.environ["WORLD_SIZE"]) #从环境变量获取当前进程组中的进程总数。
+    DEVICE = f"cuda:{ddp_local_rank}" #根据本地编号选择GPU设备。
+    torch.cuda.set_device(DEVICE) #设置当前进程的GPU设备。
+
+
+# torchrun --nproc_per_node 2 train_embedding.py
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="MiniMind Embedding Pretraining") # Changed description
+    parser.add_argument("--out_dir", type=str, default="out_embedding") # Changed default out_dir
+    # 若要以最快速度实现zero则epochs设置为1轮；否则应当利用有限的数据训练2~6个epochs。
+    parser.add_argument("--epochs", type=int, default=3)
+    parser.add_argument("--batch_size", type=int, default=32) # Smaller batch size might be needed if memory is an issue
+    parser.add_argument("--learning_rate", type=float, default=5e-4)
+    parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu") #如果GPU可用，则使用GPU，否则使用CPU。
+    parser.add_argument("--dtype", type=str, default="bfloat16")
+    parser.add_argument("--use_wandb", default=False, action="store_true")
+    parser.add_argument("--wandb_project", type=str, default="MiniMind-Embedding-Pretrain") # Changed project name
+    parser.add_argument("--num_workers", type=int, default=8)
+    parser.add_argument("--ddp", action="store_true")
+    parser.add_argument("--accumulation_steps", type=int, default=8) #梯度累积步数，用于控制梯度更新频率。
+    parser.add_argument("--grad_clip", type=float, default=1.0) #梯度裁剪阈值，用于防止梯度爆炸。
+    # parser.add_argument("--warmup_iters", type=int, default=0) #预热迭代次数，用于控制学习率预热过程。 (Can be kept or removed)
+    parser.add_argument("--log_interval", type=int, default=100) #日志打印间隔，用于控制日志打印的频率。
+    parser.add_argument("--save_interval", type=int, default=100) #模型保存间隔，用于控制模型保存的频率。
+    parser.add_argument('--local_rank', type=int, default=-1) #本地进程编号，用于分布式训练。
+    parser.add_argument('--dim', default=768, type=int) #模型维度，用于控制模型的大小。
+    # Removed n_layers, use_moe as they are not relevant for EmbeddingPretrainer
+    # parser.add_argument('--n_layers', default=8, type=int) 
+    parser.add_argument('--max_seq_len', default=512, type=int) #最大序列长度，用于控制输入序列的最大长度。
+    # parser.add_argument('--use_moe', default=False, type=bool) 
+    parser.add_argument("--data_path", type=str, default="./dataset/pretrain_hq.jsonl") #数据路径，用于控制数据集的路径。
+    # Add vocab_size to args, though it will be overridden by tokenizer if different
+    parser.add_argument('--vocab_size', default=6400, type=int)
+
+
+    args = parser.parse_args()
+
+    # Create LMConfig with relevant parameters for embedding
+    lm_config = LMConfig(
+        dim=args.dim, 
+        vocab_size=args.vocab_size, # Will be updated by tokenizer
+        max_seq_len=args.max_seq_len,
+        # n_layers, n_heads, etc. are not directly used by EmbeddingPretrainer but LMConfig requires them
+        # We can set them to default or minimal values if they cause issues, or modify LMConfig
+        # For now, using defaults from LMConfig definition for unneeded params.
+        n_layers=1, # Minimal
+        n_heads=1, # Minimal
+        n_kv_heads=1 #Minimal
+    ) 
+    args.save_dir = os.path.join(args.out_dir) #创建保存目录。
+    os.makedirs(args.save_dir, exist_ok=True) #创建保存目录。
+    os.makedirs(args.out_dir, exist_ok=True) #创建输出目录。
+    tokens_per_iter = args.batch_size * lm_config.max_seq_len #计算每个迭代步骤的token数量。
+    print(f"tokens_per_iter: {tokens_per_iter}")
+    device_type = "cuda" if "cuda" in args.device else "cpu" #确定设备类型。
+
+    # Determine the torch dtype
+    pt_dtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[args.dtype]
+
+    args.wandb_run_name = f"MiniMind-Embedding-Pretrain-Dim-{args.dim}-Vocab-{lm_config.vocab_size}" # Updated run name
+
+    ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast(dtype=pt_dtype)
+
+    ddp = int(os.environ.get("RANK", -1)) != -1  # is this a ddp run?
+    ddp_local_rank, DEVICE = 0, "cuda:0" # Default values, will be overwritten in DDP
+
+    base_seed = 1337
+    torch.manual_seed(base_seed)
+    torch.cuda.manual_seed(base_seed)
+
+    if ddp:
+        init_distributed_mode() # This sets DEVICE and ddp_local_rank
+        args.device = torch.device(DEVICE) # Ensure args.device is updated
+        rank = dist.get_rank()
+        torch.manual_seed(base_seed + rank)
+        # 同时设置 CUDA 的随机种子
+        torch.cuda.manual_seed_all(base_seed + rank) # Use seed_all for DDP
+
+    if args.use_wandb and (not ddp or dist.get_rank() == 0): # Check rank for DDP wandb init
+        import wandb
+
+        wandb.init(project=args.wandb_project, name=args.wandb_run_name, config=args)
+    else:
+        wandb = None
+
+    model, tokenizer = init_model(lm_config) # Pass the lm_config instance
+    
+    # Update lm_config vocab_size again after tokenizer to ensure consistency for save path name
+    if lm_config.vocab_size != tokenizer.vocab_size:
+        lm_config.vocab_size = tokenizer.vocab_size
+        args.wandb_run_name = f"MiniMind-Embedding-Pretrain-Dim-{args.dim}-Vocab-{lm_config.vocab_size}"
+        if wandb is not None and (not ddp or dist.get_rank() == 0):
+            wandb.config.update({'vocab_size': lm_config.vocab_size, 'wandb_run_name': args.wandb_run_name}, allow_val_change=True)
+
+
+    train_ds = PretrainDataset(args.data_path, tokenizer, max_length=lm_config.max_seq_len)
+    train_sampler = DistributedSampler(train_ds, shuffle=True, seed=base_seed) if ddp else None # Added shuffle and seed
+    train_loader = DataLoader(
+        train_ds,
+        batch_size=args.batch_size,
+        pin_memory=True,
+        drop_last=True, # Set to True for more stable training step counts
+        shuffle=(train_sampler is None), # Shuffle only if not using DDP sampler
+        num_workers=args.num_workers,
+        sampler=train_sampler
+    )
+
+    scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16'])) # bfloat16 also uses scaler
+    optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)
+
+    if ddp:
+        # model._ddp_params_and_buffers_to_ignore = {"pos_cis"} # Not relevant for EmbeddingPretrainer
+        model = DistributedDataParallel(model, device_ids=[ddp_local_rank])
+        
+    # torch.autograd.set_detect_anomaly(True) # Can be enabled for debugging
+    iter_per_epoch = len(train_loader)
+    Logger(f"Starting training for {args.epochs} epochs with {iter_per_epoch} iterations per epoch.")
+    for epoch in range(args.epochs):
+        if ddp:
+            train_sampler.set_epoch(epoch) # Important for DDP shuffling
+        train_epoch(epoch, wandb)
+
+    if wandb is not None and (not ddp or dist.get_rank() == 0) :
+        wandb.finish()
+    
+    Logger("Embedding pretraining finished.") 

train_pretrain.py

@@ -14,6 +14,7 @@ from torch.nn.parallel import DistributedDataParallel
 from torch.optim.lr_scheduler import CosineAnnealingLR
 from torch.utils.data import DataLoader, DistributedSampler
 from contextlib import nullcontext
+from typing import Optional
 
 from transformers import AutoTokenizer
 
@@ -61,6 +62,18 @@ def train_epoch(epoch, wandb):
                 loss += res.aux_loss 
                 loss = loss / args.accumulation_steps
 
+            # Print data types for debugging
+            if step == 0 and (not ddp or dist.get_rank() == 0): # Print only for the first step of the first epoch on the main process
+                Logger("---- Data Type Check ----")
+                Logger(f"X.dtype: {X.dtype}")
+                if hasattr(model, 'module'): # DDP case
+                    Logger(f"Model parameter dtype: {next(model.module.parameters()).dtype}")
+                else: # Non-DDP case
+                    Logger(f"Model parameter dtype: {next(model.parameters()).dtype}")
+                Logger(f"res.logits.dtype: {res.logits.dtype}")
+                Logger(f"loss.dtype: {loss.dtype}")
+                Logger("-------------------------")
+
             scaler.scale(loss).backward()
 
             if (step + 1) % args.accumulation_steps == 0:
@@ -106,6 +119,16 @@ def train_epoch(epoch, wandb):
 
         except Exception as e:
             print(f"Error occurred: {str(e)}")
+            save_path = f'{args.save_dir}/pretrain_{lm_config.dim}{moe_path}_nanERROR.pth'
+            if  os.path.exists(save_path):
+                os.remove(save_path)
+            
+            if isinstance(model, torch.nn.parallel.DistributedDataParallel):
+                state_dict = model.module.state_dict() 
+            else:
+                state_dict = model.state_dict()
+            torch.save(state_dict, save_path)
+
             for name, param in model.named_parameters():
                 if param.grad is not None and torch.isnan(param.grad).any():
                     print(f"NaN gradient in parameter: {name}")
@@ -118,11 +141,21 @@ def train_epoch(epoch, wandb):
             raise ValueError("NaN gradient detected")
 
 
-def init_model(lm_config):
+def init_model(lm_config, pretrained_embedding_path: Optional[str] = None):
     # 加载tokenizer
     tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer')
     # 加载模型
     model = MiniMindLM(lm_config).to(args.device)
+
+    # Load pretrained token embeddings if path is provided
+    if pretrained_embedding_path and os.path.exists(pretrained_embedding_path):
+        Logger(f"Loading pretrained token embeddings from {pretrained_embedding_path}")
+        embedding_weights = torch.load(pretrained_embedding_path, map_location=args.device)
+        model.tok_embeddings.load_state_dict(embedding_weights)
+        Logger("Successfully loaded pretrained token embeddings.")
+    elif pretrained_embedding_path:
+        Logger(f"Warning: Pretrained embedding path {pretrained_embedding_path} provided but file does not exist. Initializing embeddings from scratch.")
+
     # 打印模型参数
     Logger(f'LLM总参数量：{sum(p.numel() for p in model.parameters() if p.requires_grad) / 1e6:.3f} 百万')
     return model, tokenizer
@@ -165,6 +198,7 @@ if __name__ == "__main__":
     parser.add_argument('--max_seq_len', default=512, type=int) #最大序列长度，用于控制输入序列的最大长度。
     parser.add_argument('--use_moe', default=False, type=bool) #是否使用MOE，用于控制是否使用MOE。
     parser.add_argument("--data_path", type=str, default="./dataset/pretrain_hq.jsonl") #数据路径，用于控制数据集的路径。
+    parser.add_argument("--pretrained_embedding_path", type=str, default=None, help="Path to pretrained token embedding weights (.pth file)")
     args = parser.parse_args()
 
     lm_config = LMConfig(dim=args.dim, n_layers=args.n_layers, max_seq_len=args.max_seq_len, use_moe=args.use_moe) #创建LMConfig对象，用于控制模型配置。
@@ -175,9 +209,12 @@ if __name__ == "__main__":
     print(f"tokens_per_iter: {tokens_per_iter}")
     device_type = "cuda" if "cuda" in args.device else "cpu" #确定设备类型。
 
+    # Determine the torch dtype
+    pt_dtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[args.dtype]
+
     args.wandb_run_name = f"MiniMind-Pretrain-Epoch-{args.epochs}-BatchSize-{args.batch_size}-LearningRate-{args.learning_rate}"
 
-    ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast()
+    ctx = nullcontext() if device_type == "cpu" else torch.cuda.amp.autocast(dtype=pt_dtype)
 
     ddp = int(os.environ.get("RANK", -1)) != -1  # is this a ddp run?
     ddp_local_rank, DEVICE = 0, "cuda:0"
@@ -201,7 +238,7 @@ if __name__ == "__main__":
     else:
         wandb = None
 
-    model, tokenizer = init_model(lm_config)
+    model, tokenizer = init_model(lm_config, args.pretrained_embedding_path)
     train_ds = PretrainDataset(args.data_path, tokenizer, max_length=lm_config.max_seq_len)
     train_sampler = DistributedSampler(train_ds) if ddp else None
     train_loader = DataLoader(
@@ -214,7 +251,7 @@ if __name__ == "__main__":
         sampler=train_sampler
     )
 
-    scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16', 'bfloat16']))
+    scaler = torch.cuda.amp.GradScaler(enabled=(args.dtype in ['float16']))
     optimizer = optim.AdamW(model.parameters(), lr=args.learning_rate)
 
     if ddp: