update

model/model.py

@@ -2,6 +2,7 @@ import math
 import struct
 import inspect
 import time
+import gc
 #子空间二维分解+梯度更新
 from .LMConfig import LMConfig
 from typing import Any, Optional, Tuple, List, Union
@@ -92,6 +93,15 @@ class KnowledgeDataset(nn.Module):
         device = all_scores.device
         dtype = all_scores.dtype
 
+        # 记录进入智能选择前的内存状态
+        if hasattr(self, 'step_counter'):
+            self.step_counter += 1
+            # 禁用GPU内存监控记录以提高性能
+            # if self.step_counter % 50 == 0:  # 每50次调用记录一次
+            #     if torch.cuda.is_available():
+            #         allocated_before = torch.cuda.memory_allocated() / (1024**3)
+            #         print(f"[INTEL_SELECT_ENTER] Step {self.step_counter}: GPU Memory: {allocated_before:.2f}GB")
+
         # 对每个batch进行分层选择
         enhanced_scores = all_scores.clone()
         query_features = query.mean(dim=1)  # [batch_size, dim]
@@ -157,6 +167,24 @@ class KnowledgeDataset(nn.Module):
         all_best_tokens = torch.stack(batch_best_tokens, dim=0)
         all_best_tokens_embeddings = torch.stack(batch_best_tokens_embeddings, dim=0)
 
+        # 清理中间张量以防止内存泄漏
+        del all_candidate_indices, unique_indices, inverse_indices
+        del unique_candidate_features, normalized_candidates, normalized_queries
+        del batch_best_tokens, batch_best_tokens_embeddings
+        del flat_tokens, flat_embeddings, pre_update_embeddings
+        
+        # 记录退出智能选择后的内存状态（已禁用以提高性能）
+        # if hasattr(self, 'step_counter') and self.step_counter % 50 == 0:
+        #     if torch.cuda.is_available():
+        #         allocated_after = torch.cuda.memory_allocated() / (1024**3)
+        #         print(f"[INTEL_SELECT_EXIT] Step {self.step_counter}: GPU Memory: {allocated_after:.2f}GB")
+                
+        # 强制垃圾回收（仅在监控步骤）
+        if hasattr(self, 'step_counter') and self.step_counter % 100 == 0:
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
         return all_best_tokens, all_best_tokens_embeddings
     
         
@@ -216,6 +244,16 @@ class CrossAttention(nn.Module):
 
     def forward(self, x, db, context_mask=None, pos_emb=None):
         batch_size = x.size(0)
+        
+        # 监控交叉注意力开始时的内存（已禁用以提高性能）
+        if not hasattr(self, 'call_counter'):
+            self.call_counter = 0
+        self.call_counter += 1
+        
+        # 禁用GPU内存监控记录以提高性能
+        # if self.call_counter % 100 == 0 and torch.cuda.is_available():
+        #     allocated_before = torch.cuda.memory_allocated() / (1024**3)
+        #     print(f"[CROSS_ATTN_ENTER] Call {self.call_counter}: GPU Memory: {allocated_before:.2f}GB")
 
         # 分离多头
         q = self.to_q(x).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
@@ -241,6 +279,14 @@ class CrossAttention(nn.Module):
         context = context.transpose(1, 2).contiguous().view(batch_size, -1, self.config.dim)
 
         context = self.to_out(context)
+        
+        # 清理中间张量
+        del q, k, v, attn_scores, attn_weights
+        
+        # 监控交叉注意力结束时的内存（已禁用以提高性能）
+        # if self.call_counter % 100 == 0 and torch.cuda.is_available():
+        #     allocated_after = torch.cuda.memory_allocated() / (1024**3)
+        #     print(f"[CROSS_ATTN_EXIT] Call {self.call_counter}: GPU Memory: {allocated_after:.2f}GB")
 
         return context
 

train_pretrain_accelerate.py

@@ -22,6 +22,8 @@ from transformers import AutoTokenizer, get_cosine_schedule_with_warmup
 import numpy as np
 from sklearn.metrics.pairwise import cosine_similarity
 import swanlab  # 替换wandb导入
+import gc  # 添加垃圾回收模块
+import psutil  # 添加系统资源监控模块
 
 from model.model import MiniMindLM, RMSNorm
 from model.LMConfig import LMConfig
@@ -29,6 +31,63 @@ from model.dataset import PretrainDataset
 
 warnings.filterwarnings('ignore')
 
+# 内存监控辅助函数
+def get_memory_usage():
+    """获取当前内存使用情况"""
+    process = psutil.Process()
+    memory_info = process.memory_info()
+    return {
+        'rss_mb': memory_info.rss / 1024 / 1024,  # 物理内存使用量（MB）
+        'vms_mb': memory_info.vms / 1024 / 1024,  # 虚拟内存使用量（MB）
+    }
+
+def get_cuda_memory_usage():
+    """获取CUDA内存使用情况"""
+    if torch.cuda.is_available():
+        return {
+            'cuda_allocated_mb': torch.cuda.memory_allocated() / 1024 / 1024,
+            'cuda_reserved_mb': torch.cuda.memory_reserved() / 1024 / 1024,
+            'cuda_max_allocated_mb': torch.cuda.max_memory_allocated() / 1024 / 1024,
+        }
+    return {}
+
+def get_tensor_memory_size(tensor_list):
+    """计算tensor列表的总内存占用（MB）"""
+    total_size = 0
+    for batch in tensor_list:
+        if isinstance(batch, (list, tuple)):
+            for tensor in batch:
+                if isinstance(tensor, torch.Tensor):
+                    total_size += tensor.numel() * tensor.element_size()
+        elif isinstance(batch, torch.Tensor):
+            total_size += batch.numel() * batch.element_size()
+    return total_size / 1024 / 1024  # 转换为MB
+
+def log_memory_status(step, prefetch_batches, accelerator, stage="", detailed=False):
+    """记录内存状态"""
+    if not accelerator.is_main_process:
+        return
+        
+    memory_info = get_memory_usage()
+    cuda_info = get_cuda_memory_usage()
+    prefetch_memory = get_tensor_memory_size(prefetch_batches)
+    
+    log_msg = f"[Memory Monitor] Step {step} {stage} - "
+    log_msg += f"Prefetch batches: {len(prefetch_batches)}, "
+    log_msg += f"Prefetch memory: {prefetch_memory:.2f}MB, "
+    log_msg += f"System RSS: {memory_info['rss_mb']:.2f}MB"
+    
+    if cuda_info:
+        log_msg += f", CUDA allocated: {cuda_info['cuda_allocated_mb']:.2f}MB"
+        log_msg += f", CUDA reserved: {cuda_info['cuda_reserved_mb']:.2f}MB"
+    
+    if detailed:
+        log_msg += f", System VMS: {memory_info['vms_mb']:.2f}MB"
+        if cuda_info:
+            log_msg += f", CUDA max allocated: {cuda_info['cuda_max_allocated_mb']:.2f}MB"
+    
+    Logger(log_msg, accelerator)
+
 # 日志记录函数
 def Logger(msg, accelerator=None):
     # 如果没有提供accelerator，则只在主进程打印
@@ -227,65 +286,92 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
     moe_path = '_moe' if args.use_moe else ''
     best_loss = float('10000')
 
+    # 初始化CUDA事件变量
+    data_start = data_end = forward_start = forward_end = None
+    backward_start = backward_end = optimizer_start = optimizer_end = None
+    
     # 添加CUDA事件来分析性能 (只在主进程进行)
-    # if args.profile and accelerator.is_main_process:
+    if args.profile and accelerator.is_main_process:
-    #     data_start = torch.cuda.Event(enable_timing=True)
+        data_start = torch.cuda.Event(enable_timing=True)
-    #     data_end = torch.cuda.Event(enable_timing=True)
+        data_end = torch.cuda.Event(enable_timing=True)
-    #     forward_start = torch.cuda.Event(enable_timing=True)
+        forward_start = torch.cuda.Event(enable_timing=True)
-    #     forward_end = torch.cuda.Event(enable_timing=True)
+        forward_end = torch.cuda.Event(enable_timing=True)
-    #     backward_start = torch.cuda.Event(enable_timing=True)
+        backward_start = torch.cuda.Event(enable_timing=True)
-    #     backward_end = torch.cuda.Event(enable_timing=True)
+        backward_end = torch.cuda.Event(enable_timing=True)
-    #     optimizer_start = torch.cuda.Event(enable_timing=True)
+        optimizer_start = torch.cuda.Event(enable_timing=True)
-    #     optimizer_end = torch.cuda.Event(enable_timing=True)
+        optimizer_end = torch.cuda.Event(enable_timing=True)
 
     # 预取数据
     prefetch_factor = 2  # 预取的批次数
     data_iter = iter(train_loader)
     prefetch_batches = []
 
+    # 记录初始内存状态
+    if args.memory_monitor:
+        log_memory_status(-1, prefetch_batches, accelerator, "before_prefetch", detailed=True)
+
     # 预取初始批次
-    for _ in range(min(prefetch_factor, len(train_loader))):
+    for i in range(min(prefetch_factor, len(train_loader))):
         try:
             batch = next(data_iter)
             prefetch_batches.append(batch)
+            # 每次添加batch后记录内存变化
+            if args.memory_monitor and accelerator.is_main_process:
+                log_memory_status(-1, prefetch_batches, accelerator, f"after_adding_batch_{i+1}")
         except StopIteration:
             break
 
+    # 记录预取完成后的内存状态
+    if args.memory_monitor:
+        log_memory_status(-1, prefetch_batches, accelerator, "after_initial_prefetch", detailed=True)
+
     # 在开始循环前初始化日志记录所需变量
     last_log_time = epoch_start_time
 
     for step in range(total_steps_in_epoch):
         try:
             # 计时数据加载 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and data_start is not None:
-            #     data_start.record()
+                data_start.record()
+
+            # 记录使用batch前的内存状态（根据配置间隔记录详细信息）
+            if args.memory_monitor and step % args.memory_monitor_interval == 0:
+                log_memory_status(step, prefetch_batches, accelerator, "before_use_batch", detailed=True)
 
             # 使用预取的数据
             if prefetch_batches:
                 X, Y, loss_mask = prefetch_batches.pop(0)
+                # 记录使用batch后的内存变化
+                if args.memory_monitor and step % args.memory_monitor_interval == 0:
+                    log_memory_status(step, prefetch_batches, accelerator, "after_pop_batch")
             else:
                 # 如果预取队列为空，直接加载
                 X, Y, loss_mask = next(data_iter)
+                if args.memory_monitor and accelerator.is_main_process:
+                    Logger(f"[Memory Monitor] Step {step} - Prefetch queue empty, loading directly!", accelerator)
 
             # 异步预取下一批数据
             if step + prefetch_factor < len(train_loader):
                 try:
                     batch = next(data_iter)
                     prefetch_batches.append(batch)
+                    # 记录添加新batch后的内存变化
+                    if args.memory_monitor and step % args.memory_monitor_interval == 0:
+                        log_memory_status(step, prefetch_batches, accelerator, "after_add_batch")
                 except StopIteration:
                     pass
 
             # 计时数据加载结束 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and data_end is not None:
-            #     data_end.record()
+                data_end.record()
 
             # 更新学习率
             if scheduler is not None:
                 scheduler.step()
 
             # 计时前向传播 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and forward_start is not None:
-            #     forward_start.record()
+                forward_start.record()
 
             # 前向传播
             with ctx:
@@ -311,24 +397,24 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
                 loss = loss / args.accumulation_steps
 
             # 计时前向传播结束 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and forward_end is not None:
-            #     forward_end.record()
+                forward_end.record()
 
             # 计时反向传播 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and backward_start is not None:
-            #     backward_start.record()
+                backward_start.record()
 
             # 反向传播
             # 当使用DeepSpeed时，它会自动处理梯度累积和梯度裁剪
             accelerator.backward(loss)
 
             # 计时反向传播结束 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and backward_end is not None:
-            #     backward_end.record()
+                backward_end.record()
 
             # 计时优化器步骤 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and optimizer_start is not None:
-            #     optimizer_start.record()
+                optimizer_start.record()
 
             # 优化器步骤 - 当使用DeepSpeed时，它会自动处理梯度累积和梯度裁剪
             # 只有在达到累积步数时才会执行优化器步骤
@@ -340,40 +426,58 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
             optimizer.zero_grad()
 
             # 计时优化器步骤结束 (只在主进程进行)
-            # if args.profile and accelerator.is_main_process:
+            if args.profile and accelerator.is_main_process and optimizer_end is not None:
-            #     optimizer_end.record()
+                optimizer_end.record()
 
             # 打印训练信息 (只在主进程进行)
             if (step + 1) % args.log_interval == 0 and accelerator.is_main_process:
                 current_time = time.time()
+                
+                # 记录日志输出时的详细内存状态
+                if args.memory_monitor:
+                    log_memory_status(step, prefetch_batches, accelerator, "at_log_interval", detailed=True)
+                    
+                    # 强制垃圾回收并记录内存变化
+                    if torch.cuda.is_available():
+                        torch.cuda.empty_cache()
+                    gc.collect()
+                    log_memory_status(step, prefetch_batches, accelerator, "after_gc", detailed=True)
+                
                 # 计算性能指标
-                if args.profile:
+                if args.profile and accelerator.is_main_process:
                     torch.cuda.synchronize()
-                    # 使用自上次日志以来的时间计算性能指标，而不是总时间
+                    
-                    data_time = data_start.elapsed_time(data_end)
+                    # 确保所有事件都已记录才计算elapsed_time
-                    forward_time = forward_start.elapsed_time(forward_end)
+                    try:
-                    backward_time = backward_start.elapsed_time(backward_end)
+                        data_time = data_start.elapsed_time(data_end) if data_start is not None and data_end is not None else 0
-                    optimizer_time = optimizer_start.elapsed_time(optimizer_end)
+                        forward_time = forward_start.elapsed_time(forward_end) if forward_start is not None and forward_end is not None else 0
-                    iter_time = (current_time - last_log_time) * 1000 / args.log_interval # avg ms per iteration since last log
+                        backward_time = backward_start.elapsed_time(backward_end) if backward_start is not None and backward_end is not None else 0
-                    # total_time_ms = data_time + forward_time + backward_time + optimizer_time
+                        optimizer_time = optimizer_start.elapsed_time(optimizer_end) if optimizer_start is not None and optimizer_end is not None else 0
+                        iter_time = (current_time - last_log_time) * 1000 / args.log_interval # avg ms per iteration since last log
+                        # total_time_ms = data_time + forward_time + backward_time + optimizer_time
 
-                    # 打印性能分析
+                        # 打印性能分析
-                    if (step + 1) % (args.log_interval * args.profile_interval) == 0:
+                        if (step + 1) % (args.log_interval * args.profile_interval) == 0:
-                        Logger(f"性能分析 (Avg/iter over last {args.log_interval} steps) - "
+                            Logger(f"性能分析 (Avg/iter over last {args.log_interval} steps) - "
-                              f"Data: {data_time/args.log_interval:.2f}ms, "
+                                  f"Data: {data_time/args.log_interval:.2f}ms, "
-                              f"Fwd: {forward_time/args.log_interval:.2f}ms, "
+                                  f"Fwd: {forward_time/args.log_interval:.2f}ms, "
-                              f"Bwd: {backward_time/args.log_interval:.2f}ms, "
+                                  f"Bwd: {backward_time/args.log_interval:.2f}ms, "
-                              f"Optim: {optimizer_time/args.log_interval:.2f}ms, "
+                                  f"Optim: {optimizer_time/args.log_interval:.2f}ms, "
-                              f"Iter Time: {iter_time:.2f}ms", accelerator)
+                                  f"Iter Time: {iter_time:.2f}ms", accelerator)
-                        # 重置事件以便下次测量从0开始
+                            # 重置事件以便下次测量从0开始
-                        data_start = torch.cuda.Event(enable_timing=True)
+                            data_start = torch.cuda.Event(enable_timing=True)
-                        data_end = torch.cuda.Event(enable_timing=True)
+                            data_end = torch.cuda.Event(enable_timing=True)
-                        forward_start = torch.cuda.Event(enable_timing=True)
+                            forward_start = torch.cuda.Event(enable_timing=True)
-                        forward_end = torch.cuda.Event(enable_timing=True)
+                            forward_end = torch.cuda.Event(enable_timing=True)
-                        backward_start = torch.cuda.Event(enable_timing=True)
+                            backward_start = torch.cuda.Event(enable_timing=True)
-                        backward_end = torch.cuda.Event(enable_timing=True)
+                            backward_end = torch.cuda.Event(enable_timing=True)
-                        optimizer_start = torch.cuda.Event(enable_timing=True)
+                            optimizer_start = torch.cuda.Event(enable_timing=True)
-                        optimizer_end = torch.cuda.Event(enable_timing=True)
+                            optimizer_end = torch.cuda.Event(enable_timing=True)
+                    except RuntimeError as e:
+                        if "Both events must be recorded" in str(e):
+                            Logger(f"Warning: CUDA events not properly recorded, skipping performance analysis: {e}", accelerator)
+                        else:
+                            raise e
 
 
                 # 计算当前学习率
@@ -419,7 +523,7 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
 
             # 保存模型 (只在主进程进行)
             loss_total = loss.item() * args.accumulation_steps
-            if epoch > 1 or best_loss > loss_total and accelerator.is_main_process:
+            if epoch > 1 and best_loss > loss_total and accelerator.is_main_process:
                 best_loss = loss_total
                 # 使用函数开始处定义的moe_path变量
                 ckp = f'{args.save_dir}/pretrain_{args.dim}{moe_path}.pth'
@@ -433,8 +537,33 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
 
         except Exception as e:
             Logger(f"Error in training step: {e}", accelerator)
+            # 记录异常时的内存状态
+            if args.memory_monitor:
+                log_memory_status(step, prefetch_batches, accelerator, "at_exception", detailed=True)
             import traceback
             Logger(traceback.format_exc(), accelerator)
+            
+            # 清理prefetch_batches，防止内存泄漏
+            if args.memory_monitor and accelerator.is_main_process:
+                Logger(f"[Memory Monitor] Clearing prefetch_batches due to exception. Current length: {len(prefetch_batches)}", accelerator)
+            prefetch_batches.clear()
+            gc.collect()
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+            if args.memory_monitor:
+                log_memory_status(step, prefetch_batches, accelerator, "after_exception_cleanup", detailed=True)
+
+    # 训练epoch结束时清理prefetch_batches
+    if args.memory_monitor:
+        if accelerator.is_main_process:
+            Logger(f"[Memory Monitor] Epoch {epoch+1} finished. Clearing prefetch_batches. Final length: {len(prefetch_batches)}", accelerator)
+        log_memory_status(total_steps_in_epoch-1, prefetch_batches, accelerator, "before_epoch_end_cleanup", detailed=True)
+    prefetch_batches.clear()
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    if args.memory_monitor:
+        log_memory_status(total_steps_in_epoch-1, prefetch_batches, accelerator, "after_epoch_end_cleanup", detailed=True)
 
 def main():
     parser = argparse.ArgumentParser(description="MiniMind Pretraining with Accelerate")
@@ -468,6 +597,8 @@ def main():
     parser.add_argument("--fast_clustering", action="store_true", default=True, help="使用快速近似聚类算法（适用于大数据集）")
     parser.add_argument("--cluster_cache_path", type=str, default="./cache/cluster_tokens_single.pt", help="聚类结果缓存文件路径")
     parser.add_argument("--recompute_clusters", action="store_true", default=False, help="强制重新计算聚类，忽略缓存文件")
+    parser.add_argument("--memory_monitor", action="store_true", default=False, help="启用内存监控")
+    parser.add_argument("--memory_monitor_interval", type=int, default=10, help="内存监控间隔（步数）")
     args = parser.parse_args()
  
     #########################################################
@@ -629,7 +760,25 @@ def main():
     #########################################################
     overall_start_time = time.time() # Record overall start time
     for epoch in range(args.epochs):
+        Logger(f"开始第{epoch+1}轮训练", accelerator)
         train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, args, ctx, overall_start_time, swanlab_run) # Pass overall start time
+        
+        # 每个epoch结束后进行内存清理
+        Logger(f"第{epoch+1}轮训练完成，进行内存清理", accelerator)
+        gc.collect()
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        
+        # 记录epoch结束时的内存状态
+        if accelerator.is_main_process:
+            memory_info = get_memory_usage()
+            cuda_info = get_cuda_memory_usage()
+            log_msg = f"[Memory Monitor] Epoch {epoch+1} completed - "
+            log_msg += f"System RSS: {memory_info['rss_mb']:.2f}MB"
+            if cuda_info:
+                log_msg += f", CUDA allocated: {cuda_info['cuda_allocated_mb']:.2f}MB"
+                log_msg += f", CUDA reserved: {cuda_info['cuda_reserved_mb']:.2f}MB"
+            Logger(log_msg, accelerator)
 
     #########################################################
     # 关闭SwanLab