model/LMConfig.py

@@ -50,12 +50,6 @@ class LMConfig(PretrainedConfig):
             use_token_memory: bool = True,  # 🔥 1.4.6: 新增token-based memory flag
             freeze_ratio: float = 0.2,     # 🔥 新增: memory_bank冻结率 (0.0表示不冻结，0.2表示20%条目不更新)
             ####################################################
-            # Experiment 1.4.9: Gumbel-Softmax + Diversity Loss
-            ####################################################
-            num_candidates: int = 32,       # 🔥 实验1.4.9: 候选记忆条目数量
-            num_selected: int = 1,          # 🔥 实验1.4.9: 选中的记忆条目数量 (现在只选1个最佳)
-            gumbel_temperature: float = 1.0, # 🔥 实验1.4.9: Gumbel-Softmax温度参数
-            ####################################################
             # Triple extraction related configurations
             ####################################################
             max_subject_len: int = 8,
@@ -105,12 +99,6 @@ class LMConfig(PretrainedConfig):
         self.use_token_memory = use_token_memory  # 🔥 1.4.6: token-based memory flag
         self.freeze_ratio = freeze_ratio  # 🔥 新增: memory_bank冻结率
         ####################################################
-        # Experiment 1.4.9: Gumbel-Softmax + Diversity Loss
-        ####################################################
-        self.num_candidates = num_candidates
-        self.num_selected = num_selected
-        self.gumbel_temperature = gumbel_temperature
-        ####################################################
         # Triple extraction related configurations
         ####################################################
         self.max_subject_len = max_subject_len

model/model_memory.py

@@ -123,15 +123,14 @@ class Attention(nn.Module):
 
 
 class MemoryGate(nn.Module):
-    """Product Key Memory-based gate mechanism for memory selection with Gumbel-Softmax"""
+    """Product Key Memory-based gate mechanism for memory selection"""
     def __init__(self, config: LMConfig):
         super().__init__()
         self.config = config
         self.dim = config.dim
         self.knowledge_num = config.knowledge_num
         self.knowledge_dim = config.knowledge_dim
-        self.num_candidates = getattr(config, 'num_candidates', 32)  # Generate 32 candidates
+        self.num_selected = getattr(config, 'num_selected', 16)
-        self.num_selected = getattr(config, 'num_selected', 1)  # Select 1 best from candidates
         
         # 确保知识库数量是完全平方数
         assert int(self.knowledge_num ** 0.5) ** 2 == self.knowledge_num, \
@@ -170,28 +169,29 @@ class MemoryGate(nn.Module):
         scores_1 = torch.einsum('bsd,kd->bsk', q1, self.keys[0])  # [batch, seq_len, num_keys]
         scores_2 = torch.einsum('bsd,kd->bsk', q2, self.keys[1])  # [batch, seq_len, num_keys]
         
-        # 获取top-k candidates (now using num_candidates instead of num_selected)
+        # 获取top-k
-        topk_scores_1, topk_indices_1 = scores_1.topk(self.num_candidates, dim=-1)
+        topk_scores_1, topk_indices_1 = scores_1.topk(self.num_selected, dim=-1)
-        topk_scores_2, topk_indices_2 = scores_2.topk(self.num_candidates, dim=-1)
+        topk_scores_2, topk_indices_2 = scores_2.topk(self.num_selected, dim=-1)
         
         # 组合product key的结果
-        combined_scores = topk_scores_1.unsqueeze(-1) + topk_scores_2.unsqueeze(-2)  # [batch, seq_len, num_candidates, num_candidates]
+        combined_scores = topk_scores_1.unsqueeze(-1) + topk_scores_2.unsqueeze(-2)  # [batch, seq_len, num_selected, num_selected]
-        combined_indices = topk_indices_1.unsqueeze(-1) * self.num_keys + topk_indices_2.unsqueeze(-2)  # [batch, seq_len, num_candidates, num_candidates]
+        combined_indices = topk_indices_1.unsqueeze(-1) * self.num_keys + topk_indices_2.unsqueeze(-2)  # [batch, seq_len, num_selected, num_selected]
         
-        # 展平并选择最终的top-k candidates
+        # 展平并选择最终的top-k
         combined_scores = combined_scores.view(bsz, seq_len, -1)
         combined_indices = combined_indices.view(bsz, seq_len, -1)
         
-        candidate_scores, candidate_pk_indices = combined_scores.topk(self.num_candidates, dim=-1)
+        final_scores, final_pk_indices = combined_scores.topk(self.num_selected, dim=-1)
-        candidate_indices = combined_indices.gather(-1, candidate_pk_indices)  # [batch, seq_len, num_candidates]
+        memory_indices = combined_indices.gather(-1, final_pk_indices)
         
-        # 归一化候选分数
+        # 归一化分数
-        candidate_scores = F.softmax(candidate_scores, dim=-1)
+        memory_scores = F.softmax(final_scores, dim=-1)
-        candidate_scores = self.dropout(candidate_scores)
+        memory_scores = self.dropout(memory_scores)
         
-        # 返回候选项用于后续的相似度选择
+        # 计算平衡损失和监控统计
-        # 注意：这里返回候选项，在MiniMindBlock中进行相似度选择和多样性损失计算
+        balance_loss, stats = self._compute_balance_loss_and_stats(memory_indices, memory_scores)
-        return candidate_indices, candidate_scores, None, {}
+        
+        return memory_indices, memory_scores, balance_loss, stats
     
     def _compute_balance_loss_and_stats(self, memory_indices, memory_scores):
         """
@@ -276,10 +276,10 @@ class GatedMemoryFusion(nn.Module):
         self.config = config
         self.dim = config.dim
         self.knowledge_dim = config.knowledge_dim
-        self.num_selected = getattr(config, 'num_selected', 1)  # Now we select 1 best memory
+        self.num_selected = getattr(config, 'num_selected', 16)
         
-        # 输入维度：dim (h_attn) + num_selected * dim (选中的记忆，现在只有1个)
+        # 输入维度：dim (h_attn) + num_selected * knowledge_dim (选中的记忆)
-        # 实验1.4.9：修改为只选择1个最佳记忆
+        # 实验1.4.6：记忆解码后立即压缩回knowledge_dim避免显存爆炸
         concat_dim = self.dim + self.num_selected * self.dim
         
         # 类似SwiGLU的门控MLP结构
@@ -289,18 +289,23 @@ class GatedMemoryFusion(nn.Module):
         
         self.dropout = nn.Dropout(config.dropout)
         
-    def forward(self, h_attn: torch.Tensor, selected_memory: torch.Tensor):
+    def forward(self, h_attn: torch.Tensor, selected_memories: torch.Tensor, memory_scores: torch.Tensor):
         """
         Args:
             h_attn: [batch_size, seq_len, dim] - Self attention output
-            selected_memory: [batch_size, seq_len, dim] - Selected single best memory
+            selected_memories: [batch_size, seq_len, num_selected, knowledge_dim] - Selected memory data
+            memory_scores: [batch_size, seq_len, num_selected] - Memory selection weights (not used in concatenation approach)
         Returns:
             output: [batch_size, seq_len, dim]
         """
         bsz, seq_len, _ = h_attn.shape
         
-        # 拼接h_attn和最佳记忆
+        # 将选中的记忆展平为一维向量
-        concat_input = torch.cat([h_attn, selected_memory], dim=-1)  # [batch, seq_len, dim + dim]
+        # [batch, seq_len, num_selected, knowledge_dim] -> [batch, seq_len, num_selected * knowledge_dim]
+        memory_flat = selected_memories.reshape(bsz, seq_len, -1)
+        
+        # 拼接h_attn和记忆信息
+        concat_input = torch.cat([h_attn, memory_flat], dim=-1)  # [batch, seq_len, dim + num_selected * dim]
         
         # 门控MLP处理（类似SwiGLU）
         gate = F.silu(self.gate_proj(concat_input))  # [batch, seq_len, dim]
@@ -332,96 +337,22 @@ class MiniMindBlock(nn.Module):
         self.memory_gate = MemoryGate(config)
         self.gated_memory_fusion = GatedMemoryFusion(config)
 
-        # Gumbel-Softmax参数
-        self.gumbel_temperature = getattr(config, 'gumbel_temperature', 1.0)
-
         # self.attentionpool = nn.Linear(config.dim, 1)
 
-    def gumbel_softmax_selection(self, similarity_scores, temperature=1.0, hard=True):
-        """
-        使用Gumbel-Softmax进行可微分的离散选择
-        
-        Args:
-            similarity_scores: [batch_size, seq_len, num_candidates] - 相似度分数
-            temperature: Gumbel-Softmax温度参数
-            hard: 是否使用硬选择（one-hot）
-        
-        Returns:
-            selection_weights: [batch_size, seq_len, num_candidates] - 选择权重
-            selected_indices: [batch_size, seq_len] - 选中的索引（用于统计）
-        """
-        # 添加Gumbel噪声
-        gumbel_noise = -torch.log(-torch.log(torch.rand_like(similarity_scores) + 1e-20) + 1e-20)
-        logits = (similarity_scores + gumbel_noise) / temperature
-        
-        # Softmax
-        soft_weights = F.softmax(logits, dim=-1)
-        
-        if hard:
-            # 硬选择：创建one-hot向量
-            _, max_indices = soft_weights.max(dim=-1, keepdim=True)
-            hard_weights = torch.zeros_like(soft_weights).scatter_(-1, max_indices, 1.0)
-            # 使用straight-through estimator
-            selection_weights = hard_weights - soft_weights.detach() + soft_weights
-            selected_indices = max_indices.squeeze(-1)  # [batch_size, seq_len]
-        else:
-            # 软选择
-            selection_weights = soft_weights
-            selected_indices = torch.argmax(soft_weights, dim=-1)
-            
-        return selection_weights, selected_indices
-    
-    def compute_diversity_loss(self, candidate_memories):
-        """
-        计算候选集内部多样性损失（鼓励候选项之间的差异性）
-        
-        Args:
-            candidate_memories: [batch_size, seq_len, num_candidates, dim]
-        
-        Returns:
-            diversity_loss: 标量张量
-        """
-        bsz, seq_len, num_candidates, dim = candidate_memories.shape
-        
-        # 计算候选项之间的相似度矩阵
-        # 归一化候选记忆用于计算余弦相似度
-        normalized_memories = F.normalize(candidate_memories, p=2, dim=-1)  # [batch, seq_len, num_candidates, dim]
-        
-        # 计算相似度矩阵: [batch, seq_len, num_candidates, num_candidates]
-        similarity_matrix = torch.matmul(normalized_memories, normalized_memories.transpose(-2, -1))
-        
-        # 移除对角线（自相似度=1）
-        mask = torch.eye(num_candidates, device=candidate_memories.device).bool()
-        mask = mask.unsqueeze(0).unsqueeze(0).expand(bsz, seq_len, -1, -1)
-        
-        # 计算非对角线元素的平均相似度（希望越小越好，表示越多样）
-        off_diagonal_similarities = similarity_matrix.masked_select(~mask)
-        avg_similarity = off_diagonal_similarities.mean()
-        
-        # 多样性损失：相似度越高，损失越大
-        diversity_loss = avg_similarity
-        
-        return diversity_loss
-
     def forward(self, x, pos_cis, memory_bank, tok_embeddings, collect_ema_stats=False):
         """
-        实验1.4.9: Gumbel-Softmax + 多样性损失 + 可微分相似度损失
-        
         Args:
             x: [batch_size, seq_len, dim]
             pos_cis: positional encoding
-            memory_bank: [knowledge_num, knowledge_length] - shared memory bank with token IDs
+            memory_bank: [knowledge_num, knowledge_dim] - shared memory bank
-            tok_embeddings: token embedding layer
             collect_ema_stats: 是否收集EMA更新统计信息
         
         Returns:
             out: [batch_size, seq_len, dim]
-            balance_loss: 该层的平衡损失 (从候选项计算)
+            balance_loss: 该层的平衡损失
-            similarity_loss: 相似度损失 (可微分)
-            diversity_loss: 多样性损失
             layer_stats: 该层的监控统计信息
             ema_stats: EMA更新统计信息（如果collect_ema_stats=True）
-            cosine_stats: 查找向量与候选记忆条目的余弦相似度统计信息
+            cosine_stats: 查找向量与选中记忆条目的余弦相似度统计信息
         """
         # Self attention
         h_attn = self.attention(self.attention_norm(x), pos_cis)
@@ -430,147 +361,68 @@ class MiniMindBlock(nn.Module):
         # 使用h_attn作为门控和交叉注意力的输入（核心：self attention的输出）
         h_for_memory = self.memory_norm(h_attn)
         
-        # 🔥 新架构：生成32个候选项
+        # 门控选择记忆
-        candidate_indices, candidate_scores, _, _ = self.memory_gate(h_for_memory)
+        memory_indices, memory_scores, balance_loss, layer_stats = self.memory_gate(h_for_memory)
-        # candidate_indices: [batch, seq_len, num_candidates]
-        # candidate_scores: [batch, seq_len, num_candidates]
         
-        bsz, seq_len, num_candidates = candidate_indices.shape
+        # 根据索引获取记忆数据 - 实验1.4.6：解码token_id为特征向量
+        bsz, seq_len, num_selected = memory_indices.shape
+        memory_indices_flat = memory_indices.view(-1)
+        selected_token_ids = memory_bank[memory_indices_flat]  # [batch * seq_len * num_selected, knowledge_length]
         
-        # 解码候选token_ids为特征向量
+        # 解码token_ids为特征向量并立即压缩避免显存爆炸
-        candidate_indices_flat = candidate_indices.view(-1)  # [batch * seq_len * num_candidates]
+        selected_embeddings = tok_embeddings(selected_token_ids)  # [batch * seq_len * num_selected, knowledge_length, dim]
-        candidate_token_ids = memory_bank[candidate_indices_flat]  # [batch * seq_len * num_candidates, knowledge_length]
         
-        # 解码为embeddings并池化
+        # 立即压缩：knowledge_length * dim -> knowledge_dim 避免显存爆炸
-        candidate_embeddings = tok_embeddings(candidate_token_ids)  # [batch * seq_len * num_candidates, knowledge_length, dim]
+        # 使用平均池化压缩knowledge_length维度
-        candidate_memories = candidate_embeddings.mean(dim=1)  # [batch * seq_len * num_candidates, dim]
+        pooled_memory = selected_embeddings.mean(dim=1)  # [batch * seq_len * num_selected, dim]
-        candidate_memories = candidate_memories.view(bsz, seq_len, num_candidates, self.dim)  # [batch, seq_len, num_candidates, dim]
+        # attn_weights = self.attentionpool(selected_embeddings)  
+        # attn_weights = torch.softmax(attn_weights, dim=1)
+        # pooled_memory = torch.sum(selected_embeddings * attn_weights, dim=1)
             
-        # 🔥 核心改进: 计算可微分的相似度分数 (移除no_grad)
+        selected_memory = pooled_memory.view(bsz, seq_len, num_selected, self.dim)  # [batch, seq_len, num_selected, dim]
-        h_expanded = h_for_memory.unsqueeze(2).expand(-1, -1, num_candidates, -1)  # [batch, seq_len, num_candidates, dim]
-        similarity_scores = F.cosine_similarity(h_expanded, candidate_memories, dim=-1)  # [batch, seq_len, num_candidates]
         
-        # 🔥 使用Gumbel-Softmax选择最佳候选项
+        # 门控MLP融合：串型连接h_attn和选中的记忆
-        selection_weights, selected_indices = self.gumbel_softmax_selection(
+        memory_output = self.gated_memory_fusion(h_for_memory, selected_memory, memory_scores)
-            similarity_scores, 
-            temperature=self.gumbel_temperature,
-            hard=True
-        )  # selection_weights: [batch, seq_len, num_candidates], selected_indices: [batch, seq_len]
-        
-        # 🔥 计算相似度损失 (现在是可微分的!)
-        # 相似度损失：希望选中的记忆与查询向量相似度尽可能高
-        selected_similarities = (similarity_scores * selection_weights).sum(dim=-1)  # [batch, seq_len]
-        similarity_loss = -selected_similarities.mean()  # 负号：相似度越高，损失越小
-        
-        # 🔥 计算候选集多样性损失
-        diversity_loss = self.compute_diversity_loss(candidate_memories)
-        
-        # 🔥 使用selection_weights进行加权选择最终记忆
-        selected_memory = (candidate_memories * selection_weights.unsqueeze(-1)).sum(dim=2)  # [batch, seq_len, dim]
-        
-        # 门控MLP融合：只融合选中的单个最佳记忆
-        memory_output = self.gated_memory_fusion(h_for_memory, selected_memory)
         
         # 残差连接
         out = h + memory_output
         
-        # 🔥 计算平衡损失和统计信息 (基于候选项的选择分布)
+        # 🔍 新增: 计算查找向量与选中记忆条目的余弦相似度
-        balance_loss, layer_stats = self._compute_candidate_balance_stats(candidate_indices, selection_weights)
+        with torch.no_grad():
             
-        # 🔥 计算详细的相似度统计信息
+            # 扩展查找向量维度以匹配selected_memory
-        cosine_stats = {
+            h_expanded = h_for_memory.unsqueeze(2).expand(-1, -1, num_selected, -1)  # [batch, seq_len, num_selected, dim]
-            'similarity_scores': similarity_scores,                                    # [batch, seq_len, num_candidates]
-            'selected_similarities': selected_similarities,                           # [batch, seq_len]
-            'avg_similarity': similarity_scores.mean().item(),                        # 平均相似度
-            'max_similarity': similarity_scores.max().item(),                         # 最大相似度
-            'min_similarity': similarity_scores.min().item(),                         # 最小相似度
-            'selected_avg_similarity': selected_similarities.mean().item(),           # 选中记忆的平均相似度
-            'selection_entropy': -torch.sum(selection_weights * torch.log(selection_weights + 1e-10), dim=-1).mean().item()  # 选择熵
-        }
             
-        # 收集EMA更新统计信息（现在基于选中的记忆）
+            # 计算余弦相似度：cosine_sim(query, memory) for each selected memory
+            cosine_similarities = F.cosine_similarity(
+                h_expanded,           # [batch, seq_len, num_selected, dim]
+                selected_memory,      # [batch, seq_len, num_selected, knowledge_dim]
+                dim=-1               # 在knowledge_dim维度计算余弦相似度
+            )  # [batch, seq_len, num_selected]
+            
+            # 计算余弦相似度统计信息
+            cosine_stats = {
+                'cosine_similarities': cosine_similarities,                    # [batch, seq_len, num_selected]
+                'avg_cosine_similarity': cosine_similarities.mean().item(),   # 平均余弦相似度
+                'max_cosine_similarity': cosine_similarities.max().item(),    # 最大余弦相似度
+                'min_cosine_similarity': cosine_similarities.min().item(),    # 最小余弦相似度
+                'std_cosine_similarity': cosine_similarities.std().item(),    # 余弦相似度标准差
+            }
+        
+        # 收集EMA更新统计信息（仅在训练时且启用时）
         ema_stats = None
         if collect_ema_stats and self.training:
-            # 扩展选中的索引以匹配EMA更新的期望格式
-            selected_memory_indices = candidate_indices.gather(2, selected_indices.unsqueeze(-1))  # [batch, seq_len, 1]
             ema_stats = {
-                'memory_indices': selected_memory_indices,          # [batch, seq_len, 1]
+                'memory_indices': memory_indices,      # [batch, seq_len, num_selected]
-                'memory_scores': torch.ones_like(selected_memory_indices.float()),   # [batch, seq_len, 1] - 选中的权重为1
+                'memory_scores': memory_scores,        # [batch, seq_len, num_selected]  
-                'h_for_memory': h_for_memory,                       # [batch, seq_len, dim]
+                'h_for_memory': h_for_memory,          # [batch, seq_len, dim]
-                'selected_memory': selected_memory.unsqueeze(2),    # [batch, seq_len, 1, dim]
+                'selected_memory': selected_memory,     # [batch, seq_len, num_selected, knowledge_dim]
             }
         
         if collect_ema_stats:
-            return out, balance_loss, similarity_loss, diversity_loss, layer_stats, ema_stats, cosine_stats
+            return out, balance_loss, layer_stats, ema_stats, cosine_stats
         else:
-            return out, balance_loss, similarity_loss, diversity_loss, layer_stats, cosine_stats
+            return out, balance_loss, layer_stats, cosine_stats
-    
-    def _compute_candidate_balance_stats(self, candidate_indices, selection_weights):
-        """
-        计算基于候选项选择的平衡损失和统计信息
-        
-        Args:
-            candidate_indices: [batch_size, seq_len, num_candidates]
-            selection_weights: [batch_size, seq_len, num_candidates] - Gumbel-Softmax权重
-        
-        Returns:
-            balance_loss: 标量张量
-            stats: 统计信息字典
-        """
-        bsz, seq_len, num_candidates = candidate_indices.shape
-        device = candidate_indices.device
-        
-        # 使用加权统计每个记忆条目被选中的概率
-        flat_indices = candidate_indices.view(-1)  # [batch * seq_len * num_candidates]
-        flat_weights = selection_weights.view(-1)  # [batch * seq_len * num_candidates]
-        
-        # 统计每个记忆条目被选中的加权次数
-        memory_counts = torch.zeros(self.config.knowledge_num, device=device)
-        memory_counts.scatter_add_(0, flat_indices, flat_weights)
-        
-        # 计算选择概率分布
-        total_selections = memory_counts.sum()
-        memory_probs = memory_counts / (total_selections + 1e-10)
-        
-        # 计算KL散度损失（与均匀分布的KL散度）
-        uniform_prob = 1.0 / self.config.knowledge_num
-        memory_probs_safe = memory_probs + 1e-10
-        kl_loss = F.kl_div(
-            torch.log(memory_probs_safe),
-            torch.full_like(memory_probs, uniform_prob),
-            reduction='sum'
-        )
-        
-        # 计算基尼系数损失
-        sorted_probs, _ = torch.sort(memory_probs)
-        n = self.config.knowledge_num
-        index = torch.arange(1, n + 1, device=device, dtype=torch.float)
-        gini_coeff = (2 * torch.sum(index * sorted_probs) / (n * torch.sum(sorted_probs))) - (n + 1) / n
-        gini_loss = gini_coeff
-        
-        # 组合平衡损失
-        balance_loss = 0.5 * kl_loss + 0.5 * gini_loss
-        
-        # 计算统计信息
-        with torch.no_grad():
-            coverage_rate = (memory_counts > 0.01).float().mean().item()  # 被选中概率>1%的记忆比例
-            top10_threshold = torch.quantile(memory_counts, 0.9)
-            hot_memories = (memory_counts >= top10_threshold).sum().item()
-            dead_memories = (memory_counts < 0.01).sum().item()  # 几乎从未被选中的记忆
-            selection_variance = memory_counts.var().item()
-            
-            stats = {
-                'gini_coefficient': gini_coeff.item(),
-                'kl_divergence': kl_loss.item(),
-                'coverage_rate': coverage_rate,
-                'hot_memories': hot_memories,
-                'dead_memories': dead_memories,
-                'selection_variance': selection_variance,
-                'max_selections': memory_counts.max().item(),
-                'min_selections': memory_counts.min().item(),
-            }
-        
-        return balance_loss, stats
 
 
 class MiniMindLM(PreTrainedModel):
@@ -622,12 +474,10 @@ class MiniMindLM(PreTrainedModel):
             # 固定冻结前面的条目
             freeze_mask[:freeze_num] = True
             self.register_buffer('freeze_mask', freeze_mask, persistent=False)
-            print(f"🔥 Memory bank freezing enabled: {freeze_num}/{params.knowledge_num} entries ({params.freeze_ratio*100:.1f}%) frozen", flush=True)
+            print(f"🔥 Memory bank freezing enabled: {freeze_num}/{params.knowledge_num} entries ({params.freeze_ratio*100:.1f}%) frozen")
-            import sys; sys.stdout.flush()
         else:
             self.register_buffer('freeze_mask', torch.zeros(params.knowledge_num, dtype=torch.bool), persistent=False)
-            print(f"🔥 Memory bank freezing disabled: all entries can be updated", flush=True)
+            print(f"🔥 Memory bank freezing disabled: all entries can be updated")
-            import sys; sys.stdout.flush()
         
         self.OUT = CausalLMOutputWithPast()
     
@@ -689,26 +539,20 @@ class MiniMindLM(PreTrainedModel):
         h = self.dropout(self.tok_embeddings(input_ids))
         pos_cis = self.pos_cis[start_pos:start_pos + input_ids.size(1)]
         
-        # 收集所有层的损失和统计信息 - 实验1.4.9: 四损失系统
+        # 收集所有层的平衡损失和统计信息
         total_balance_loss = 0
-        total_similarity_loss = 0
-        total_diversity_loss = 0
         all_layer_stats = {}
         all_ema_stats = {}
         all_cosine_stats = {}
         
         for layer_idx, layer in enumerate(self.layers):
             if collect_ema_stats:
-                h, balance_loss, similarity_loss, diversity_loss, layer_stats, ema_stats, cosine_stats = layer(h, pos_cis, self.memory_bank, self.tok_embeddings, collect_ema_stats=True)
+                h, balance_loss, layer_stats, ema_stats, cosine_stats = layer(h, pos_cis, self.memory_bank, self.tok_embeddings, collect_ema_stats=True)
                 all_ema_stats[f'layer_{layer_idx}'] = ema_stats
             else:
-                h, balance_loss, similarity_loss, diversity_loss, layer_stats, cosine_stats = layer(h, pos_cis, self.memory_bank, self.tok_embeddings, collect_ema_stats=False)
+                h, balance_loss, layer_stats, cosine_stats = layer(h, pos_cis, self.memory_bank, self.tok_embeddings, collect_ema_stats=False)
             
-            # 累加四种损失
             total_balance_loss += balance_loss
-            total_similarity_loss += similarity_loss
-            total_diversity_loss += diversity_loss
-            
             # 为每层的统计信息添加前缀
             for key, value in layer_stats.items():
                 all_layer_stats[f'layer_{layer_idx}_{key}'] = value
@@ -719,12 +563,8 @@ class MiniMindLM(PreTrainedModel):
 
         logits = self.output(self.norm(h))
         
-        # 🔥 新的四损失结构
+        # 使用总的平衡损失作为aux_loss
-        aux_loss = {
+        aux_loss = total_balance_loss
-            'balance_loss': total_balance_loss,
-            'similarity_loss': total_similarity_loss,
-            'diversity_loss': total_diversity_loss,
-        }
         
         self.OUT.__setitem__('last_hidden_state', h)
         self.OUT.__setitem__('logits', logits)

model/model_memory_1_4_9.py

@@ -1,770 +0,0 @@
-import math
-import struct
-import inspect
-import time
-
-from .LMConfig import LMConfig
-from typing import Any, Optional, Tuple, List, Union
-import numpy as np
-import torch
-import torch.nn.functional as F
-from torch import nn
-from transformers import PreTrainedModel
-from transformers.modeling_outputs import CausalLMOutputWithPast
-
-
-class RMSNorm(torch.nn.Module):
-    def __init__(self, dim: int, eps: float = 1e-6):
-        super().__init__()
-        self.eps = eps
-        self.weight = nn.Parameter(torch.ones(dim))
-
-    def _norm(self, x):
-        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
-
-    def forward(self, x):
-        return self.weight * self._norm(x.float()).type_as(x)
-
-
-def precompute_pos_cis(dim: int, end: int = int(32 * 1024), theta: float = 1e6):
-    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
-    t = torch.arange(end, device=freqs.device)  # type: ignore
-    freqs = torch.outer(t, freqs).float()  # type: ignore
-    pos_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
-    return pos_cis
-
-
-def apply_rotary_emb(xq, xk, pos_cis):
-    def unite_shape(pos_cis, x):
-        ndim = x.ndim
-        assert 0 <= 1 < ndim
-        assert pos_cis.shape == (x.shape[1], x.shape[-1])
-        shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
-        return pos_cis.view(*shape)
-
-    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
-    xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
-    pos_cis = unite_shape(pos_cis, xq_)
-    xq_out = torch.view_as_real(xq_ * pos_cis).flatten(3)
-    xk_out = torch.view_as_real(xk_ * pos_cis).flatten(3)
-    return xq_out.type_as(xq), xk_out.type_as(xk)
-
-
-def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
-    bs, slen, n_kv_heads, head_dim = x.shape
-    if n_rep == 1:
-        return x
-    return (
-        x[:, :, :, None, :]
-        .expand(bs, slen, n_kv_heads, n_rep, head_dim)
-        .reshape(bs, slen, n_kv_heads * n_rep, head_dim)
-    )
-
-
-class Attention(nn.Module):
-    """Self attention module without KV cache"""
-    def __init__(self, args: LMConfig):
-        super().__init__()
-        self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
-        assert args.n_heads % self.n_kv_heads == 0
-        self.n_local_heads = args.n_heads
-        self.n_local_kv_heads = self.n_kv_heads
-        self.n_rep = self.n_local_heads // self.n_local_kv_heads
-        self.head_dim = args.dim // args.n_heads
-        self.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False)
-        self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False)
-        self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False)
-        self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False)
-        self.attn_dropout = nn.Dropout(args.dropout)
-        self.resid_dropout = nn.Dropout(args.dropout)
-        self.dropout = args.dropout
-        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention') and args.flash_attn
-        # print("WARNING: using slow attention. Flash Attention requires PyTorch >= 2.0")
-        mask = torch.full((1, 1, args.max_seq_len, args.max_seq_len), float("-inf"))
-        mask = torch.triu(mask, diagonal=1)
-        self.register_buffer("mask", mask, persistent=False)
-
-    def forward(self, x: torch.Tensor, pos_cis: torch.Tensor):
-        """Forward pass without KV cache"""
-        bsz, seq_len, _ = x.shape
-        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
-        xq = xq.view(bsz, seq_len, self.n_local_heads, self.head_dim)
-        xk = xk.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim)
-        xv = xv.view(bsz, seq_len, self.n_local_kv_heads, self.head_dim)
-
-        xq, xk = apply_rotary_emb(xq, xk, pos_cis)
-        
-        # 注意：完全去除了KV cache相关代码
-        
-        xq, xk, xv = (
-            xq.transpose(1, 2),
-            repeat_kv(xk, self.n_rep).transpose(1, 2),
-            repeat_kv(xv, self.n_rep).transpose(1, 2)
-        )
-        if self.flash and seq_len != 1:
-            dropout_p = self.dropout if self.training else 0.0
-            output = F.scaled_dot_product_attention(
-                xq, xk, xv,
-                attn_mask=None,
-                dropout_p=dropout_p,
-                is_causal=True
-            )
-        else:
-            scores = (xq @ xk.transpose(-2, -1)) / math.sqrt(self.head_dim)
-            scores += self.mask[:, :, :seq_len, :seq_len]
-            scores = F.softmax(scores.float(), dim=-1).type_as(xq)
-            scores = self.attn_dropout(scores)
-            output = scores @ xv
-
-        output = output.transpose(1, 2).reshape(bsz, seq_len, -1)
-        output = self.resid_dropout(self.wo(output))
-        return output
-
-
-class MemoryGate(nn.Module):
-    """Product Key Memory-based gate mechanism for memory selection"""
-    def __init__(self, config: LMConfig):
-        super().__init__()
-        self.config = config
-        self.dim = config.dim
-        self.knowledge_num = config.knowledge_num
-        self.knowledge_dim = config.knowledge_dim
-        self.num_selected = getattr(config, 'num_selected', 16)
-        
-        # 确保知识库数量是完全平方数
-        assert int(self.knowledge_num ** 0.5) ** 2 == self.knowledge_num, \
-            f"knowledge_num ({self.knowledge_num}) must be a perfect square for product key memory"
-        
-        self.num_keys = int(self.knowledge_num ** 0.5)
-        
-        # 查询投影：将输入维度映射到knowledge_dim * 2（用于两个product key）
-        self.gate_proj = nn.Linear(self.dim, self.knowledge_dim, bias=False)
-        
-        # Product Key Memory: 两个独立的键集合
-        self.keys = nn.Parameter(torch.randn(2, self.num_keys, self.knowledge_dim // 2))
-        
-        self.dropout = nn.Dropout(config.dropout)
-        
-    def forward(self, x: torch.Tensor):
-        """
-        Args:
-            x: [batch_size, seq_len, dim]
-        Returns:
-            memory_indices: [batch_size, seq_len, num_selected]
-            memory_scores: [batch_size, seq_len, num_selected]
-            balance_loss: 平衡损失（KL散度 + 基尼系数）
-            stats: 监控统计信息字典
-        """
-        bsz, seq_len, _ = x.shape
-        
-        # 生成查询向量
-        queries = self.gate_proj(x)  # [batch, seq_len, knowledge_dim]
-        
-        # 分割为两部分用于product key
-        q1 = queries[:, :, :self.knowledge_dim // 2]  # [batch, seq_len, knowledge_dim // 2]
-        q2 = queries[:, :, self.knowledge_dim // 2:]  # [batch, seq_len, knowledge_dim // 2]
-        
-        # 计算与两个键集合的相似度
-        scores_1 = torch.einsum('bsd,kd->bsk', q1, self.keys[0])  # [batch, seq_len, num_keys]
-        scores_2 = torch.einsum('bsd,kd->bsk', q2, self.keys[1])  # [batch, seq_len, num_keys]
-        
-        # 获取top-k
-        topk_scores_1, topk_indices_1 = scores_1.topk(self.num_selected, dim=-1)
-        topk_scores_2, topk_indices_2 = scores_2.topk(self.num_selected, dim=-1)
-        
-        # 组合product key的结果
-        combined_scores = topk_scores_1.unsqueeze(-1) + topk_scores_2.unsqueeze(-2)  # [batch, seq_len, num_selected, num_selected]
-        combined_indices = topk_indices_1.unsqueeze(-1) * self.num_keys + topk_indices_2.unsqueeze(-2)  # [batch, seq_len, num_selected, num_selected]
-        
-        # 展平并选择最终的top-k
-        combined_scores = combined_scores.view(bsz, seq_len, -1)
-        combined_indices = combined_indices.view(bsz, seq_len, -1)
-        
-        final_scores, final_pk_indices = combined_scores.topk(self.num_selected, dim=-1)
-        memory_indices = combined_indices.gather(-1, final_pk_indices)
-        
-        # 归一化分数
-        memory_scores = F.softmax(final_scores, dim=-1)
-        memory_scores = self.dropout(memory_scores)
-        
-        # 计算平衡损失和监控统计
-        balance_loss, stats = self._compute_balance_loss_and_stats(memory_indices, memory_scores)
-        
-        return memory_indices, memory_scores, balance_loss, stats
-    
-    def _compute_balance_loss_and_stats(self, memory_indices, memory_scores):
-        """
-        计算平衡损失和监控统计信息
-        
-        Args:
-            memory_indices: [batch_size, seq_len, num_selected]
-            memory_scores: [batch_size, seq_len, num_selected]
-        
-        Returns:
-            balance_loss: 标量张量
-            stats: 统计信息字典
-        """
-        bsz, seq_len, num_selected = memory_indices.shape
-        device = memory_indices.device
-        
-        # 1. 计算记忆选择分布
-        # 将所有选择的记忆索引展平
-        flat_indices = memory_indices.view(-1)  # [batch_size * seq_len * num_selected]
-        
-        # 统计每个记忆条目被选中的次数
-        memory_counts = torch.zeros(self.knowledge_num, device=device)
-        memory_counts.scatter_add_(0, flat_indices, torch.ones_like(flat_indices, dtype=torch.float))
-        
-        # 计算选择概率分布
-        total_selections = bsz * seq_len * num_selected
-        memory_probs = memory_counts / total_selections
-        
-        # 2. 计算KL散度损失（与均匀分布的KL散度）
-        uniform_prob = 1.0 / self.knowledge_num
-        # 避免log(0)的问题
-        memory_probs_safe = memory_probs + 1e-10
-        kl_loss = F.kl_div(
-            torch.log(memory_probs_safe),
-            torch.full_like(memory_probs, uniform_prob),
-            reduction='sum'
-        )
-        
-        # 3. 计算基尼系数损失（衡量分布不平等程度）
-        sorted_probs, _ = torch.sort(memory_probs)
-        n = self.knowledge_num
-        index = torch.arange(1, n + 1, device=device, dtype=torch.float)
-        gini_coeff = (2 * torch.sum(index * sorted_probs) / (n * torch.sum(sorted_probs))) - (n + 1) / n
-        gini_loss = gini_coeff  # 基尼系数越大，分布越不均匀
-        
-        # 4. 组合平衡损失
-        balance_loss = 0.5 * kl_loss + 0.5 * gini_loss
-        
-        # 5. 计算监控统计信息
-        with torch.no_grad():
-            # 记忆覆盖率：被选中的记忆条目占总数的比例
-            coverage_rate = (memory_counts > 0).float().mean().item()
-            
-            # 热点记忆：选择次数前10%的记忆条目
-            top10_threshold = torch.quantile(memory_counts, 0.9)
-            hot_memories = (memory_counts >= top10_threshold).sum().item()
-            
-            # 死记忆：从未被选中的记忆条目
-            dead_memories = (memory_counts == 0).sum().item()
-            
-            # 记忆选择方差（衡量不平衡程度）
-            selection_variance = memory_counts.var().item()
-            
-            stats = {
-                'gini_coefficient': gini_coeff.item(),
-                'kl_divergence': kl_loss.item(),
-                'coverage_rate': coverage_rate,
-                'hot_memories': hot_memories,
-                'dead_memories': dead_memories,
-                'selection_variance': selection_variance,
-                'max_selections': memory_counts.max().item(),
-                'min_selections': memory_counts.min().item(),
-            }
-        
-        return balance_loss, stats
-
-
-class GatedMemoryFusion(nn.Module):
-    """Gated MLP fusion for concatenated h_attn and selected memories"""
-    def __init__(self, config: LMConfig):
-        super().__init__()
-        self.config = config
-        self.dim = config.dim
-        self.knowledge_dim = config.knowledge_dim
-        self.num_selected = getattr(config, 'num_selected', 16)
-        
-        # 输入维度：dim (h_attn) + num_selected * knowledge_dim (选中的记忆)
-        # 实验1.4.6：记忆解码后立即压缩回knowledge_dim避免显存爆炸
-        concat_dim = self.dim + self.num_selected * self.dim
-        
-        # 类似SwiGLU的门控MLP结构
-        self.gate_proj = nn.Linear(concat_dim, self.dim, bias=False)
-        self.up_proj = nn.Linear(concat_dim, self.dim, bias=False)
-        self.down_proj = nn.Linear(self.dim, self.dim, bias=False)
-        
-        self.dropout = nn.Dropout(config.dropout)
-        
-    def forward(self, h_attn: torch.Tensor, selected_memories: torch.Tensor, memory_scores: torch.Tensor):
-        """
-        Args:
-            h_attn: [batch_size, seq_len, dim] - Self attention output
-            selected_memories: [batch_size, seq_len, num_selected, knowledge_dim] - Selected memory data
-            memory_scores: [batch_size, seq_len, num_selected] - Memory selection weights (not used in concatenation approach)
-        Returns:
-            output: [batch_size, seq_len, dim]
-        """
-        bsz, seq_len, _ = h_attn.shape
-        
-        # 将选中的记忆展平为一维向量
-        # [batch, seq_len, num_selected, knowledge_dim] -> [batch, seq_len, num_selected * knowledge_dim]
-        memory_flat = selected_memories.reshape(bsz, seq_len, -1)
-        
-        # 拼接h_attn和记忆信息
-        concat_input = torch.cat([h_attn, memory_flat], dim=-1)  # [batch, seq_len, dim + num_selected * dim]
-        
-        # 门控MLP处理（类似SwiGLU）
-        gate = F.silu(self.gate_proj(concat_input))  # [batch, seq_len, dim]
-        up = self.up_proj(concat_input)              # [batch, seq_len, dim]
-        fusion_output = gate * up                    # Element-wise multiplication
-        
-        # 输出投影
-        output = self.down_proj(fusion_output)       # [batch, seq_len, dim]
-        output = self.dropout(output)
-        
-        return output
-
-
-class MiniMindBlock(nn.Module):
-    """Transformer block with memory-based cross attention instead of FFN"""
-    def __init__(self, layer_id: int, config: LMConfig):
-        super().__init__()
-        self.config = config  # 保存config引用
-        self.n_heads = config.n_heads
-        self.dim = config.dim
-        self.head_dim = config.dim // config.n_heads
-        self.attention = Attention(config)
-        
-        self.layer_id = layer_id
-        self.attention_norm = RMSNorm(config.dim, eps=config.norm_eps)
-        self.memory_norm = RMSNorm(config.dim, eps=config.norm_eps)
-        
-        # 记忆相关模块
-        self.memory_gate = MemoryGate(config)
-        self.gated_memory_fusion = GatedMemoryFusion(config)
-
-        # self.attentionpool = nn.Linear(config.dim, 1)
-
-    def forward(self, x, pos_cis, memory_bank, tok_embeddings, collect_ema_stats=False):
-        """
-        Args:
-            x: [batch_size, seq_len, dim]
-            pos_cis: positional encoding
-            memory_bank: [knowledge_num, knowledge_dim] - shared memory bank
-            collect_ema_stats: 是否收集EMA更新统计信息
-        
-        Returns:
-            out: [batch_size, seq_len, dim]
-            balance_loss: 该层的平衡损失
-            layer_stats: 该层的监控统计信息
-            ema_stats: EMA更新统计信息（如果collect_ema_stats=True）
-            cosine_stats: 查找向量与选中记忆条目的余弦相似度统计信息
-        """
-        # Self attention
-        h_attn = self.attention(self.attention_norm(x), pos_cis)
-        h = x + h_attn
-        
-        # 使用h_attn作为门控和交叉注意力的输入（核心：self attention的输出）
-        h_for_memory = self.memory_norm(h_attn)
-        
-        # 门控选择记忆
-        memory_indices, memory_scores, balance_loss, layer_stats = self.memory_gate(h_for_memory)
-        
-        # 根据索引获取记忆数据 - 实验1.4.6：解码token_id为特征向量
-        bsz, seq_len, num_selected = memory_indices.shape
-        memory_indices_flat = memory_indices.view(-1)
-        selected_token_ids = memory_bank[memory_indices_flat]  # [batch * seq_len * num_selected, knowledge_length]
-        
-        # 解码token_ids为特征向量并立即压缩避免显存爆炸
-        selected_embeddings = tok_embeddings(selected_token_ids)  # [batch * seq_len * num_selected, knowledge_length, dim]
-        
-        # 立即压缩：knowledge_length * dim -> knowledge_dim 避免显存爆炸
-        # 使用平均池化压缩knowledge_length维度
-        pooled_memory = selected_embeddings.mean(dim=1)  # [batch * seq_len * num_selected, dim]
-        # attn_weights = self.attentionpool(selected_embeddings)  
-        # attn_weights = torch.softmax(attn_weights, dim=1)
-        # pooled_memory = torch.sum(selected_embeddings * attn_weights, dim=1)
-            
-        selected_memory = pooled_memory.view(bsz, seq_len, num_selected, self.dim)  # [batch, seq_len, num_selected, dim]
-        
-        # 门控MLP融合：串型连接h_attn和选中的记忆
-        memory_output = self.gated_memory_fusion(h_for_memory, selected_memory, memory_scores)
-        
-        # 残差连接
-        out = h + memory_output
-        
-        # 🔍 新增: 计算查找向量与选中记忆条目的余弦相似度
-        with torch.no_grad():
-            
-            # 扩展查找向量维度以匹配selected_memory
-            h_expanded = h_for_memory.unsqueeze(2).expand(-1, -1, num_selected, -1)  # [batch, seq_len, num_selected, dim]
-            
-            # 计算余弦相似度：cosine_sim(query, memory) for each selected memory
-            cosine_similarities = F.cosine_similarity(
-                h_expanded,           # [batch, seq_len, num_selected, dim]
-                selected_memory,      # [batch, seq_len, num_selected, knowledge_dim]
-                dim=-1               # 在knowledge_dim维度计算余弦相似度
-            )  # [batch, seq_len, num_selected]
-            
-            # 计算余弦相似度统计信息
-            cosine_stats = {
-                'cosine_similarities': cosine_similarities,                    # [batch, seq_len, num_selected]
-                'avg_cosine_similarity': cosine_similarities.mean().item(),   # 平均余弦相似度
-                'max_cosine_similarity': cosine_similarities.max().item(),    # 最大余弦相似度
-                'min_cosine_similarity': cosine_similarities.min().item(),    # 最小余弦相似度
-                'std_cosine_similarity': cosine_similarities.std().item(),    # 余弦相似度标准差
-            }
-        
-        # 收集EMA更新统计信息（仅在训练时且启用时）
-        ema_stats = None
-        if collect_ema_stats and self.training:
-            ema_stats = {
-                'memory_indices': memory_indices,      # [batch, seq_len, num_selected]
-                'memory_scores': memory_scores,        # [batch, seq_len, num_selected]  
-                'h_for_memory': h_for_memory,          # [batch, seq_len, dim]
-                'selected_memory': selected_memory,     # [batch, seq_len, num_selected, knowledge_dim]
-            }
-        
-        if collect_ema_stats:
-            return out, balance_loss, layer_stats, ema_stats, cosine_stats
-        else:
-            return out, balance_loss, layer_stats, cosine_stats
-
-
-class MiniMindLM(PreTrainedModel):
-    config_class = LMConfig
-
-    def __init__(self, params: LMConfig = None):
-        self.params = params or LMConfig()
-        super().__init__(self.params)
-        self.vocab_size, self.n_layers = params.vocab_size, params.n_layers
-        self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)
-        self.dropout = nn.Dropout(params.dropout)
-        self.layers = nn.ModuleList([MiniMindBlock(l, params) for l in range(self.n_layers)])
-        self.norm = RMSNorm(params.dim, eps=params.norm_eps)
-        self.output = nn.Linear(params.dim, params.vocab_size, bias=False)
-        self.tok_embeddings.weight = self.output.weight
-        self.register_buffer("pos_cis",
-                             precompute_pos_cis(dim=params.dim // params.n_heads, theta=params.rope_theta),
-                             persistent=False)
-        
-        # 初始化共享记忆库 - 实验1.4.6：存储token_id而非特征向量
-        # VQ-VAE风格：memory_bank作为codebook，使用EMA更新而非梯度更新
-        if params.use_ema_update:
-            self.memory_bank = nn.Parameter(
-                torch.randint(0, params.vocab_size, (params.knowledge_num, params.knowledge_length)),
-                requires_grad=False  # 禁用梯度更新，使用EMA更新
-            )
-        else:
-            self.memory_bank = nn.Parameter(
-                torch.randint(0, params.vocab_size, (params.knowledge_num, params.knowledge_length)),
-                requires_grad=True   # 传统梯度更新
-            )
-        
-        # EMA更新相关缓冲区
-        if params.use_ema_update:
-            # 记录每个memory条目的更新统计
-            self.register_buffer('ema_update_count', torch.zeros(params.knowledge_num), persistent=False)
-            # 注意：现在memory_bank存储token_id，但EMA在特征空间进行，所以不需要sum_buffer了
-            # self.register_buffer('ema_sum_buffer', torch.zeros_like(self.memory_bank), persistent=False)
-            # EMA更新频率计数器
-            self.register_buffer('ema_step_counter', torch.zeros(1, dtype=torch.long), persistent=False)
-        
-        # 记录上一步的记忆库状态，用于计算更新统计
-        self.register_buffer('prev_memory_bank', torch.zeros_like(self.memory_bank), persistent=False)
-        
-        # 🔥 新增: 冻结mask - 标记哪些memory_bank条目被冻结（不更新）
-        if params.freeze_ratio > 0.0:
-            freeze_num = int(params.knowledge_num * params.freeze_ratio)
-            freeze_mask = torch.zeros(params.knowledge_num, dtype=torch.bool)
-            # 固定冻结前面的条目
-            freeze_mask[:freeze_num] = True
-            self.register_buffer('freeze_mask', freeze_mask, persistent=False)
-            print(f"🔥 Memory bank freezing enabled: {freeze_num}/{params.knowledge_num} entries ({params.freeze_ratio*100:.1f}%) frozen")
-        else:
-            self.register_buffer('freeze_mask', torch.zeros(params.knowledge_num, dtype=torch.bool), persistent=False)
-            print(f"🔥 Memory bank freezing disabled: all entries can be updated")
-        
-        self.OUT = CausalLMOutputWithPast()
-    
-    def get_memory_update_stats(self):
-        """
-        计算记忆库更新统计信息
-        
-        Returns:
-            update_stats: 包含更新统计的字典
-        """
-        with torch.no_grad():
-            if hasattr(self, 'prev_memory_bank') and self.prev_memory_bank.numel() > 0:
-                # 计算L2距离变化
-                l2_distance = torch.norm(self.memory_bank - self.prev_memory_bank, p=2, dim=-1)
-                avg_l2_distance = l2_distance.mean().item()
-                max_l2_distance = l2_distance.max().item()
-                
-                # 计算余弦相似度
-                cos_sim = F.cosine_similarity(
-                    self.memory_bank.view(-1), 
-                    self.prev_memory_bank.view(-1), 
-                    dim=0
-                ).item()
-                
-                # 计算更新率（发生显著变化的记忆条目比例）
-                threshold = 0.01  # 更新阈值
-                updated_memories = (l2_distance > threshold).sum().item()
-                update_rate = updated_memories / self.memory_bank.size(0)
-                
-                update_stats = {
-                    'memory_avg_l2_change': avg_l2_distance,
-                    'memory_max_l2_change': max_l2_distance,
-                    'memory_cosine_similarity': cos_sim,
-                    'memory_update_rate': update_rate,
-                    'memory_updated_count': updated_memories
-                }
-            else:
-                # 第一次调用时的默认值
-                update_stats = {
-                    'memory_avg_l2_change': 0.0,
-                    'memory_max_l2_change': 0.0,
-                    'memory_cosine_similarity': 1.0,
-                    'memory_update_rate': 0.0,
-                    'memory_updated_count': 0
-                }
-            
-            # 更新prev_memory_bank
-            self.prev_memory_bank.copy_(self.memory_bank)
-            
-            return update_stats
-
-    def forward(self,
-                input_ids: Optional[torch.Tensor] = None,
-                **args):
-        """Forward pass without KV cache support"""
-        start_pos = args.get('start_pos', 0)
-        collect_ema_stats = args.get('collect_ema_stats', self.params.use_ema_update and self.training)
-        
-        h = self.dropout(self.tok_embeddings(input_ids))
-        pos_cis = self.pos_cis[start_pos:start_pos + input_ids.size(1)]
-        
-        # 收集所有层的平衡损失和统计信息
-        total_balance_loss = 0
-        all_layer_stats = {}
-        all_ema_stats = {}
-        all_cosine_stats = {}
-        
-        for layer_idx, layer in enumerate(self.layers):
-            if collect_ema_stats:
-                h, balance_loss, layer_stats, ema_stats, cosine_stats = layer(h, pos_cis, self.memory_bank, self.tok_embeddings, collect_ema_stats=True)
-                all_ema_stats[f'layer_{layer_idx}'] = ema_stats
-            else:
-                h, balance_loss, layer_stats, cosine_stats = layer(h, pos_cis, self.memory_bank, self.tok_embeddings, collect_ema_stats=False)
-            
-            total_balance_loss += balance_loss
-            # 为每层的统计信息添加前缀
-            for key, value in layer_stats.items():
-                all_layer_stats[f'layer_{layer_idx}_{key}'] = value
-            
-            # 为每层的余弦相似度统计信息添加前缀
-            for key, value in cosine_stats.items():
-                all_cosine_stats[f'layer_{layer_idx}_{key}'] = value
-
-        logits = self.output(self.norm(h))
-        
-        # 使用总的平衡损失作为aux_loss
-        aux_loss = total_balance_loss
-        
-        self.OUT.__setitem__('last_hidden_state', h)
-        self.OUT.__setitem__('logits', logits)
-        self.OUT.__setitem__('aux_loss', aux_loss)
-        self.OUT.__setitem__('layer_stats', all_layer_stats)  # 添加层级统计信息
-        self.OUT.__setitem__('ema_stats', all_ema_stats if collect_ema_stats else None)  # 添加EMA统计信息
-        self.OUT.__setitem__('cosine_stats', all_cosine_stats)  # 添加余弦相似度统计信息
-        self.OUT.__setitem__('past_key_values', None)  # 不支持KV cache
-        return self.OUT
-
-    @torch.inference_mode()
-    def generate(self, input_ids, eos_token_id=2, max_new_tokens=1024, temperature=0.75, top_p=0.90,
-                 stream=False, rp=1., pad_token_id=0, num_return_sequences=1, **args):
-        """Generate without KV cache"""
-        # 流式生成
-        if stream:
-            return self._stream(input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, **args)
-
-        # 直接生成
-        generated = []
-        for i in range(input_ids.size(0)):
-            non_pad = input_ids[i][input_ids[i] != pad_token_id].unsqueeze(0)
-            for _ in range(num_return_sequences):
-                out = self._stream(non_pad, eos_token_id, max_new_tokens, temperature, top_p, rp, **args)
-                tokens_list = [tokens[:, -1:] for tokens in out]
-                gen = torch.cat(tokens_list, dim=-1) if tokens_list else non_pad
-                full_sequence = torch.cat([non_pad, gen], dim=-1)
-                generated.append(full_sequence)
-
-        max_length = max(seq.size(1) for seq in generated)
-        generated = [
-            torch.cat(
-                [seq, torch.full((1, max_length - seq.size(1)), pad_token_id, dtype=seq.dtype, device=seq.device)],
-                dim=-1)
-            for seq in generated
-        ]
-        output = torch.cat(generated, dim=0)
-        res = output.view(input_ids.size(0) * num_return_sequences, -1)
-        return res
-
-    def _stream(self, input_ids, eos_token_id, max_new_tokens, temperature, top_p, rp, **args):
-        """Stream generation without KV cache - regenerates full sequence each time"""
-        start = input_ids.shape[1]
-        while input_ids.shape[1] < start + max_new_tokens:
-            # 每次都重新计算整个序列（因为没有KV cache）
-            out = self(input_ids, **args)
-            logits = out.logits[:, -1, :]
-            
-            # 重复惩罚
-            logits[:, list(set(input_ids.tolist()[0]))] /= rp
-            logits /= (temperature + 1e-9)
-            
-            # Top-p采样
-            if top_p is not None and top_p < 1.0:
-                sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
-                sorted_probs = F.softmax(sorted_logits, dim=-1)
-                cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
-                sorted_indices_to_remove = cumulative_probs > top_p
-                sorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()
-                sorted_indices_to_remove[:, 0] = False
-                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
-                logits[indices_to_remove] = -float('Inf')
-                
-            input_ids_next = torch.multinomial(F.softmax(logits, dim=-1), num_samples=1)
-            input_ids = torch.cat((input_ids, input_ids_next), dim=1)
-            yield input_ids[:, start:]
-            if input_ids_next.item() == eos_token_id:
-                break
-    
-    def apply_ema_update(self, ema_stats):
-        """
-        应用token-based EMA更新到memory_bank
-        实验1.4.6：批量化tensor操作优化版本
-        
-        Args:
-            ema_stats: 从forward pass收集的EMA统计信息，格式为：
-                {'layer_0': {'memory_indices': ..., 'h_for_memory': ...}, 'layer_1': ...}
-        """
-        if not self.params.use_ema_update:
-            return {}
-        
-        # 增加EMA步数计数器
-        self.ema_step_counter += 1
-        
-        # 检查是否需要进行EMA更新
-        if self.ema_step_counter % self.params.ema_update_freq != 0:
-            return {'ema_update_applied': False, 'reason': 'frequency_check_failed'}
-        
-        with torch.no_grad():
-            device = self.memory_bank.device
-            knowledge_num, knowledge_length = self.memory_bank.shape
-            dim = self.params.dim
-            
-            # 🚀 批量收集所有层的数据（避免字典操作）
-            all_indices = []
-            all_features = []
-            total_selections = 0
-            total_layers = 0
-            
-            # 收集所有层的EMA统计信息
-            for layer_ema_stats in ema_stats.values():
-                if layer_ema_stats is None:
-                    continue
-                    
-                total_layers += 1
-                memory_indices = layer_ema_stats['memory_indices']  # [batch, seq_len, num_selected]
-                h_for_memory = layer_ema_stats['h_for_memory']      # [batch, seq_len, dim]
-                
-                bsz, seq_len, num_selected = memory_indices.shape
-                total_selections += bsz * seq_len * num_selected
-                
-                # 展平索引和对应的h_for_memory
-                flat_indices = memory_indices.view(-1)  # [batch * seq_len * num_selected]
-                
-                # 为每个选择位置复制对应的h_for_memory
-                h_expanded = h_for_memory.unsqueeze(2).expand(-1, -1, num_selected, -1)  # [batch, seq_len, num_selected, dim]
-                flat_h = h_expanded.reshape(-1, dim)  # [batch * seq_len * num_selected, dim]
-                
-                all_indices.append(flat_indices)
-                all_features.append(flat_h)
-            
-            if not all_indices:
-                return {'ema_update_applied': False, 'reason': 'no_ema_stats'}
-                
-            # 🚀 合并所有数据
-            all_indices = torch.cat(all_indices, dim=0)  # [total_selections]
-            all_features = torch.cat(all_features, dim=0)  # [total_selections, dim]
-            
-            # 🚀 批量计算每个memory的平均特征（避免循环）
-            unique_indices, inverse_indices = torch.unique(all_indices, return_inverse=True)
-            
-            # 使用scatter_add批量聚合（确保数据类型一致）
-            aggregated_features = torch.zeros(unique_indices.size(0), dim, device=device, dtype=all_features.dtype)
-            count_per_memory = torch.zeros(unique_indices.size(0), device=device, dtype=all_features.dtype)
-            
-            aggregated_features.scatter_add_(0, inverse_indices.unsqueeze(1).expand(-1, dim), all_features)
-            count_per_memory.scatter_add_(0, inverse_indices, torch.ones_like(inverse_indices, dtype=all_features.dtype))
-            
-            # 计算平均值
-            avg_features = aggregated_features / count_per_memory.unsqueeze(1)  # [unique_count, dim]
-            
-            # 🚀 分批EMA更新（控制显存使用）
-            batch_size = 4096  # 每批处理4096个memory，控制显存
-            updated_memories = 0
-            
-            for i in range(0, unique_indices.size(0), batch_size):
-                end_i = min(i + batch_size, unique_indices.size(0))
-                batch_indices = unique_indices[i:end_i]
-                batch_avg_features = avg_features[i:end_i]
-                
-                # 当前批次的token解码
-                current_tokens_batch = self.memory_bank[batch_indices]  # [batch_size, knowledge_length]
-                current_embeddings_batch = self.tok_embeddings(current_tokens_batch.view(-1)).view(
-                    batch_indices.size(0), knowledge_length, dim)  # [batch_size, knowledge_length, dim]
-                
-                old_features_batch = current_embeddings_batch.view(batch_indices.size(0), -1)  # [batch_size, knowledge_length * dim]
-                expanded_new_features = batch_avg_features.repeat(1, knowledge_length)  # [batch_size, knowledge_length * dim]
-                
-                # EMA更新：new = γ * old + (1-γ) * new_avg
-                updated_features_batch = (
-                    self.params.ema_decay * old_features_batch + 
-                    (1 - self.params.ema_decay) * expanded_new_features
-                )
-                
-                # 分批编码为token_ids（关键：控制输出层的输入大小）
-                updated_reshaped = updated_features_batch.view(-1, dim)  # [batch_size * knowledge_length, dim]
-                logits_batch = self.output(updated_reshaped)  # [batch_size * knowledge_length, vocab_size]
-                new_token_ids_batch = torch.argmax(logits_batch, dim=-1).view(batch_indices.size(0), knowledge_length)
-                
-                # 🔥 新增: 应用冻结mask，只更新未冻结的条目
-                # 检查哪些batch_indices对应的条目没有被冻结
-                unfrozen_mask_batch = ~self.freeze_mask[batch_indices]  # [batch_size] - True表示未冻结
-                
-                # 只更新未冻结的条目
-                if unfrozen_mask_batch.any():
-                    unfrozen_indices = batch_indices[unfrozen_mask_batch]
-                    unfrozen_tokens = new_token_ids_batch[unfrozen_mask_batch]
-                    self.memory_bank[unfrozen_indices] = unfrozen_tokens
-                    updated_memories += unfrozen_indices.size(0)
-                else:
-                    # 如果这个batch中的所有条目都被冻结，则跳过更新
-                    pass
-            
-            update_ratio = updated_memories / knowledge_num
-            
-            # 🔥 新增: 计算冻结统计信息
-            frozen_count = self.freeze_mask.sum().item()
-            total_memories = knowledge_num
-            
-            update_stats = {
-                'ema_update_applied': True,
-                'ema_step': self.ema_step_counter.item(),
-                'total_selections': total_selections,
-                'total_layers': total_layers,
-                'updated_memories': updated_memories,
-                'update_ratio': update_ratio,
-                'frozen_memories': frozen_count,
-                'frozen_ratio': frozen_count / total_memories,
-                'ema_decay': self.params.ema_decay,
-                'selected_memory_coverage': updated_memories / knowledge_num,
-            }
-            
-            return update_stats

run_file/experiment_1_4_10.sh

@@ -1,414 +0,0 @@
-#!/bin/bash
-
-# ============================================================================
-# MiniMind 实验脚本 - Experiment 1.4.10
-# ============================================================================
-# 
-# 🎯 实验目标: 
-# 基于实验1.4.9，实现四损失系统：CE + Balance + Similarity + Diversity
-# 核心创新：Gumbel-Softmax选择机制 + 可微分相似度损失 + 候选集多样性约束
-# 
-# 使用方法:
-# bash run_file/experiment_1_4_10.sh
-# ============================================================================
-
-# ----------------------------------------------------------------------------
-# 🧑‍🔬 实验基本信息
-# ----------------------------------------------------------------------------
-EXPERIMENT_VERSION="1.4.10"
-EXPERIMENT_DESCRIPTION="四损失系统实验 - Gumbel-Softmax + 可微分相似度损失 + 多样性约束"
-RESEARCHER_NAME="AI Assistant"
-EXPERIMENT_DATE="$(date '+%Y-%m-%d %H:%M:%S')"
-
-# ----------------------------------------------------------------------------
-# 🤖 环境配置
-# ----------------------------------------------------------------------------
-
-# 调试和监控环境变量
-export NCCL_DEBUG=INFO
-export PYTHONFAULTHANDLER=1
-export CUDA_LAUNCH_BLOCKING=1
-
-# SwanLab 配置
-export SWANLAB_PROJECT="MiniMind-Experiment-1.4.10"
-
-# 日志配置
-LOG_DIR="out/experiment_${EXPERIMENT_VERSION}"
-mkdir -p "$LOG_DIR"
-LOG_FILE="$LOG_DIR/experiment.log"
-
-# ----------------------------------------------------------------------------
-# 🤖 硬件配置
-# ----------------------------------------------------------------------------
-CUDA_VISIBLE_DEVICES="0,1,2,3"
-NUM_PROCESSES="4"
-MIXED_PRECISION="bf16"
-MAIN_PROCESS_PORT="29500"
-
-# ----------------------------------------------------------------------------
-# 🤖 模型架构参数
-# ----------------------------------------------------------------------------
-MODEL_TYPE="model_memory"  # 🔥 使用Token-based Memory模型
-MODEL_SIZE="50.0"
-DIM="512"
-N_LAYERS="16"
-N_HEADS="32"
-MAX_SEQ_LEN="512"
-USE_MOE="false"
-
-# 🔥 知识库配置（四损失系统优化）
-KNOWLEDGE_NUM="1048576"  # 1M entries
-KNOWLEDGE_LENGTH="8"    # 🔥 增加到32个token提升表达能力
-KNOWLEDGE_DIM="128"      # 保留兼容性
-DISABLE_DB="false"
-
-# ----------------------------------------------------------------------------
-# 🤖 训练超参数
-# ----------------------------------------------------------------------------
-EPOCHS="3"
-EMBEDDING_EPOCH="42"
-BATCH_SIZE="4"          # 🔥 降低批次大小以适应更复杂的计算
-ACCUMULATION_STEPS="4"  # 🔥 增加累积步数保持有效批次大小
-LEARNING_RATE="2e-4"   # 🔥 适度降低学习率提升稳定性
-DTYPE="bfloat16"
-GRAD_CLIP="1.0"
-WARMUP_ITERS="0"
-
-# 🔥 四损失系统配置
-BALANCE_LOSS_COEF="0.01"      # 平衡损失系数
-SIMILARITY_LOSS_COEF="0.15"   # 🔥 相似度损失系数（核心损失）
-DIVERSITY_LOSS_COEF="0.08"    # 🔥 多样性损失系数（避免候选重复）
-
-# 数据和缓存路径
-DATA_PATH="/home/zym/Code/stable/merged_pretrain.jsonl"
-DATABASE_INIT_PATH="/home/zym/Code/stable/sentence_trex_data.json"
-CLUSTER_CACHE_PATH="None"  # 禁用聚类缓存
-VAL_DATA_PATH="dataset/stable/eval_data.json"
-
-# 训练配置
-NUM_WORKERS="8"
-LOG_INTERVAL="100"        # 🔥 更频繁的日志记录观察四个损失
-VAL_INTERVAL="100"
-SAVE_INTERVAL="10000"
-
-# 性能分析配置
-USE_PROFILE="true"
-PROFILE_INTERVAL="10"
-MEMORY_MONITOR_INTERVAL="100"
-
-# 高级功能
-USE_FLASH_ATTN="true"
-FAST_CLUSTERING="true"
-
-# 冻结率
-FREEZE_RATIO="0.2"
-
-# ----------------------------------------------------------------------------
-# 🤖 预检查函数
-# ----------------------------------------------------------------------------
-check_environment() {
-    echo "🔍 环境检查中..."
-    
-    # 检查GPU可用性
-    if ! nvidia-smi &> /dev/null; then
-        echo "❌ 错误: 未检测到GPU或nvidia-smi不可用"
-        exit 1
-    fi
-    
-    # 检查CUDA设备
-    if ! nvidia-smi -i "$CUDA_VISIBLE_DEVICES" &> /dev/null; then
-        echo "❌ 错误: GPU $CUDA_VISIBLE_DEVICES 不可用"
-        exit 1
-    fi
-    
-    # 检查数据文件
-    if [[ ! -f "$DATA_PATH" ]]; then
-        echo "❌ 错误: 训练数据文件不存在: $DATA_PATH"
-        exit 1
-    fi
-    
-    if [[ ! -f "$DATABASE_INIT_PATH" ]]; then
-        echo "❌ 错误: 数据库初始化文件不存在: $DATABASE_INIT_PATH"
-        exit 1
-    fi
-
-    
-    echo "✅ 环境检查通过"
-}
-
-# ----------------------------------------------------------------------------
-# 🤖 实验信息记录
-# ----------------------------------------------------------------------------
-log_experiment_info() {
-    echo "📝 记录实验信息..."
-    cat > "$LOG_DIR/experiment_info.txt" << EOF
-========================================
-MiniMind 实验信息
-========================================
-实验版本: $EXPERIMENT_VERSION
-实验描述: $EXPERIMENT_DESCRIPTION
-研究者: $RESEARCHER_NAME
-开始时间: $EXPERIMENT_DATE
-========================================
-硬件配置:
-GPU设备: $CUDA_VISIBLE_DEVICES
-进程数: $NUM_PROCESSES
-混合精度: $MIXED_PRECISION
-========================================
-模型配置:
-模型类型: $MODEL_TYPE (Token-based Memory + 四损失系统)
-模型大小: $MODEL_SIZE MB
-维度: $DIM
-层数: $N_LAYERS
-注意力头数: $N_HEADS
-最大序列长度: $MAX_SEQ_LEN
-知识库大小: $KNOWLEDGE_NUM (1M entries)
-知识长度: $KNOWLEDGE_LENGTH (增强表达能力)
-知识维度: $KNOWLEDGE_DIM (兼容性保留)
-========================================
-训练配置:
-训练轮次: $EPOCHS
-批次大小: $BATCH_SIZE (优化显存使用)
-学习率: $LEARNING_RATE (稳定性优化)
-梯度累积: $ACCUMULATION_STEPS (保持有效批次)
-数据类型: $DTYPE
-========================================
-🔥 四损失系统配置:
-平衡损失系数: $BALANCE_LOSS_COEF (记忆选择平衡)
-相似度损失系数: $SIMILARITY_LOSS_COEF (语义匹配优化)
-多样性损失系数: $DIVERSITY_LOSS_COEF (候选集多样性)
-========================================
-🔥 Gumbel-Softmax配置:
-候选项数量: 32 (Product Key生成)
-选择数量: 1 (Gumbel-Softmax选择最佳)
-温度参数: 1.0 (平衡探索与利用)
-选择机制: 硬选择 + Straight-Through Estimator
-========================================
-🔥 核心创新对比:
-传统方法: 16个记忆平均融合 (缺乏语义针对性)
-新方法: 32候选→1最佳 (语义相似度驱动)
-旧相似度损失: no_grad计算 (不可微分)
-新相似度损失: 可微分优化 (直接指导学习)
-新增多样性: 候选集内部差异性约束
-========================================
-数据路径:
-训练数据: $DATA_PATH
-验证数据: $VAL_DATA_PATH
-数据库初始化: $DATABASE_INIT_PATH
-聚类缓存: $CLUSTER_CACHE_PATH
-========================================
-预期改进:
-1. 相似度损失收敛: 从震荡→稳定下降
-2. 记忆选择质量: 更精准的语义匹配
-3. 生成文本质量: 更好的连贯性和相关性
-4. 四损失平衡: CE主导，其他损失辅助
-========================================
-EOF
-}
-
-# ----------------------------------------------------------------------------
-# 🤖 主执行函数
-# ----------------------------------------------------------------------------
-run_experiment() {
-    echo "🚀 开始执行实验 $EXPERIMENT_VERSION"
-    echo "📄 实验描述: $EXPERIMENT_DESCRIPTION"
-    echo "⏰ 开始时间: $EXPERIMENT_DATE"
-    
-    # 构建训练命令
-    local train_cmd="CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES accelerate launch --config_file accelerate_config.yaml train_pretrain_accelerate.py"
-    
-    # 添加训练参数
-    train_cmd+=" --out_dir \"$LOG_DIR\""
-    train_cmd+=" --epochs $EPOCHS"
-    train_cmd+=" --embedding_epoch $EMBEDDING_EPOCH"
-    train_cmd+=" --batch_size $BATCH_SIZE"
-    train_cmd+=" --learning_rate $LEARNING_RATE"
-    train_cmd+=" --dtype $DTYPE"
-    train_cmd+=" --num_workers $NUM_WORKERS"
-    train_cmd+=" --accumulation_steps $ACCUMULATION_STEPS"
-    train_cmd+=" --grad_clip $GRAD_CLIP"
-    train_cmd+=" --warmup_iters $WARMUP_ITERS"
-    train_cmd+=" --log_interval $LOG_INTERVAL"
-    train_cmd+=" --val_interval $VAL_INTERVAL"
-    train_cmd+=" --save_interval $SAVE_INTERVAL"
-    train_cmd+=" --dim $DIM"
-    train_cmd+=" --n_layers $N_LAYERS"
-    train_cmd+=" --n_heads $N_HEADS"
-    train_cmd+=" --max_seq_len $MAX_SEQ_LEN"
-    train_cmd+=" --data_path \"$DATA_PATH\""
-    train_cmd+=" --val_data_path \"$VAL_DATA_PATH\""
-    train_cmd+=" --knowledge_num $KNOWLEDGE_NUM"
-    train_cmd+=" --knowledge_length $KNOWLEDGE_LENGTH"
-    train_cmd+=" --database_init_path \"$DATABASE_INIT_PATH\""
-    train_cmd+=" --memory_monitor_interval $MEMORY_MONITOR_INTERVAL"
-    train_cmd+=" --model_type \"$MODEL_TYPE\""
-    train_cmd+=" --model_size $MODEL_SIZE"
-    train_cmd+=" --freeze_ratio $FREEZE_RATIO"
-    
-    # 🔥 四损失系统参数
-    train_cmd+=" --balance_loss_coef $BALANCE_LOSS_COEF"
-    train_cmd+=" --similarity_loss_coef $SIMILARITY_LOSS_COEF"
-    train_cmd+=" --diversity_loss_coef $DIVERSITY_LOSS_COEF"
-    
-    # 可选参数
-    if [[ "$USE_PROFILE" == "true" ]]; then
-        train_cmd+=" --profile"
-        train_cmd+=" --profile_interval $PROFILE_INTERVAL"
-    fi
-    
-    if [[ "$USE_FLASH_ATTN" == "true" ]]; then
-        train_cmd+=" --use_flash_attn"
-    fi
-    
-    if [[ "$FAST_CLUSTERING" == "true" ]]; then
-        train_cmd+=" --fast_clustering"
-    fi
-    
-    if [[ "$CLUSTER_CACHE_PATH" != "None" ]]; then
-        train_cmd+=" --cluster_cache_path \"$CLUSTER_CACHE_PATH\""
-    fi
-    
-    # SwanLab配置
-    train_cmd+=" --use_swanlab"
-    train_cmd+=" --swanlab_project \"$SWANLAB_PROJECT\""
-    # train_cmd+=" --swanlab_online True"
-    
-    echo "📋 执行命令:"
-    echo "$train_cmd"
-    echo
-    
-    # 记录命令到日志文件
-    echo "执行命令: $train_cmd" >> "$LOG_FILE"
-    echo "开始时间: $(date)" >> "$LOG_FILE"
-    
-    # 使用nohup执行训练（后台运行，输出写入日志文件）
-    echo "🔄 使用nohup后台运行训练，输出将写入日志文件: $LOG_FILE"
-    
-    # 创建训练脚本
-    train_script="/tmp/train_${EXPERIMENT_VERSION}.sh"
-    cat > "$train_script" << EOF
-#!/bin/bash
-# cd /home/pci/nas/AI_Large_Model_Team/ycz/Minimind
-# source .venv/bin/activate
-$train_cmd
-echo "结束时间: \$(date)"
-echo "退出代码: \$?"
-EOF
-    chmod +x "$train_script"
-    
-    # 使用nohup后台运行
-    nohup bash "$train_script" >> "$LOG_FILE" 2>&1 &
-    local train_pid=$!
-    
-    echo "🔥 训练进程已启动，PID: $train_pid"
-    echo "训练PID: $train_pid" >> "$LOG_FILE"
-    echo "训练脚本: $train_script" >> "$LOG_FILE"
-    
-    # 等待几秒确保进程启动
-    sleep 5
-    
-    # 检查进程是否还在运行
-    if kill -0 $train_pid 2>/dev/null; then
-        echo "✅ 训练进程正在后台运行"
-        echo "📋 实时查看日志: tail -f $LOG_FILE"
-        echo "📋 检查进程状态: ps -p $train_pid"
-        echo "🛑 停止训练: kill $train_pid"        
-        echo "📈 SwanLab: https://swanlab.cn/project/$SWANLAB_PROJECT"
-        echo ""
-        echo "🧠 四损失系统机制正在测试中..."
-        echo "   🔥 损失结构: CE + Balance + Similarity + Diversity"
-        echo "   🔥 候选机制: 32个候选 → Gumbel-Softmax选择1个最佳"
-        echo "   🔥 相似度损失: 可微分优化 (修复震荡问题)"
-        echo "   🔥 多样性约束: 候选集内部差异性正则化"
-        echo "   🔥 选择策略: 语义相似度驱动 vs 随机平均"
-        echo ""
-        echo "📊 与实验1.4.9对比:"
-        echo "   - 选择机制: 平均融合 → 最优选择"
-        echo "   - 相似度损失: 不可微分 → 可微分"  
-        echo "   - 候选多样性: 无约束 → 多样性正则化"
-        echo "   - 损失系统: 三损失 → 四损失平衡"
-        echo ""
-        echo "训练正在后台运行，可以安全关闭终端。"
-        echo ""
-        echo "🎯 预期改进:"
-        echo "   - 相似度损失: 稳定收敛 (不再震荡)"
-        echo "   - CE Loss: < 0.8 (改善语言建模)"
-        echo "   - 生成质量: 更连贯的文本输出"
-        echo "   - 记忆选择: 更精准的语义匹配"
-        echo ""
-        echo "⏱️ 预计训练时间: 18-22小时 (额外计算开销)"
-        echo "📊 预计GPU占用: ~24GB (Gumbel-Softmax + 多样性计算)"
-        echo ""
-        echo "🔍 关键监控指标:"
-        echo "   - Similarity Loss: 期望从1.9震荡→稳定下降"
-        echo "   - Diversity Loss: 保持适中值避免过度惩罚"
-        echo "   - Selection Entropy: 监控选择多样性"
-        echo "   - Selected Similarity: 观察选中记忆的相似度"
-        echo ""
-    else
-        echo "❌ 训练进程启动失败"
-        echo "📋 查看日志: $LOG_FILE"
-        exit 1
-    fi
-}
-
-# ----------------------------------------------------------------------------
-# 🤖 清理函数
-# ----------------------------------------------------------------------------
-cleanup() {
-    echo "🧹 清理临时文件..."
-    # 删除临时验证文件
-    rm -f /tmp/temp_val.jsonl
-}
-
-# ----------------------------------------------------------------------------
-# 🤖 信号处理
-# ----------------------------------------------------------------------------
-trap cleanup EXIT
-trap 'echo "❌ 实验被中断"; cleanup; exit 130' INT TERM
-
-# ----------------------------------------------------------------------------
-# 🤖 主程序入口
-# ----------------------------------------------------------------------------
-main() {
-    echo "============================================================================"
-    echo "🧠 MiniMind 预训练实验 1.4.10"
-    echo "🎯 四损失系统 - Gumbel-Softmax + 可微分相似度损失 + 多样性约束"
-    echo "============================================================================"
-    echo ""
-    echo "🔥 核心创新:"
-    echo "  ► 四损失架构: CE + Balance + Similarity + Diversity"
-    echo "  ► Gumbel-Softmax: 32候选→1最佳 (可微分离散选择)"
-    echo "  ► 相似度损失: 可微分优化 (修复震荡)"
-    echo "  ► 多样性约束: 候选集内部差异性正则化"
-    echo "  ► 语义选择: 相似度驱动 vs 平均融合"
-    echo ""
-    echo "🎯 实验假设:"
-    echo "  ✓ 可微分相似度损失解决震荡问题"
-    echo "  ✓ 语义驱动选择改善记忆利用质量"
-    echo "  ✓ 多样性约束防止候选集退化"
-    echo "  ✓ 四损失平衡提升整体模型性能"
-    echo ""
-    echo "🔧 关键技术细节:"
-    echo "  ► Straight-Through Estimator确保梯度流"
-    echo "  ► 候选集多样性通过余弦相似度矩阵计算"
-    echo "  ► Gumbel噪声增强选择随机性"
-    echo "  ► 硬选择保持离散性，软梯度保持可微性"
-    echo ""
-    echo "============================================================================"
-    
-    # 执行检查和初始化
-    check_environment
-    log_experiment_info
-    
-    # 运行实验
-    run_experiment
-    
-    echo "============================================================================"
-    echo "✅ 实验 $EXPERIMENT_VERSION 启动完成"
-    echo "📅 启动时间: $(date)"
-    echo "============================================================================"
-}
-
-# 执行主程序
-main "$@"

run_file/experiment_template.sh

@@ -37,13 +37,9 @@ EXPERIMENT_DATE="$(date '+%Y-%m-%d %H:%M:%S')"  # 自动记录实验开始时间
 # source "$VIRTUAL_ENV/bin/activate"
 
 # 调试和监控环境变量
+export NCCL_DEBUG=INFO                          # NCCL 调试信息
 export PYTHONFAULTHANDLER=1                     # Python 故障处理
-# export NCCL_DEBUG=INFO                        # NCCL 调试信息（仅调试时启用）
+export CUDA_LAUNCH_BLOCKING=1                   # CUDA 同步执行（调试用）
-# export CUDA_LAUNCH_BLOCKING=1                 # CUDA 同步执行（严重影响性能，仅调试时启用）
-
-# 🔥 强制禁用输出缓冲，确保日志立即写入（不影响GPU性能）
-export PYTHONUNBUFFERED=1                       # Python 解释器不缓冲输出
-export PYTHONIOENCODING=utf-8                   # 确保编码一致性
 
 # SwanLab 配置
 export SWANLAB_API_KEY="[SWANLAB_API_KEY]"     # 🤖 [AI构建] SwanLab API密钥
@@ -296,8 +292,8 @@ echo "退出代码: \$?"
 EOF
     chmod +x "$train_script"
     
-    # 使用nohup后台运行，并使用stdbuf禁用缓冲
+    # 使用nohup后台运行
-    nohup stdbuf -oL -eL bash "$train_script" >> "$LOG_FILE" 2>&1 &
+    nohup bash "$train_script" >> "$LOG_FILE" 2>&1 &
     local train_pid=$!
     
     echo "🔥 训练进程已启动，PID: $train_pid"

train_pretrain_accelerate.py

@@ -1,10 +1,6 @@
 import os
 # 设置环境变量 - 将wandb替换为SwanLab
 # os.environ["SWANLAB_MODE"] = "online"  # SwanLab使用在线模式
-
-# 🔥 强制禁用输出缓冲，确保日志立即写入
-os.environ['PYTHONUNBUFFERED'] = '1'  # Python 解释器不缓冲输出
-os.environ['PYTHONIOENCODING'] = 'utf-8'  # 确保编码一致性
 import platform
 import argparse
 from tqdm import tqdm
@@ -96,9 +92,7 @@ def log_memory_status(step, prefetch_batches, accelerator, stage="", detailed=Fa
 def Logger(msg, accelerator=None):
     # 如果没有提供accelerator，则只在主进程打印
     if accelerator is None or accelerator.is_main_process:
-        print(f"[{time.strftime('%Y-%m-%d %H:%M:%S')}] {msg}", flush=True)  # 强制刷新输出缓冲
+        print(f"[{time.strftime('%Y-%m-%d %H:%M:%S')}] {msg}")
-        import sys
-        sys.stdout.flush()  # 确保立即写入
 
 # Helper function to format seconds into HH:MM:SS
 def format_time(seconds):
@@ -1009,42 +1003,25 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
                 ).view(Y.size())
                 ce_loss = (ce_loss * loss_mask).sum() / loss_mask.sum()
                 
-                # 🔥 实验1.4.9: 四损失系统处理
+                # 获取平衡损失（如果模型支持）
                 balance_loss = 0
-                similarity_loss = 0
-                diversity_loss = 0
-                
                 if hasattr(res, 'aux_loss') and res.aux_loss is not None:
-                    aux_loss = res.aux_loss
+                    balance_loss = res.aux_loss
-                    if isinstance(aux_loss, dict):
-                        # 新的四损失结构
-                        balance_loss = aux_loss.get('balance_loss', 0)
-                        similarity_loss = aux_loss.get('similarity_loss', 0) 
-                        diversity_loss = aux_loss.get('diversity_loss', 0)
-                    else:
-                        # 向后兼容：旧的单一aux_loss
-                        balance_loss = aux_loss
 
                 # 获取余弦相似度统计信息（如果模型支持）
                 cosine_stats = {}
-                avg_selected_similarity = 0.0
                 if hasattr(res, 'cosine_stats') and res.cosine_stats is not None:
                     cosine_stats = res.cosine_stats
-                    # 🔥 使用选中记忆的平均相似度（更精确的指标）
+                # 计算平均余弦相似度
-                    selected_similarities = [v for k, v in cosine_stats.items() if k.endswith('_selected_avg_similarity')]
+                avg_cosine_similarity = 0.0
-                    if selected_similarities:
+                if cosine_stats:
-                        avg_selected_similarity = np.mean(selected_similarities)
+                    # 计算所有层的平均余弦相似度
+                    cosine_similarities = [v for k, v in cosine_stats.items() if k.endswith('_avg_cosine_similarity')]
+                    if cosine_similarities:
+                        avg_cosine_similarity = np.mean(cosine_similarities)
                 
-                # 🔥 四损失系统：CE + Balance + Similarity + Diversity
+                # 计算总损失
-                # 损失系数可以通过命令行参数调整
+                total_loss = ce_loss + args.balance_loss_coef * balance_loss + 2 * (1-avg_cosine_similarity)
-                balance_coef = getattr(args, 'balance_loss_coef', 0.01)
-                similarity_coef = getattr(args, 'similarity_loss_coef', 0.1) 
-                diversity_coef = getattr(args, 'diversity_loss_coef', 0.05)
-                
-                total_loss = (ce_loss + 
-                             balance_coef * balance_loss + 
-                             similarity_coef * similarity_loss +
-                             diversity_coef * diversity_loss)
                 loss = total_loss / args.accumulation_steps
 
             # 计时前向传播结束 (只在主进程进行)
@@ -1243,15 +1220,13 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
                     layer_stats = res.layer_stats
                 
 
-                # 🔥 构建四损失系统的日志字典
+                # 构建日志字典
                 log_dict = {
                     "epoch": epoch + 1,
                     "step": step + 1,
                     "total_steps_in_epoch": total_steps_in_epoch,
                     "train/loss_ce": ce_loss.item(),
                     "train/loss_balance": balance_loss.item() if isinstance(balance_loss, torch.Tensor) else balance_loss,
-                    "train/loss_similarity": similarity_loss.item() if isinstance(similarity_loss, torch.Tensor) else similarity_loss,
-                    "train/loss_diversity": diversity_loss.item() if isinstance(diversity_loss, torch.Tensor) else diversity_loss,
                     "train/loss_total": total_loss.item(),
                     "lr": current_lr,
                     "tokens_per_sec": tokens_per_sec,
@@ -1280,20 +1255,17 @@ def train_epoch(epoch, accelerator, model, train_loader, optimizer, scheduler, a
                         'memory/avg_coverage_rate': avg_coverage,
                         'memory/total_dead_memories': total_dead,
                         'memory/total_hot_memories': total_hot,
-                        'train/avg_selected_similarity': avg_selected_similarity,  # 🔥 使用选中记忆的相似度
+                        'train/avg_cosine_similarity': avg_cosine_similarity,
                     })
 
-                # 🔥 四损失系统的控制台输出
                 Logger(f"Epoch {epoch+1}/{args.epochs}, Step {step+1}/{total_steps_in_epoch}, "
-                      f"CE: {log_dict['train/loss_ce']:.4f}, "
+                      f"CE Loss: {log_dict['train/loss_ce']:.4f}, "
-                      f"Bal: {log_dict['train/loss_balance']:.4f}, "
+                      f"Balance Loss: {log_dict['train/loss_balance']:.4f}, "
-                      f"Sim: {log_dict['train/loss_similarity']:.4f}, "
+                      f"Avg Cosine Sim: {avg_cosine_similarity:.4f}, "
-                      f"Div: {log_dict['train/loss_diversity']:.4f}, "
+                      f"Total Loss: {log_dict['train/loss_total']:.4f}, "
-                      f"Total: {log_dict['train/loss_total']:.4f}, "
+                      f"Val Loss: {log_dict.get('val/loss', 'N/A')}, "
-                      f"Val: {log_dict.get('val/loss', 'N/A')}, "
                       f"LR: {log_dict['lr']:.6f}, "
                       f"Speed: {log_dict['tokens_per_sec']:.2f} tokens/sec | "
-                      f"Sel.Sim: {avg_selected_similarity:.4f} | "
                       f"Epoch Time Left: {format_time(epoch_remaining_time)} | "
                       f"Total Time Left: {format_time(total_remaining_time)}", accelerator)
 
@@ -1376,7 +1348,7 @@ def main():
     parser.add_argument("--knowledge_dim", type=int, default=128,help="知识库的向量维度")
     parser.add_argument("--database_init_path", type=str, default="/home/iomgaa/Code/Minimind/dataset/stable/sentence_trex_data.json", help="数据库初始化路径")
     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("--cluster_cache_path", type=str, default="/home/iomgaa/Code/Minimind/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）")
@@ -1384,11 +1356,8 @@ def main():
     parser.add_argument("--model_size", type=float, default=50.0, help="模型大小")
     parser.add_argument("--swanlab_online", type=bool, default=False, help="是否使用在线SwanLab服务")
     parser.add_argument("--balance_loss_coef", type=float, default=0.01, help="平衡损失系数")
-    parser.add_argument("--similarity_loss_coef", type=float, default=0.1, help="相似度损失系数（实验1.4.9）")
-    parser.add_argument("--diversity_loss_coef", type=float, default=0.05, help="多样性损失系数（实验1.4.9）") 
     parser.add_argument("--val_data_path", type=str, default="/home/zym/Code/stable/eval_data.json", help="验证数据集路径")
     parser.add_argument("--val_interval", type=int, default=100, help="验证评估间隔")
-    parser.add_argument("--freeze_ratio", type=float, default=0.2, help="冻结率")
     args = parser.parse_args()
  
     #########################################################
@@ -1427,8 +1396,7 @@ def main():
         flash_attn=args.use_flash_attn,
         knowledge_num=args.knowledge_num,
         knowledge_length=args.knowledge_length,
-        embeddings_epoch=args.embedding_epoch,
+        embeddings_epoch=args.embedding_epoch
-        freeze_ratio=args.freeze_ratio
     )
 
     #########################################################