Update modeling_densebackward_olmoe0125.py

modeling_densebackward_olmoe0125.py

@@ -23,140 +23,173 @@ class DenseBackwardOlmoeSparseMoeBlock(OlmoeSparseMoeBlock):
         router_logits: Tensor, shape (batch_size * sequence_length, num_experts)
     """
     def forward(self, hidden_states: torch.Tensor):
-        """
-        输入:
-        hidden_states: Tensor, shape (batch_size, sequence_length, hidden_dim)
-        输出:
-            final_output: Tensor, shape (batch_size, sequence_length, hidden_dim)
-            router_logits: Tensor, shape (batch_size * sequence_length, num_experts)
-        实现思路：
-          1. 将输入展平为 (B*seq_len, hidden_dim)，通过 self.gate 得到 router_logits，
-             并计算全专家的 routing 权重（softmax 后）。
-          2. 对 routing 权重取 top-k，得到 routing_weights_topk 与 selected_experts；
-             如配置要求，归一化 top-k 概率。
-          3. 稀疏计算部分：仅计算每个 token 对于 top-k 专家的输出，
-             并累加得到 sparse_output（保留原版计算流程，同时记录激活专家的实际输出）。
-          4. Dense 估计部分：先计算所有专家对所有 token 的输出（all_expert_outputs），
-             再逐 token 调用 estimate_dense_output 得到 dense 输出（dense_estimated）。
-          5. 使用直通梯度技巧：前向输出用 sparse_output，但梯度来源于 dense_estimated。
-          6. 最后 reshape 为 (batch_size, sequence_length, hidden_dim) 并返回 final_output 及 router_logits.
-        """
-        #determine the shape of hidden_states
-        batch_size, seq_length, hidden_dim = hidden_states.shape
-        flat_hidden = hidden_states.view(-1, hidden_dim)  # (B*seq_len, hidden_dim)
-
-        # 计算路由 logits 和全专家 routing 权重
-        router_logits = self.gate(flat_hidden)  # (B*seq_len, num_experts)
-        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)  # (B*seq_len, num_experts)
-
-        # Top-k 选择
-        routing_weights_topk, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
-        if self.norm_topk_prob:
-            routing_weights_topk = routing_weights_topk / routing_weights_topk.sum(dim=-1, keepdim=True)
-        routing_weights_topk = routing_weights_topk.to(flat_hidden.dtype)
-
-        # ---------- 稀疏计算部分 ----------
-        # 初始化稀疏输出，shape: (B*seq_len, hidden_dim)
-        sparse_output = torch.zeros((flat_hidden.size(0), hidden_dim), dtype=flat_hidden.dtype, device=flat_hidden.device)
-        # 用于记录每个 token 对激活专家的实际输出
-        activated_outputs = [{} for _ in range(flat_hidden.size(0))]
-        # one-hot 编码 top-k 专家，shape: (B*seq_len, top_k, num_experts)
-        expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)  # (B*seq_len, top_k, num_experts)
-        expert_mask = expert_mask.permute(2, 1, 0)  # (num_experts, top_k, B*seq_len)
-
-        for expert_idx in range(self.num_experts):
-            expert_layer = self.experts[expert_idx]
-            idx, top_x = torch.where(expert_mask[expert_idx])
-            if top_x.numel() > 0:
-                current_state = flat_hidden[top_x]  # (n, hidden_dim)
-                current_output = expert_layer(current_state)  # (n, hidden_dim)
-                weight = routing_weights_topk[top_x, idx].unsqueeze(-1)  # (n, 1)
-                weighted_output = current_output * weight
-                sparse_output.index_add_(0, top_x, weighted_output.to(flat_hidden.dtype))
-                # 保存当前 token 对该专家的实际输出
-                for pos, token_idx in enumerate(top_x.tolist()):
-                    activated_outputs[token_idx][expert_idx] = current_output[pos]
-        # ---------- 稀疏计算结束 ----------
-
-        # ---------- Dense估计部分 ----------
-        # 计算所有专家对所有 token 的 dense 输出，shape: (B*seq_len, num_experts, hidden_dim)
-        # 创建全零张量，只填入已激活专家的输出
-        all_expert_outputs = torch.zeros((flat_hidden.size(0), self.num_experts, hidden_dim), 
+    # determine the shape of hidden_states
+    batch_size, seq_length, hidden_dim = hidden_states.shape
+    flat_hidden = hidden_states.view(-1, hidden_dim)  # (B*seq_len, hidden_dim)
+    total_tokens = flat_hidden.size(0)
+
+    # 计算路由 logits 和全专家 routing 权重
+    router_logits = self.gate(flat_hidden)  # (B*seq_len, num_experts)
+    routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)  # (B*seq_len, num_experts)
+
+    # Top-k 选择
+    routing_weights_topk, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+    if self.norm_topk_prob:
+        routing_weights_topk = routing_weights_topk / routing_weights_topk.sum(dim=-1, keepdim=True)
+    routing_weights_topk = routing_weights_topk.to(flat_hidden.dtype)
+
+    # ---------- 稀疏计算部分 ----------
+    # 初始化稀疏输出，shape: (B*seq_len, hidden_dim)
+    sparse_output = torch.zeros((total_tokens, hidden_dim), dtype=flat_hidden.dtype, device=flat_hidden.device)
+    
+    # 创建一个张量存储激活专家的输出，避免使用Python字典
+    # shape: (B*seq_len, num_experts, hidden_dim)
+    all_expert_outputs = torch.zeros((total_tokens, self.num_experts, hidden_dim), 
                                    dtype=flat_hidden.dtype, device=flat_hidden.device)
-        # 填入已激活专家的输出
-        for i in range(flat_hidden.size(0)):
-            for expert_idx in activated_outputs[i].keys():
-                all_expert_outputs[i, expert_idx] = activated_outputs[i][expert_idx]
-        # 将 selected_experts 转换为 list，每个 token 的激活专家列表
-        all_routing = selected_experts.tolist()  # 长度为 (B*seq_len)
+    
+    # 使用张量掩码跟踪哪些专家被激活
+    # shape: (B*seq_len, num_experts)
+    expert_activated = torch.zeros((total_tokens, self.num_experts), 
+                                  dtype=torch.bool, device=flat_hidden.device)
+    
+    # one-hot 编码 top-k 专家，shape: (B*seq_len, top_k, num_experts)
+    expert_mask = F.one_hot(selected_experts, num_classes=self.num_experts)  # (B*seq_len, top_k, num_experts)
+    expert_mask = expert_mask.permute(2, 1, 0)  # (num_experts, top_k, B*seq_len)
+
+    for expert_idx in range(self.num_experts):
+        expert_layer = self.experts[expert_idx]
+        idx, top_x = torch.where(expert_mask[expert_idx])
+        if top_x.numel() > 0:
+            current_state = flat_hidden[top_x]  # (n, hidden_dim)
+            current_output = expert_layer(current_state)  # (n, hidden_dim)
+            weight = routing_weights_topk[top_x, idx].unsqueeze(-1)  # (n, 1)
+            weighted_output = current_output * weight
+            sparse_output.index_add_(0, top_x, weighted_output.to(flat_hidden.dtype))
+            
+            # 保存专家输出到张量中，而不是使用字典
+            all_expert_outputs.index_copy_(0, top_x, 
+                                         torch.zeros_like(all_expert_outputs[0:top_x.size(0)]).scatter_(
+                                             1, expert_idx * torch.ones((top_x.size(0), 1), 
+                                                                     dtype=torch.long, 
+                                                                     device=flat_hidden.device), 
+                                             current_output.unsqueeze(1)))
+            
+            # 标记哪些专家被激活
+            expert_activated.index_copy_(0, top_x,
+                                       torch.zeros_like(expert_activated[0:top_x.size(0)]).scatter_(
+                                           1, expert_idx * torch.ones((top_x.size(0), 1), 
+                                                                   dtype=torch.long, 
+                                                                   device=flat_hidden.device), 
+                                           torch.ones((top_x.size(0), 1), 
+                                                    dtype=torch.bool, 
+                                                    device=flat_hidden.device)))
+    # ---------- 稀疏计算结束 ----------
+
+    # ---------- Dense估计部分 ----------
+    # 从GPU获取必要信息，避免过多的tensor->list转换
+    selected_experts_gpu = selected_experts  # 保持在GPU上
+
+    # 预分配结果张量，避免在循环中append
+    dense_outputs = torch.zeros_like(sparse_output)
+    
+    # 使用向量化的estimate_dense_output函数
+    dense_outputs = self.estimate_dense_output_batch(
+        total_tokens=total_tokens,
+        selected_experts=selected_experts_gpu,
+        routing_weights=routing_weights,
+        expert_activated=expert_activated,
+        all_expert_outputs=all_expert_outputs
+    )
+    # ---------- Dense估计结束 ----------
+
+    # 使用直通梯度：前向输出用稀疏结果，但反向传播时梯度来源于 dense 估计
+    final_flat = sparse_output.detach() + (dense_outputs - dense_outputs.detach())
+    final_output = final_flat.view(batch_size, seq_length, hidden_dim)
+    return final_output, router_logits
+
+def estimate_dense_output_batch(self, total_tokens, selected_experts, routing_weights, 
+                              expert_activated, all_expert_outputs):
+    """
+    批量估计所有token的dense输出，优化版本。
+    
+    参数：
+        total_tokens: token总数
+        selected_experts: 每个token激活的专家索引，形状 (total_tokens, top_k)
+        routing_weights: 路由权重，形状 (total_tokens, num_experts)
+        expert_activated: 掩码张量，标记每个token激活了哪些专家，形状 (total_tokens, num_experts)
+        all_expert_outputs: 专家输出，形状 (total_tokens, num_experts, hidden_dim)
         
-        dense_outputs = []
-        for i in range(flat_hidden.size(0)):
-            dense_est = self.estimate_dense_output(
-                token_idx=i,
-                activated=all_routing[i],              # 当前 token 激活的专家列表，例如 [a, b]
-                gate_prob=routing_weights[i],            # 当前 token 的完整 routing 权重 (num_experts,)
-                activated_outputs=activated_outputs[i],  # 当前 token 对激活专家的实际输出
-                all_routing=all_routing,                 # 全 batch 每个 token 的激活专家列表（list of lists）
-                all_expert_outputs=all_expert_outputs      # (B*seq_len, num_experts, hidden_dim)
-            )
-            dense_outputs.append(dense_est.unsqueeze(0))
-        dense_outputs = torch.cat(dense_outputs, dim=0)  # (B*seq_len, hidden_dim)
-        # ---------- Dense估计结束 ----------
-
-        # 使用直通梯度：前向输出用稀疏结果，但反向传播时梯度来源于 dense 估计
-        final_flat = sparse_output.detach() + (dense_outputs - dense_outputs.detach())
-        final_output = final_flat.view(batch_size, seq_length, hidden_dim)
-        return final_output, router_logits
-
-    def estimate_dense_output(self, token_idx, activated, gate_prob, activated_outputs, all_routing, all_expert_outputs):
-        """
-        对于当前 token，根据 mini-batch 中的信息估计 dense 输出。
-        参数：
-            token_idx: 当前 token 的索引（标量）
-            activated: 当前 token 激活的专家列表，例如 [1, 3]
-            gate_prob: 当前 token 的 routing 权重，形状 (num_experts,)
-            activated_outputs: dict，当前 token 对激活专家的实际输出，形状 (hidden_dim,)
-            all_routing: list，每个 token 的激活专家列表（长度为 N，每个元素为 list）
-            all_expert_outputs: Tensor, (N, num_experts, hidden_dim)
-        返回：
-            estimated_dense: Tensor, (hidden_dim,)
-        """
-        num_experts = gate_prob.size(0)
-        dense_parts = {}
-        # 对于激活的专家，直接使用其实际输出
-        for idx in activated:
-            dense_parts[idx] = activated_outputs[idx]
-        # 对于未激活的专家，使用 mini-batch 中其他 token 的输出估计
-        non_activated = [i for i in range(num_experts) if i not in activated]
-        for i in non_activated:
-            indices = []
-            for idx, r_dec in enumerate(all_routing):
-                if (i in r_dec) and (len(set(r_dec) & set(activated)) > 0):
-                    indices.append(idx)
-            if indices:
-                selected_outputs = all_expert_outputs[indices, i, :]  # (n, hidden_dim)
-                 # 只计算非零值的平均值
-                mask = (selected_outputs.sum(dim=-1) != 0).to(selected_outputs.dtype).unsqueeze(-1)
-                if mask.sum() > 0:
-                    estimated = (selected_outputs * mask).sum(dim=0) / mask.sum()
-                else:
-                   # 如果全是零，返回零向量
-                    estimated = torch.zeros_like(selected_outputs[0])
+    返回：
+        dense_outputs: 形状 (total_tokens, hidden_dim)
+    """
+    hidden_dim = all_expert_outputs.size(-1)
+    num_experts = routing_weights.size(1)
+    device = all_expert_outputs.device
+    
+    # 预分配结果张量
+    dense_outputs = torch.zeros((total_tokens, hidden_dim), dtype=all_expert_outputs.dtype, device=device)
+    
+    # 对每个token单独处理（此处仍需循环，但后续可进一步优化）
+    for token_idx in range(total_tokens):
+        # 对于激活的专家，直接使用输出
+        activated_mask = expert_activated[token_idx]  # (num_experts,)
+        
+        # 对于未激活的专家，找到估计值
+        for expert_idx in range(num_experts):
+            if activated_mask[expert_idx]:
+                # 直接使用激活专家的输出
+                expert_output = all_expert_outputs[token_idx, expert_idx]
             else:
-                all_outputs = all_expert_outputs[:, i, :]
-                mask = (all_outputs.sum(dim=-1) != 0).to(all_outputs.dtype).unsqueeze(-1)
-                if mask.sum() > 0:
-                    estimated = (all_outputs * mask).sum(dim=0) / mask.sum()
+                # 寻找可以用于估计的输出
+                # 找出其他激活了当前专家的token
+                tokens_with_expert = expert_activated[:, expert_idx]
+                
+                # 找出同时激活了当前token的某些专家和当前专家的其他token
+                # 首先获取当前token激活的专家
+                current_activated = selected_experts[token_idx]
+                
+                # 在其他token中寻找同时激活了current_activated中专家和expert_idx的token
+                valid_tokens = torch.zeros(total_tokens, dtype=torch.bool, device=device)
+                
+                # 对于每个其他token，检查它是否同时激活了当前token的某个专家和当前专家
+                for other_token in range(total_tokens):
+                    if other_token == token_idx:
+                        continue
+                    
+                    # 检查其他token是否激活了当前专家
+                    if expert_activated[other_token, expert_idx]:
+                        # 检查是否有共同激活的专家
+                        other_experts = selected_experts[other_token]
+                        common = torch.any(torch.isin(other_experts, current_activated))
+                        if common:
+                            valid_tokens[other_token] = True
+                
+                # 如果找到了有效token
+                if valid_tokens.any():
+                    # 获取有效token对当前专家的输出
+                    valid_outputs = all_expert_outputs[valid_tokens, expert_idx]
+                    # 只计算非零值的平均值
+                    mask = (valid_outputs.sum(dim=-1) != 0).to(valid_outputs.dtype).unsqueeze(-1)
+                    if mask.sum() > 0:
+                        expert_output = (valid_outputs * mask).sum(dim=0) / mask.sum()
+                    else:
+                        expert_output = torch.zeros(hidden_dim, dtype=all_expert_outputs.dtype, device=device)
                 else:
-                    # 如果全是零，返回零向量
-                    estimated = torch.zeros_like(all_outputs[0])
-            dense_parts[i] = estimated
-        # 按 gate_prob 加权求和各专家输出
-        estimated_dense = 0
-        for i in range(num_experts):
-            estimated_dense += gate_prob[i] * dense_parts[i]
-        return estimated_dense
+                    # 如果没有找到有效token，使用所有激活了当前专家的token的输出
+                    if tokens_with_expert.any():
+                        all_valid_outputs = all_expert_outputs[tokens_with_expert, expert_idx]
+                        mask = (all_valid_outputs.sum(dim=-1) != 0).to(all_valid_outputs.dtype).unsqueeze(-1)
+                        if mask.sum() > 0:
+                            expert_output = (all_valid_outputs * mask).sum(dim=0) / mask.sum()
+                        else:
+                            expert_output = torch.zeros(hidden_dim, dtype=all_expert_outputs.dtype, device=device)
+                    else:
+                        expert_output = torch.zeros(hidden_dim, dtype=all_expert_outputs.dtype, device=device)
+            
+            # 根据routing权重加权
+            dense_outputs[token_idx] += routing_weights[token_idx, expert_idx] * expert_output
+    
+    return dense_outputs
 
 
 class DenseBackwardOLMoEForCausalLM(OlmoeForCausalLM):