# my_custom_olmoe/modeling_custom.py import torch import torch.nn as nn import torch.nn.functional as F # 导入官方实现(注意根据你的 transformers 版本调整导入路径) from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM, OlmoeSparseMoeBlock, OlmoeMLP from configuration_custom import DenseBackwardOLMoEConfig class DenseBackwardOlmoeSparseMoeBlock(OlmoeSparseMoeBlock): """ 继承自官方 OlmoeSparseMoeBlock,实现 dense backward 功能: 前向输出依旧保持与官方相同(即稀疏计算结果), 但在反向传播时,通过直通梯度让 dense 计算的梯度传递回来, dense 输出通过对每个专家在所有 token 上进行计算,并利用全 routing 权重加权获得。 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) """ 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) all_expert_outputs = torch.stack([expert(flat_hidden) for expert in self.experts], dim=1) # 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) 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) estimated = selected_outputs.mean(dim=0) else: estimated = all_expert_outputs[:, i, :].mean(dim=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 class DenseBackwardOLMoEForCausalLM(OlmoeForCausalLM): """ 自定义的 Olmoe ForCausalLM 模型,使用新的 DenseBackwardOlmoeSparseMoeBlock 替换原版的 MoE 模块, 以实现 dense backward 功能。 配置类:DenseBackwardOLMoEConfig """ config_class = DenseBackwardOLMoEConfig base_model_prefix = "olmoe" def __init__(self, config): # 首先调用父类初始化方法 super().__init__(config) # 不要尝试重新赋值self,而是从预训练模型加载并更新当前模型 pretrained_model = OlmoeForCausalLM.from_pretrained("allenai/OLMoE-1B-7B-0125") # 复制预训练模型的状态到当前模型 self.config = pretrained_model.config self.model = pretrained_model.model self.vocab_size = pretrained_model.vocab_size self.router_aux_loss_coef = pretrained_model.router_aux_loss_coef self.num_experts = pretrained_model.num_experts self.lm_head = pretrained_model.lm_head # 遍历模型中所有 decoder 层,替换每个 OlmoeSparseMoeBlock 为 DenseBackward 版本 # 此处假设官方模型在 self.model.layers 中组织 decoder 层, # 且每层中 mlp 模块包含属性 sparse_moe_block。 for layer in self.model.layers: if hasattr(layer.mlp, "gate"): print("111") orig_block = layer.mlp # 通过直接复制原版属性创建新的块 new_block = DenseBackwardOlmoeSparseMoeBlock(config) # 或其他适当参数 # 然后手动复制需要共享的属性: new_block.gate = orig_block.gate new_block.experts = orig_block.experts new_block.num_experts = orig_block.num_experts new_block.top_k = orig_block.top_k new_block.norm_topk_prob = orig_block.norm_topk_prob layer.mlp = new_block print(type(layer.mlp)) # 在调用post_init()前 test_param = self.model.layers[0].mlp.experts[0].up_proj.weight.data[0, 0].item() print(f"权重示例值(前): {test_param}") self.post_init() # 在调用post_init()后 test_param_after = self.model.layers[0].mlp.experts[0].up_proj.weight.data[0, 0].item() print(f"权重示例值(后): {test_param_after}") def main(): config = DenseBackwardOLMoEConfig( # 官方模型参数 model_marker="DenseBackward_olmoe_marker", ) # 创建自定义模型实例 model = DenseBackwardOLMoEForCausalLM(config) print(type(model)) print(type(model.model)) print(type(model.model.layers[0])) print(type(model.model.layers[0].mlp)) print(type(model.model.layers[0].mlp.experts)) if __name__ == "__main__": main()