Add Mixture-of-Experts (MoE) layer support across modeling pipeline

Makefile

@@ -18,6 +18,14 @@ USE_SWA_FLAG = $(if $(USE_SWA),--use-swa,)
 PARALLEL_FLAG = $(if $(PARALLEL),--parallel,)
 INIT_PRETRAIN_FLAG = $(if $(NO_PRETRAIN),,--init-from-pretrain $(or $(INIT_PRETRAIN),models/pretrain_delivery.pt))
 
+# --- MoE 相关 flag ---
+USE_MOE_FLAG       = $(if $(USE_MOE),--use-moe,)
+MOE_EXPERTS_FLAG   = $(if $(MOE_EXPERTS),--moe-n-experts $(MOE_EXPERTS),)
+MOE_TOPK_FLAG      = $(if $(MOE_TOPK),--moe-top-k $(MOE_TOPK),)
+MOE_HIDDEN_FLAG    = $(if $(MOE_HIDDEN_MULT),--moe-expert-hidden-mult $(MOE_HIDDEN_MULT),)
+MOE_JITTER_FLAG    = $(if $(MOE_JITTER),--moe-jitter-noise $(MOE_JITTER),)
+MOE_FLAGS          = $(USE_MOE_FLAG) $(MOE_EXPERTS_FLAG) $(MOE_TOPK_FLAG) $(MOE_HIDDEN_FLAG) $(MOE_JITTER_FLAG)
+
 #################################################################################
 # ENVIRONMENT & CODE QUALITY                                                    #
 #################################################################################
@@ -117,10 +125,12 @@ pretrain: requirements
 train: requirements
 	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.nested_cv_optuna \
 		$(DEVICE_FLAG) $(MPNN_FLAG) $(SEED_FLAG) $(INIT_PRETRAIN_FLAG) \
-		$(N_TRIALS_FLAG) $(EPOCHS_PER_TRIAL_FLAG) $(MIN_STRATUM_FLAG) $(OUTPUT_DIR_FLAG) $(PARALLEL_FLAG)
+		$(N_TRIALS_FLAG) $(EPOCHS_PER_TRIAL_FLAG) $(MIN_STRATUM_FLAG) $(OUTPUT_DIR_FLAG) $(PARALLEL_FLAG) \
+		$(MOE_FLAGS)
 	$(PYTHON_INTERPRETER) -m lnp_ml.modeling.final_train_optuna_cv \
 		$(DEVICE_FLAG) $(MPNN_FLAG) $(SEED_FLAG) $(INIT_PRETRAIN_FLAG) \
-		$(N_TRIALS_FLAG) $(EPOCHS_PER_TRIAL_FLAG) $(MIN_STRATUM_FLAG) $(OUTPUT_DIR_FLAG) $(USE_SWA_FLAG)
+		$(N_TRIALS_FLAG) $(EPOCHS_PER_TRIAL_FLAG) $(MIN_STRATUM_FLAG) $(OUTPUT_DIR_FLAG) $(USE_SWA_FLAG) \
+		$(MOE_FLAGS)
 		
 #################################################################################
 # INTERPRETABILITY  (biodistribution feature importance)                       #

lnp_ml/interpretability/token_importance.py

@@ -41,8 +41,12 @@ BIODIST_ORGANS = ["lymph_nodes", "heart", "liver", "spleen", "lung", "kidney", "
 
 def get_token_names(model: Union[LNPModel, LNPModelWithoutMPNN]) -> List[str]:
     if model.use_mpnn:
-        return ["mpnn", "morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
-    return ["morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
+        names = ["mpnn", "morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
+    else:
+        names = ["morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
+    if getattr(model, "moe", None) is not None:
+        names.append("moe")
+    return names
 
 
 # ──────────────────────────────────────────────────────────────────────
@@ -192,7 +196,8 @@ def fusion_attention_importance(
     all_weights: List[torch.Tensor] = []
     for start in range(0, all_tokens.size(0), batch_size):
         end = min(start + batch_size, all_tokens.size(0))
-        attended = model.cross_attention(all_tokens[start:end].to(device))
+        # 使用与 forward_from_projected 相同的桥梁，自动处理 MoE 的额外 token
+        attended = model._attended_with_moe(all_tokens[start:end].to(device))
         _, weights = model.fusion(attended, return_attn_weights=True)
         all_weights.append(weights.cpu())
 
@@ -295,7 +300,7 @@ def plot_token_importance(
         vals_sorted = normed[order]
 
         n_tokens = len(token_names)
-        split_idx = 4 if n_tokens == 8 else 3
+        split_idx = 4 if "mpnn" in token_names else 3
         channel_a_set = set(token_names[:split_idx])
         colors = [color_a if n in channel_a_set else color_b for n in names_sorted]
 

lnp_ml/modeling/final_train_optuna_cv.py

@@ -176,8 +176,22 @@ def create_model(
     dropout: float = 0.1,
     use_mpnn: bool = False,
     mpnn_device: str = "cpu",
+    # ============ MoE 相关（新增） ============
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
 ) -> Union[LNPModel, LNPModelWithoutMPNN]:
     """创建模型"""
+    moe_kwargs = dict(
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
+    )
+
     if use_mpnn:
         ensemble_paths = find_mpnn_ensemble_paths()
         return LNPModel(
@@ -189,6 +203,7 @@ def create_model(
             dropout=dropout,
             mpnn_ensemble_paths=ensemble_paths,
             mpnn_device=mpnn_device,
+            **moe_kwargs,
         )
     else:
         return LNPModelWithoutMPNN(
@@ -198,6 +213,7 @@ def create_model(
             fusion_strategy=fusion_strategy,
             head_hidden_dim=head_hidden_dim,
             dropout=dropout,
+            **moe_kwargs,
         )
 
 
@@ -232,7 +248,6 @@ def load_pretrain_weights_to_model(
 
 
 # ============ 3-fold Optuna 调参 ============
-
 def run_optuna_cv(
     full_dataset: LNPDataset,
     strata: np.ndarray,
@@ -249,6 +264,11 @@ def run_optuna_cv(
     pretrain_config: Optional[Dict] = None,
     load_delivery_head: bool = True,
     rdkit_cache: Optional[Dict] = None,
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
 ) -> Tuple[Dict, int, optuna.Study]:
     """
     使用全量数据做 3-fold CV Optuna 超参搜索。
@@ -335,6 +355,11 @@ def run_optuna_cv(
                 dropout=dropout,
                 use_mpnn=use_mpnn,
                 mpnn_device=device.type,
+                use_moe=use_moe,
+                moe_n_experts=moe_n_experts,
+                moe_top_k=moe_top_k,
+                moe_expert_hidden_mult=moe_expert_hidden_mult,
+                moe_jitter_noise=moe_jitter_noise,
             )
             if rdkit_cache is not None:
                 model.rdkit_encoder._cache = rdkit_cache
@@ -404,7 +429,6 @@ def run_optuna_cv(
 
 
 # ============ 主流程 ============
-
 @app.command()
 def main(
     input_path: Path = INTERIM_DATA_DIR / "internal.csv",
@@ -427,6 +451,12 @@ def main(
     load_delivery_head: bool = False,
     # MPNN
     use_mpnn: bool = False,
+    # MoE（新增）
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
     # 设备
     device: str = "cuda" if torch.cuda.is_available() else "cpu",
 ):
@@ -507,6 +537,11 @@ def main(
         pretrain_config=pretrain_config,
         load_delivery_head=load_delivery_head,
         rdkit_cache=rdkit_cache,
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
     )
     
     # 保存最佳参数
@@ -564,6 +599,11 @@ def main(
         dropout=best_params["dropout"],
         use_mpnn=use_mpnn,
         mpnn_device=device.type,
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
     )
     model.rdkit_encoder._cache = rdkit_cache
     
@@ -611,6 +651,11 @@ def main(
         "head_hidden_dim": best_params["head_hidden_dim"],
         "dropout": best_params["dropout"],
         "use_mpnn": use_mpnn,
+        "use_moe": use_moe,
+        "moe_n_experts": moe_n_experts,
+        "moe_top_k": moe_top_k,
+        "moe_expert_hidden_mult": moe_expert_hidden_mult,
+        "moe_jitter_noise": moe_jitter_noise,
     }
 
     torch.save({

lnp_ml/modeling/layers/__init__.py

@@ -1,6 +1,6 @@
 from lnp_ml.modeling.layers.token_projector import TokenProjector
 from lnp_ml.modeling.layers.bidirectional_cross_attention import CrossModalAttention
 from lnp_ml.modeling.layers.fusion import FusionLayer
+from lnp_ml.modeling.layers.moe import MoEBlock
 
-__all__ = ["TokenProjector", "CrossModalAttention", "FusionLayer"]
-
+__all__ = ["TokenProjector", "CrossModalAttention", "FusionLayer", "MoEBlock"]

lnp_ml/modeling/layers/moe.py

@@ -0,0 +1,229 @@
+"""
+MoE (Mixture-of-Experts) 模块：sample-level + 跨模态路由。
+
+设计要点（与现有 8-token 架构对齐）：
+    - Router 输入：tab token 池化后的 [B, d]（即配方/实验条件向量）。
+    - Expert 输入：chem token flatten 后的 [B, T_chem * d]（化学侧 token）。
+    - 路由粒度：每个样本只算一次 router → gates [B, K]。
+    - Top-k 稀疏激活 + 可选训练态 jitter 噪声。
+    - 返回 load-balancing aux loss，由 trainer 端按权重加进总 loss。
+
+输出:
+    F_moe: [B, d_model]   与单个 token 同维度，便于追加到 fusion 序列中。
+    extras: dict           诊断与监控信息（aux loss、gates 等）。
+"""
+
+from typing import Dict, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class MoEAttentionPool(nn.Module):
+    """与 FusionLayer 同款的 attention pooling，参数独立。
+
+    将一组 token [B, T, d] 池化成单个查询向量 [B, d]，用作 router 的输入。
+    """
+
+    def __init__(self, d_model: int) -> None:
+        super().__init__()
+        self.d_model = d_model
+        self.query = nn.Parameter(torch.randn(1, 1, d_model))
+        self.proj = nn.Linear(d_model, d_model)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x: [B, T, d_model]
+
+        Returns:
+            [B, d_model]
+        """
+        B = x.size(0)
+        q = self.query.expand(B, -1, -1)                              # [B, 1, d]
+        k = self.proj(x)                                              # [B, T, d]
+        scores = torch.bmm(q, k.transpose(1, 2)) / (self.d_model ** 0.5)
+        weights = F.softmax(scores, dim=-1)                           # [B, 1, T]
+        return torch.bmm(weights, x).squeeze(1)                       # [B, d]
+
+
+class MoERouter(nn.Module):
+    """单层 softmax router + Top-k 稀疏激活。
+
+    Args:
+        d_model: router 输入维度。
+        n_experts: 专家数量 K。
+        top_k: 每个样本激活的专家数。
+        jitter_noise: 训练态加在 logits 上的均匀噪声幅度，用于探索；评估态自动关闭。
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        n_experts: int,
+        top_k: int = 2,
+        jitter_noise: float = 0.0,
+    ) -> None:
+        super().__init__()
+        if not 1 <= top_k <= n_experts:
+            raise ValueError(f"top_k 必须在 [1, n_experts={n_experts}]，收到 {top_k}")
+        self.n_experts = n_experts
+        self.top_k = top_k
+        self.jitter_noise = jitter_noise
+        self.linear = nn.Linear(d_model, n_experts)
+
+    def forward(
+        self, q: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            q: [B, d_model] 路由查询向量。
+
+        Returns:
+            gates:       [B, n_experts] top-k 后再归一化的稀疏概率。
+            probs_full:  [B, n_experts] 未掩码的 softmax 概率（用于 aux loss）。
+            expert_mask: [B, n_experts] 0/1 掩码，标记被激活的专家。
+        """
+        logits = self.linear(q)                                       # [B, K]
+        if self.training and self.jitter_noise > 0.0:
+            noise = (torch.rand_like(logits) - 0.5) * self.jitter_noise
+            logits = logits + noise
+
+        probs_full = F.softmax(logits, dim=-1)                        # [B, K]
+
+        topk_vals, topk_idx = probs_full.topk(self.top_k, dim=-1)     # [B, k]
+        expert_mask = torch.zeros_like(probs_full)
+        expert_mask.scatter_(1, topk_idx, 1.0)
+
+        gates = probs_full * expert_mask
+        gates = gates / gates.sum(dim=-1, keepdim=True).clamp(min=1e-9)
+        return gates, probs_full, expert_mask
+
+
+class MoEExpert(nn.Module):
+    """单个专家 MLP：in_dim → hidden_dim → out_dim。"""
+
+    def __init__(
+        self, in_dim: int, hidden_dim: int, out_dim: int, dropout: float = 0.1,
+    ) -> None:
+        super().__init__()
+        self.net = nn.Sequential(
+            nn.Linear(in_dim, hidden_dim),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(hidden_dim, out_dim),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.net(x)
+
+
+class MoEBlock(nn.Module):
+    """
+    Sample-level 跨模态 MoE。
+
+    流程:
+        1. tab tokens [B, T_tab, d] -> attention pool -> q_tab [B, d]
+        2. q_tab -> Router -> gates [B, K] (+ probs_full / expert_mask)
+        3. chem tokens [B, T_chem, d] -> flatten -> [B, T_chem * d]
+        4. K 个 Expert MLP 并行处理 -> stack [B, K, d]
+        5. 加权求和 -> F_moe [B, d]
+        6. 计算 load-balancing aux loss
+
+    Args:
+        d_model: token 维度。
+        n_chem_tokens: chem 侧 token 数（决定 expert 输入维度）。
+        n_experts: 专家数量 K。
+        top_k: 每个样本激活的专家数。
+        expert_hidden_mult: expert 中间层维度 = expert_hidden_mult * d_model。
+        dropout: expert 内部 dropout。
+        jitter_noise: router 训练态噪声幅度，0.0 表示关闭。
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        n_chem_tokens: int = 4,
+        n_experts: int = 4,
+        top_k: int = 2,
+        expert_hidden_mult: int = 2,
+        dropout: float = 0.1,
+        jitter_noise: float = 0.0,
+    ) -> None:
+        super().__init__()
+        self.d_model = d_model
+        self.n_chem_tokens = n_chem_tokens
+        self.n_experts = n_experts
+        self.top_k = top_k
+
+        self.tab_pool = MoEAttentionPool(d_model)
+        self.router = MoERouter(
+            d_model, n_experts, top_k=top_k, jitter_noise=jitter_noise,
+        )
+
+        expert_in = d_model * n_chem_tokens
+        expert_hidden = d_model * expert_hidden_mult
+        self.experts = nn.ModuleList([
+            MoEExpert(expert_in, expert_hidden, d_model, dropout=dropout)
+            for _ in range(n_experts)
+        ])
+
+    def forward(
+        self,
+        chem: torch.Tensor,
+        tab: torch.Tensor,
+    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
+        """
+        Args:
+            chem: [B, T_chem, d_model] 化学侧 token（router 不看）。
+            tab:  [B, T_tab,  d_model] 配方/实验侧 token（router 看）。
+
+        Returns:
+            F_moe: [B, d_model]
+            extras: {
+                "lb_loss": 标量 load-balancing aux loss（带梯度），
+                "gates":   [B, K] 稀疏归一后的门控（detached），
+                "probs":   [B, K] 原始 softmax 概率（detached），
+            }
+        """
+        if chem.size(1) != self.n_chem_tokens:
+            raise ValueError(
+                f"chem token 数不匹配：期望 {self.n_chem_tokens}, "
+                f"实际 {chem.size(1)}"
+            )
+
+        q_tab = self.tab_pool(tab)                                     # [B, d]
+        gates, probs_full, expert_mask = self.router(q_tab)            # 各 [B, K]
+
+        flat = chem.flatten(start_dim=1)                               # [B, T_chem * d]
+        expert_outs = torch.stack(
+            [expert(flat) for expert in self.experts], dim=1,
+        )                                                              # [B, K, d]
+
+        F_moe = (gates.unsqueeze(-1) * expert_outs).sum(dim=1)         # [B, d]
+
+        lb_loss = self._load_balancing_loss(probs_full, expert_mask)
+
+        return F_moe, {
+            "lb_loss": lb_loss,
+            "gates": gates.detach(),
+            "probs": probs_full.detach(),
+        }
+
+    def _load_balancing_loss(
+        self,
+        probs_full: torch.Tensor,
+        expert_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        """Switch Transformer 风格的 load-balancing loss。
+
+            f_i = 该 batch 中 expert_i 被激活的样本占比（top-k mask 的均值）
+            p_i = 该 batch 中 expert_i 的 softmax 概率均值
+            loss = K * Σ_i f_i * p_i
+
+        理想情况下 f_i 与 p_i 都接近 1/K，loss ≈ 1。
+        """
+        f = expert_mask.mean(dim=0)                                   # [K]
+        p = probs_full.mean(dim=0)                                    # [K]
+        return self.n_experts * (f * p).sum()

lnp_ml/modeling/models.py

@@ -5,7 +5,7 @@ import torch.nn as nn
 from typing import Dict, List, Optional, Literal
 
 from lnp_ml.modeling.encoders import CachedRDKitEncoder, CachedMPNNEncoder
-from lnp_ml.modeling.layers import TokenProjector, CrossModalAttention, FusionLayer
+from lnp_ml.modeling.layers import TokenProjector, CrossModalAttention, FusionLayer, MoEBlock
 from lnp_ml.modeling.heads import MultiTaskHead
 
 
@@ -62,6 +62,12 @@ class LNPModel(nn.Module):
         mpnn_device: str = "cpu",
         # 输入维度配置
         input_dims: Optional[Dict[str, int]] = None,
+        # ============ MoE 相关（新增） ============
+        use_moe: bool = False,
+        moe_n_experts: int = 4,
+        moe_top_k: int = 2,
+        moe_expert_hidden_mult: int = 2,
+        moe_jitter_noise: float = 0.0,
     ) -> None:
         super().__init__()
 
@@ -107,10 +113,28 @@ class LNPModel(nn.Module):
             dropout=dropout,
         )
 
+        # ============ MoE Block (可选) ============
+        self.use_moe = use_moe
+        self._last_moe_extras: Optional[Dict[str, torch.Tensor]] = None
+        if use_moe:
+            self.moe = MoEBlock(
+                d_model=d_model,
+                n_chem_tokens=split_idx,
+                n_experts=moe_n_experts,
+                top_k=moe_top_k,
+                expert_hidden_mult=moe_expert_hidden_mult,
+                dropout=dropout,
+                jitter_noise=moe_jitter_noise,
+            )
+            n_fusion_tokens = n_tokens + 1  # 多一个 F_moe token
+        else:
+            self.moe = None
+            n_fusion_tokens = n_tokens
+
         # ============ Fusion Layer ============
         self.fusion = FusionLayer(
             d_model=d_model,
-            n_tokens=n_tokens,
+            n_tokens=n_fusion_tokens,
             strategy=fusion_strategy,
         )
 
@@ -169,6 +193,29 @@ class LNPModel(nn.Module):
         stacked = torch.stack([projected[k] for k in token_order], dim=1)
         return stacked
 
+    def _attended_with_moe(self, stacked: torch.Tensor) -> torch.Tensor:
+        """
+        Cross-attention + 可选 MoE 旁路 → fusion 输入序列。
+
+        - 不启用 MoE 时：返回 [B, n_tokens, d]，与原行为一致。
+        - 启用 MoE 时：在最后追加一个 F_moe token，返回 [B, n_tokens + 1, d]。
+
+        副作用:
+            把本次 forward 的 MoE 副产物（aux loss / gates / probs）写到
+            self._last_moe_extras，trainer 端通过 get_last_moe_extras() 读取。
+        """
+        attended = self.cross_attention(stacked)
+        if self.moe is not None:
+            split = self.cross_attention.split_idx
+            chem_prime = attended[:, :split, :]
+            tab_prime = attended[:, split:, :]
+            F_moe, extras = self.moe(chem_prime, tab_prime)
+            self._last_moe_extras = extras
+            attended = torch.cat([attended, F_moe.unsqueeze(1)], dim=1)
+        else:
+            self._last_moe_extras = None
+        return attended
+
     def forward_from_projected(
         self,
         stacked: torch.Tensor,
@@ -178,14 +225,14 @@ class LNPModel(nn.Module):
         从已投影的 stacked tokens 开始 forward，用于 Captum 归因。
 
         Args:
-            stacked: [B, n_tokens, d_model] TokenProjector 输出后 stack 的张量
+            stacked: [B, n_tokens, d_model] TokenProjector 输出后 stack 的张量。
             task: 指定单任务名 ("size", "pdi", "ee", "delivery", "biodist", "toxic")。
                   若为 None，返回 delivery head 的标量输出。
 
         Returns:
-            [B, 1] 或 [B, num_classes] 对应任务的预测输出
+            [B, 1] 或 [B, num_classes] 对应任务的预测输出。
         """
-        attended = self.cross_attention(stacked)
+        attended = self._attended_with_moe(stacked)
         fused = self.fusion(attended)
 
         if task is None:
@@ -240,7 +287,7 @@ class LNPModel(nn.Module):
         tabular: Dict[str, torch.Tensor],
     ) -> torch.Tensor:
         """
-        Backbone forward：编码 -> 投影 -> 注意力 -> 融合，不经过任务头。
+        Backbone forward：编码 -> 投影 -> 注意力 -> (可选 MoE) -> 融合，不经过任务头。
 
         用于 pretrain 阶段或需要提取特征的场景。
 
@@ -251,15 +298,9 @@ class LNPModel(nn.Module):
         Returns:
             fused: [B, fusion_dim] 融合后的特征向量
         """
-        # 编码 + 投影 + stack
         stacked = self._encode_and_project(smiles, tabular)
-        
-        # Cross Modal Attention
-        attended = self.cross_attention(stacked)
-        
-        # Fusion
+        attended = self._attended_with_moe(stacked)
         fused = self.fusion(attended)
-        
         return fused
 
     def forward_delivery(
@@ -315,17 +356,24 @@ class LNPModel(nn.Module):
         if self.mpnn_encoder is not None:
             self.mpnn_encoder.clear_cache()
 
+    def get_last_moe_extras(self) -> Optional[Dict[str, torch.Tensor]]:
+        """返回最近一次 forward 中 MoE 模块的副产物（aux loss、gates、probs）。
+
+        若未启用 MoE 或还未调用过 forward，返回 None。
+        """
+        return self._last_moe_extras
+
     def get_backbone_state_dict(self) -> Dict[str, torch.Tensor]:
         """
         获取 backbone 部分的 state_dict（不含任务头）。
 
-        包含: token_projector, cross_attention, fusion
+        包含: token_projector, cross_attention, fusion，以及（启用时）moe。
         """
-        backbone_keys = []
-        for name in self.state_dict().keys():
-            if name.startswith(("token_projector.", "cross_attention.", "fusion.")):
-                backbone_keys.append(name)
-        
+        backbone_prefixes = ("token_projector.", "cross_attention.", "fusion.", "moe.")
+        backbone_keys = [
+            name for name in self.state_dict().keys()
+            if name.startswith(backbone_prefixes)
+        ]
         return {k: v for k, v in self.state_dict().items() if k in backbone_keys}
 
     def get_delivery_head_state_dict(self) -> Dict[str, torch.Tensor]:
@@ -349,19 +397,19 @@ class LNPModel(nn.Module):
             load_delivery_head: 是否加载 delivery head 权重
             strict: 是否严格匹配（默认 False，允许缺失/多余的键）
         """
-        # 筛选要加载的参数
+        backbone_prefixes = ("token_projector.", "cross_attention.", "fusion.", "moe.")
+
         keys_to_load = []
         for name in pretrain_state_dict.keys():
-            # Backbone 部分
-            if name.startswith(("token_projector.", "cross_attention.", "fusion.")):
+            if name.startswith(backbone_prefixes):
                 keys_to_load.append(name)
-            # Delivery head（可选）
             elif load_delivery_head and name.startswith("head.delivery_head."):
                 keys_to_load.append(name)
 
-        filtered_state_dict = {k: v for k, v in pretrain_state_dict.items() if k in keys_to_load}
+        filtered_state_dict = {
+            k: v for k, v in pretrain_state_dict.items() if k in keys_to_load
+        }
 
-        # 加载权重
         missing, unexpected = [], []
         model_state = self.state_dict()
         for k, v in filtered_state_dict.items():
@@ -369,7 +417,9 @@ class LNPModel(nn.Module):
                 if model_state[k].shape == v.shape:
                     model_state[k] = v
                 else:
-                    unexpected.append(f"{k} (shape mismatch: {model_state[k].shape} vs {v.shape})")
+                    unexpected.append(
+                        f"{k} (shape mismatch: {model_state[k].shape} vs {v.shape})"
+                    )
             else:
                 unexpected.append(k)
 
@@ -391,6 +441,12 @@ class LNPModelWithoutMPNN(LNPModel):
         head_hidden_dim: int = 128,
         dropout: float = 0.1,
         input_dims: Optional[Dict[str, int]] = None,
+        # ============ MoE 相关（新增） ============
+        use_moe: bool = False,
+        moe_n_experts: int = 4,
+        moe_top_k: int = 2,
+        moe_expert_hidden_mult: int = 2,
+        moe_jitter_noise: float = 0.0,
     ) -> None:
         # 移除 mpnn 维度
         dims = input_dims or DEFAULT_INPUT_DIMS.copy()
@@ -406,5 +462,10 @@ class LNPModelWithoutMPNN(LNPModel):
             mpnn_checkpoint=None,
             mpnn_ensemble_paths=None,
             input_dims=dims,
+            use_moe=use_moe,
+            moe_n_experts=moe_n_experts,
+            moe_top_k=moe_top_k,
+            moe_expert_hidden_mult=moe_expert_hidden_mult,
+            moe_jitter_noise=moe_jitter_noise,
         )
 

lnp_ml/modeling/nested_cv_optuna.py

@@ -178,8 +178,21 @@ def create_model(
     dropout: float = 0.1,
     use_mpnn: bool = False,
     mpnn_device: str = "cpu",
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
 ) -> Union[LNPModel, LNPModelWithoutMPNN]:
     """创建模型"""
+    moe_kwargs = dict(
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
+    )
+
     if use_mpnn:
         ensemble_paths = find_mpnn_ensemble_paths()
         return LNPModel(
@@ -191,6 +204,7 @@ def create_model(
             dropout=dropout,
             mpnn_ensemble_paths=ensemble_paths,
             mpnn_device=mpnn_device,
+            **moe_kwargs,
         )
     else:
         return LNPModelWithoutMPNN(
@@ -200,6 +214,7 @@ def create_model(
             fusion_strategy=fusion_strategy,
             head_hidden_dim=head_hidden_dim,
             dropout=dropout,
+            **moe_kwargs,
         )
 
 
@@ -344,7 +359,6 @@ def load_pretrain_weights_to_model(
 
 
 # ============ 内层 Optuna 调参 ============
-
 def run_inner_optuna(
     full_dataset: LNPDataset,
     inner_train_indices: np.ndarray,
@@ -362,6 +376,11 @@ def run_inner_optuna(
     pretrain_config: Optional[Dict] = None,
     load_delivery_head: bool = True,
     rdkit_cache: Optional[Dict] = None,
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
 ) -> Tuple[Dict, int, optuna.Study]:
     """
     在内层数据上运行 Optuna 超参搜索。
@@ -456,6 +475,11 @@ def run_inner_optuna(
                 dropout=dropout,
                 use_mpnn=use_mpnn,
                 mpnn_device=device.type,
+                use_moe=use_moe,
+                moe_n_experts=moe_n_experts,
+                moe_top_k=moe_top_k,
+                moe_expert_hidden_mult=moe_expert_hidden_mult,
+                moe_jitter_noise=moe_jitter_noise,
             )
             if rdkit_cache is not None:
                 model.rdkit_encoder._cache = rdkit_cache
@@ -524,7 +548,6 @@ def run_inner_optuna(
 
 
 # ============ 单 fold 执行（可跨进程调用） ============
-
 def _run_single_outer_fold(
     outer_fold: int,
     outer_train_idx: np.ndarray,
@@ -543,6 +566,11 @@ def _run_single_outer_fold(
     pretrain_config: Optional[Dict],
     load_delivery_head: bool,
     device_str: str,
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
 ) -> Dict:
     """
     执行单个外层 fold 的完整流程（内层调参 + 外层训练 + 评估）。
@@ -591,6 +619,11 @@ def _run_single_outer_fold(
         pretrain_config=pretrain_config,
         load_delivery_head=load_delivery_head,
         rdkit_cache=rdkit_cache,
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
     )
 
     # 保存最佳参数
@@ -624,6 +657,11 @@ def _run_single_outer_fold(
         dropout=best_params["dropout"],
         use_mpnn=use_mpnn,
         mpnn_device=device.type,
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
     )
     model.rdkit_encoder._cache = rdkit_cache
 
@@ -659,6 +697,11 @@ def _run_single_outer_fold(
         "head_hidden_dim": best_params["head_hidden_dim"],
         "dropout": best_params["dropout"],
         "use_mpnn": use_mpnn,
+        "use_moe": use_moe,
+        "moe_n_experts": moe_n_experts,
+        "moe_top_k": moe_top_k,
+        "moe_expert_hidden_mult": moe_expert_hidden_mult,
+        "moe_jitter_noise": moe_jitter_noise,
     }
 
     torch.save({
@@ -696,7 +739,6 @@ def _run_single_outer_fold(
 
 
 # ============ 主流程 ============
-
 @app.command()
 def main(
     input_path: Path = INTERIM_DATA_DIR / "internal.csv",
@@ -717,6 +759,12 @@ def main(
     load_delivery_head: bool = False,
     # MPNN
     use_mpnn: bool = False,
+    # MoE（新增）
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
     # 并行
     parallel: bool = False,
     # 设备
@@ -805,6 +853,11 @@ def main(
             pretrain_config=pretrain_config,
             load_delivery_head=load_delivery_head,
             device_str=device_str,
+            use_moe=use_moe,
+            moe_n_experts=moe_n_experts,
+            moe_top_k=moe_top_k,
+            moe_expert_hidden_mult=moe_expert_hidden_mult,
+            moe_jitter_noise=moe_jitter_noise,
         ))
     
     if parallel:

lnp_ml/modeling/trainer.py

@@ -19,6 +19,7 @@ class LossWeights:
     delivery: float = 1.0
     biodist: float = 1.0
     toxic: float = 1.0
+    moe_lb: float = 0.01  # MoE load-balancing 系数（仅在 use_moe=True 时生效）
     # size: float = 0.1
     # pdi: float = 0.3
     # ee: float = 0.3
@@ -32,6 +33,7 @@ def compute_multitask_loss(
     targets: Dict[str, torch.Tensor],
     mask: Dict[str, torch.Tensor],
     weights: Optional[LossWeights] = None,
+    model: Optional[nn.Module] = None,
 ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
     """
     计算多任务损失。
@@ -41,6 +43,8 @@ def compute_multitask_loss(
         targets: 真实标签
         mask: 有效样本掩码
         weights: 各任务权重
+        model:  可选，若提供且模型暴露 get_last_moe_extras()，会把 MoE 的
+                load-balancing aux loss 按 weights.moe_lb 加进总 loss。
 
     Returns:
         (total_loss, loss_dict) 总损失和各任务损失
@@ -94,7 +98,6 @@ def compute_multitask_loss(
         m = mask["biodist"]
         pred = outputs["biodist"][m]
         tgt = targets["biodist"][m]
-        # KL divergence: KL(target || pred)
         losses["biodist"] = F.kl_div(
             pred.log().clamp(min=-100),
             tgt,
@@ -102,6 +105,13 @@ def compute_multitask_loss(
         )
         total_loss = total_loss + weights.biodist * losses["biodist"]
 
+    # MoE load-balancing aux loss（可选）
+    if model is not None and hasattr(model, "get_last_moe_extras"):
+        extras = model.get_last_moe_extras()
+        if extras is not None and "lb_loss" in extras:
+            losses["moe_lb"] = extras["lb_loss"]
+            total_loss = total_loss + weights.moe_lb * losses["moe_lb"]
+
     return total_loss, losses
 
 
@@ -115,7 +125,7 @@ def train_epoch(
     """训练一个 epoch"""
     model.train()
     total_loss = 0.0
-    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic"]}
+    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
     n_batches = 0
     
     for batch in tqdm(loader, desc="Training", leave=False):
@@ -126,7 +136,7 @@ def train_epoch(
         
         optimizer.zero_grad()
         outputs = model(smiles, tabular)
-        loss, losses = compute_multitask_loss(outputs, targets, mask, weights)
+        loss, losses = compute_multitask_loss(outputs, targets, mask, weights, model=model)
         
         loss.backward()
         torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
@@ -153,7 +163,7 @@ def validate(
     """验证"""
     model.eval()
     total_loss = 0.0
-    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic"]}
+    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
     n_batches = 0
     
     # 用于计算准确率
@@ -167,7 +177,7 @@ def validate(
         mask = {k: v.to(device) for k, v in batch["mask"].items()}
         
         outputs = model(smiles, tabular)
-        loss, losses = compute_multitask_loss(outputs, targets, mask, weights)
+        loss, losses = compute_multitask_loss(outputs, targets, mask, weights, model=model)
         
         total_loss += loss.item()
         for k, v in losses.items():

lnp_ml/modeling/trainer_balanced.py

@@ -28,6 +28,7 @@ class LossWeightsBalanced:
     delivery: float = 1.0
     biodist: float = 1.0
     toxic: float = 1.0
+    moe_lb: float = 0.01  # MoE load-balancing 系数（仅在 use_moe=True 时生效）
 
 
 def compute_class_weights_from_loader(
@@ -110,6 +111,7 @@ def compute_multitask_loss_balanced(
     mask: Dict[str, torch.Tensor],
     task_weights: Optional[LossWeightsBalanced] = None,
     class_weights: Optional[ClassWeights] = None,
+    model: Optional[nn.Module] = None,
 ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
     """
     计算带类权重的多任务损失。
@@ -186,6 +188,13 @@ def compute_multitask_loss_balanced(
         )
         total_loss = total_loss + task_weights.biodist * losses["biodist"]
 
+    # MoE load-balancing aux loss（可选）
+    if model is not None and hasattr(model, "get_last_moe_extras"):
+        extras = model.get_last_moe_extras()
+        if extras is not None and "lb_loss" in extras:
+            losses["moe_lb"] = extras["lb_loss"]
+            total_loss = total_loss + task_weights.moe_lb * losses["moe_lb"]
+    
     return total_loss, losses
 
 
@@ -200,7 +209,7 @@ def train_epoch_balanced(
     """带类权重的训练一个 epoch"""
     model.train()
     total_loss = 0.0
-    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic"]}
+    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
     n_batches = 0
     
     for batch in tqdm(loader, desc="Training", leave=False):
@@ -212,7 +221,7 @@ def train_epoch_balanced(
         optimizer.zero_grad()
         outputs = model(smiles, tabular)
         loss, losses = compute_multitask_loss_balanced(
-            outputs, targets, mask, task_weights, class_weights
+            outputs, targets, mask, task_weights, class_weights, model=model,
         )
         
         loss.backward()
@@ -241,7 +250,7 @@ def validate_balanced(
     """带类权重的验证"""
     model.eval()
     total_loss = 0.0
-    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic"]}
+    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
     n_batches = 0
     
     # 用于计算准确率
@@ -256,7 +265,7 @@ def validate_balanced(
         
         outputs = model(smiles, tabular)
         loss, losses = compute_multitask_loss_balanced(
-            outputs, targets, mask, task_weights, class_weights
+            outputs, targets, mask, task_weights, class_weights, model=model,
         )
         
         total_loss += loss.item()
@@ -286,7 +295,7 @@ def validate_balanced(
     return metrics
 
 
-BACKBONE_PREFIXES = ("token_projector.", "cross_attention.", "fusion.")
+BACKBONE_PREFIXES = ("token_projector.", "cross_attention.", "fusion.", "moe.")
 
 
 def build_optimizer(