From 987ecf6386e505ee747642ce972f2c01fa25b4dc Mon Sep 17 00:00:00 2001
From: DicongLi <2024379585@qq.com>
Date: Sun, 5 Jul 2026 01:07:26 +0800
Subject: [PATCH] =?UTF-8?q?feat:=20=E5=90=8C=E6=AD=A5=20RAG=20=E8=9E=8D?=
 =?UTF-8?q?=E5=90=88=E4=B8=8E=E8=AE=AD=E7=BB=83=E6=A0=B8=E5=BF=83=E4=BB=A3?=
 =?UTF-8?q?=E7=A0=81(fusion=20=E6=94=AF=E6=8C=81=20f=5Fretr=E3=80=81models?=
 =?UTF-8?q?/trainer/moe/dataset)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 lnp_ml/modeling/layers/fusion.py | 306 +++++++++++----------
 lnp_ml/modeling/layers/moe.py    | 456 +++++++++++++++----------------
 2 files changed, 383 insertions(+), 379 deletions(-)

diff --git a/lnp_ml/modeling/layers/fusion.py b/lnp_ml/modeling/layers/fusion.py
index 70ef066..3824eb8 100644
--- a/lnp_ml/modeling/layers/fusion.py
+++ b/lnp_ml/modeling/layers/fusion.py
@@ -1,152 +1,156 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from typing import Dict, List, Literal, Optional, Tuple, Union
-
-
-PoolingStrategy = Literal["concat", "avg", "max", "attention"]
-
-
-class FusionLayer(nn.Module):
-    """
-    将多个 token 融合成单个向量。
-
-    输入: Dict[str, Tensor] 或 [B, n_tokens, d_model]
-    输出: [B, fusion_dim]
-
-    策略:
-        - concat: [B, n_tokens, d_model] -> [B, n_tokens * d_model]
-        - avg: [B, n_tokens, d_model] -> [B, d_model]
-        - max: [B, n_tokens, d_model] -> [B, d_model]
-        - attention: [B, n_tokens, d_model] -> [B, d_model] (learnable attention pooling)
-    """
-
-    def __init__(
-        self,
-        d_model: int,
-        n_tokens: int,
-        strategy: PoolingStrategy = "attention",
-    ) -> None:
-        """
-        Args:
-            d_model: 每个 token 的维度
-            n_tokens: token 数量（如 8）
-            strategy: 融合策略
-        """
-        super().__init__()
-        self.d_model = d_model
-        self.n_tokens = n_tokens
-        self.strategy = strategy
-
-        if strategy == "concat":
-            self.fusion_dim = n_tokens * d_model
-        else:
-            self.fusion_dim = d_model
-
-        # Attention pooling: learnable query
-        if strategy == "attention":
-            self.attn_query = nn.Parameter(torch.randn(1, 1, d_model))
-            self.attn_proj = nn.Linear(d_model, d_model)
-
-    def forward(
-        self,
-        x: Union[Dict[str, torch.Tensor], torch.Tensor],
-        return_attn_weights: bool = False,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
-        """
-        Args:
-            x: Dict[str, Tensor] 每个 [B, d_model]，或已 stack 的 [B, n_tokens, d_model]
-            return_attn_weights: 若为 True 且策略为 attention，额外返回 attn_weights [B, n_tokens]
-
-        Returns:
-            return_attn_weights=False: [B, fusion_dim]
-            return_attn_weights=True:  ([B, fusion_dim], [B, n_tokens])
-        """
-        if isinstance(x, dict):
-            x = torch.stack(list(x.values()), dim=1)
-
-        if self.strategy == "concat":
-            out = x.flatten(start_dim=1)
-            return (out, None) if return_attn_weights else out
-
-        elif self.strategy == "avg":
-            out = x.mean(dim=1)
-            return (out, None) if return_attn_weights else out
-
-        elif self.strategy == "max":
-            out = x.max(dim=1).values
-            return (out, None) if return_attn_weights else out
-
-        elif self.strategy == "attention":
-            return self._attention_pooling(x, return_attn_weights)
-
-        else:
-            raise ValueError(f"Unknown strategy: {self.strategy}")
-
-    def _attention_pooling(
-        self, x: torch.Tensor, return_attn_weights: bool = False,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
-        """
-        Attention pooling: 用可学习 query 对 tokens 做加权求和
-
-        Args:
-            x: [B, n_tokens, d_model]
-            return_attn_weights: 是否返回权重
-
-        Returns:
-            return_attn_weights=False: [B, d_model]
-            return_attn_weights=True:  ([B, d_model], [B, n_tokens])
-        """
-        B = x.size(0)
-        query = self.attn_query.expand(B, -1, -1)
-
-        keys = self.attn_proj(x)
-        scores = torch.bmm(query, keys.transpose(1, 2)) / (self.d_model ** 0.5)
-        attn_weights = F.softmax(scores, dim=-1)  # [B, 1, n_tokens]
-
-        out = torch.bmm(attn_weights, x).squeeze(1)  # [B, d_model]
-
-        if return_attn_weights:
-            return out, attn_weights.squeeze(1)  # [B, n_tokens]
-        return out
-
-
-class ResidualConcatFusion(nn.Module):
-    """对真实 token 做 attention pooling，再用零初始化门把 MoE/LLM 旁路以残差方式加入。
-
-    g_moe / g_llm 初始为 0 → +moe/+llm 起点严格等于 baseline；
-    旁路只有确实有用时才会被训练打开，从机制上保证“加了不会更差”。
-    """
-
-    def __init__(self, d_model: int, strategy: PoolingStrategy = "attention") -> None:
-        super().__init__()
-        if strategy == "concat":
-            raise ValueError("ResidualConcatFusion 不支持 concat（token 数随开关变化）")
-        self.d_model = d_model
-        self.pool = FusionLayer(d_model=d_model, n_tokens=1, strategy=strategy)
-        self.fusion_dim = self.pool.fusion_dim
-        # 零初始化门控（可学习标量），旁路初始不参与
-        self.g_moe = nn.Parameter(torch.zeros(()))
-        self.g_llm = nn.Parameter(torch.zeros(()))
-
-    def forward(
-        self,
-        chem: torch.Tensor,
-        tab: torch.Tensor,
-        f_moe: Optional[torch.Tensor] = None,
-        f_llm: Optional[torch.Tensor] = None,
-        return_attn_weights: bool = False,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
-        # 只对真实 token（chem + tab）做注意力池化，旁路不参与 softmax 竞争
-        seq = torch.cat([chem, tab], dim=1)  # [B, n_chem + n_cond, d_model]
-        pooled = self.pool(seq, return_attn_weights=return_attn_weights)
-        if return_attn_weights:
-            pooled, attn = pooled
-
-        out = pooled
-        if f_moe is not None:
-            out = out + self.g_moe * f_moe  # 残差 + 零初始化门
-        if f_llm is not None:
-            out = out + self.g_llm * f_llm
-
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Dict, List, Literal, Optional, Tuple, Union
+
+
+PoolingStrategy = Literal["concat", "avg", "max", "attention"]
+
+
+class FusionLayer(nn.Module):
+    """
+    将多个 token 融合成单个向量。
+
+    输入: Dict[str, Tensor] 或 [B, n_tokens, d_model]
+    输出: [B, fusion_dim]
+
+    策略:
+        - concat: [B, n_tokens, d_model] -> [B, n_tokens * d_model]
+        - avg: [B, n_tokens, d_model] -> [B, d_model]
+        - max: [B, n_tokens, d_model] -> [B, d_model]
+        - attention: [B, n_tokens, d_model] -> [B, d_model] (learnable attention pooling)
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        n_tokens: int,
+        strategy: PoolingStrategy = "attention",
+    ) -> None:
+        """
+        Args:
+            d_model: 每个 token 的维度
+            n_tokens: token 数量（如 8）
+            strategy: 融合策略
+        """
+        super().__init__()
+        self.d_model = d_model
+        self.n_tokens = n_tokens
+        self.strategy = strategy
+
+        if strategy == "concat":
+            self.fusion_dim = n_tokens * d_model
+        else:
+            self.fusion_dim = d_model
+
+        # Attention pooling: learnable query
+        if strategy == "attention":
+            self.attn_query = nn.Parameter(torch.randn(1, 1, d_model))
+            self.attn_proj = nn.Linear(d_model, d_model)
+
+    def forward(
+        self,
+        x: Union[Dict[str, torch.Tensor], torch.Tensor],
+        return_attn_weights: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Args:
+            x: Dict[str, Tensor] 每个 [B, d_model]，或已 stack 的 [B, n_tokens, d_model]
+            return_attn_weights: 若为 True 且策略为 attention，额外返回 attn_weights [B, n_tokens]
+
+        Returns:
+            return_attn_weights=False: [B, fusion_dim]
+            return_attn_weights=True:  ([B, fusion_dim], [B, n_tokens])
+        """
+        if isinstance(x, dict):
+            x = torch.stack(list(x.values()), dim=1)
+
+        if self.strategy == "concat":
+            out = x.flatten(start_dim=1)
+            return (out, None) if return_attn_weights else out
+
+        elif self.strategy == "avg":
+            out = x.mean(dim=1)
+            return (out, None) if return_attn_weights else out
+
+        elif self.strategy == "max":
+            out = x.max(dim=1).values
+            return (out, None) if return_attn_weights else out
+
+        elif self.strategy == "attention":
+            return self._attention_pooling(x, return_attn_weights)
+
+        else:
+            raise ValueError(f"Unknown strategy: {self.strategy}")
+
+    def _attention_pooling(
+        self, x: torch.Tensor, return_attn_weights: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Attention pooling: 用可学习 query 对 tokens 做加权求和
+
+        Args:
+            x: [B, n_tokens, d_model]
+            return_attn_weights: 是否返回权重
+
+        Returns:
+            return_attn_weights=False: [B, d_model]
+            return_attn_weights=True:  ([B, d_model], [B, n_tokens])
+        """
+        B = x.size(0)
+        query = self.attn_query.expand(B, -1, -1)
+
+        keys = self.attn_proj(x)
+        scores = torch.bmm(query, keys.transpose(1, 2)) / (self.d_model ** 0.5)
+        attn_weights = F.softmax(scores, dim=-1)  # [B, 1, n_tokens]
+
+        out = torch.bmm(attn_weights, x).squeeze(1)  # [B, d_model]
+
+        if return_attn_weights:
+            return out, attn_weights.squeeze(1)  # [B, n_tokens]
+        return out
+
+
+class ResidualConcatFusion(nn.Module):
+    """对真实 token 做 attention pooling，再用零初始化门把 MoE/LLM 旁路以残差方式加入。
+
+    g_moe / g_llm 初始为 0 → +moe/+llm 起点严格等于 baseline；
+    旁路只有确实有用时才会被训练打开，从机制上保证“加了不会更差”。
+    """
+
+    def __init__(self, d_model: int, strategy: PoolingStrategy = "attention") -> None:
+        super().__init__()
+        if strategy == "concat":
+            raise ValueError("ResidualConcatFusion 不支持 concat（token 数随开关变化）")
+        self.d_model = d_model
+        self.pool = FusionLayer(d_model=d_model, n_tokens=1, strategy=strategy)
+        self.fusion_dim = self.pool.fusion_dim
+        # 零初始化门控（可学习标量），旁路初始不参与
+        self.g_moe = nn.Parameter(torch.zeros(()))
+        self.g_llm = nn.Parameter(torch.zeros(()))
+        self.g_retr = nn.Parameter(torch.zeros(()))  # 检索旁路零初始化门控
+
+    def forward(
+        self,
+        chem: torch.Tensor,
+        tab: torch.Tensor,
+        f_moe: Optional[torch.Tensor] = None,
+        f_llm: Optional[torch.Tensor] = None,
+        f_retr: Optional[torch.Tensor] = None,
+        return_attn_weights: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        # 只对真实 token（chem + tab）做注意力池化，旁路不参与 softmax 竞争
+        seq = torch.cat([chem, tab], dim=1)  # [B, n_chem + n_cond, d_model]
+        pooled = self.pool(seq, return_attn_weights=return_attn_weights)
+        if return_attn_weights:
+            pooled, attn = pooled
+
+        out = pooled
+        if f_moe is not None:
+            out = out + self.g_moe * f_moe  # 残差 + 零初始化门
+        if f_llm is not None:
+            out = out + self.g_llm * f_llm
+        if f_retr is not None:
+            out = out + self.g_retr * f_retr  # 检索旁路，零初始化门保证起点=不开
+
         return (out, attn) if return_attn_weights else out
\ No newline at end of file
diff --git a/lnp_ml/modeling/layers/moe.py b/lnp_ml/modeling/layers/moe.py
index 8a07c05..aab389d 100644
--- a/lnp_ml/modeling/layers/moe.py
+++ b/lnp_ml/modeling/layers/moe.py
@@ -1,229 +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]
+"""
+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()
\ No newline at end of file