perf: RepEncoder 融合 28-slot 查表+池化为单次(减per-batch kernel启动,无新增同步)

延续 dense MoE 的胜因(消 per-batch 开销在评测端被放大见效)。28次embedding
+28次segment_reduce 融合为1次；用 numel 读shape避免同步；base累加无同步。
保留 _rep_forward_perslot 作等价对照。CONFIG.fuse_embedding 默认 True。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

代码/code/infer.py

@@ -48,6 +48,7 @@ CONFIG = {
     # synchronize，去掉同步点可能让被计时的 model(batch) 大幅缩短。本地force-sync看不出，
     # 须靠提交验证。AUC中性、MoE仅占2%算力故风险极低。
     "vectorize_moe": True,    # True=稠密向量化MoE(无同步点)；False=原逐expert循环(.nonzero同步)
+    "fuse_embedding": True,   # True=28个slot的查表+池化融合为1次(减per-batch kernel启动)
     "compile": False,         # 是否 torch.compile（实测慢5×，勿开）
 }
 
@@ -327,6 +328,22 @@ def move_batch_to_device(batch, device):
         return batch
 
 
+def _rep_forward_perslot(enc, batch):
+    """原始逐 slot 实现（保留作数值等价对照/回退）。"""
+    pooled_embs = []
+    max_idx = enc.emb.num_embeddings - 1
+    target_dtype = enc.input_norm.weight.dtype
+    for i in range(enc.slot_num):
+        values, offsets = batch[i + 1]
+        offsets = offsets.to(values.device)
+        values = enc._signid(values, max_idx)
+        sign_emb = enc.emb(values).to(target_dtype)
+        res = torch.segment_reduce(sign_emb, reduce='sum', offsets=offsets, initial=0)
+        pooled_embs.append(res)
+    fused_embs = torch.cat(pooled_embs, dim=1)
+    return enc.linear(enc.input_norm(fused_embs))
+
+
 class RepEncoder(nn.Module):
     def __init__(self, vocab_size, emb_dim, padding_idx=0, slot_num=0, d_model=0):
         super().__init__()
@@ -336,24 +353,35 @@ class RepEncoder(nn.Module):
         self.input_norm = nn.LayerNorm(slot_num * emb_dim)
         self.linear = nn.Linear(in_features=slot_num * emb_dim, out_features=d_model)
 
+    def _signid(self, values, max_idx):
+        if CONFIG["signid_mode"] == "modulo":
+            return values % self.emb.num_embeddings  # 取模哈希（与训练一致时用）
+        return values.clamp(0, max_idx)               # 超界 sign id 截断
+
     def forward(self, batch):
-        pooled_embs = []
+        if not CONFIG.get("fuse_embedding", True):
+            return _rep_forward_perslot(self, batch)
+
         max_idx = self.emb.num_embeddings - 1
-        target_dtype = self.input_norm.weight.dtype  # 后续层 dtype（FP16 时为 torch.float16）
+        target_dtype = self.input_norm.weight.dtype
+        N = batch[1][1].numel() - 1  # 样本数（slot1 的 offsets 段数）
+
+        # 把 28 个 slot 的 values 拼成一条，offsets 平移拼成覆盖 28*N 段的单一 offsets
+        parts, ends, base = [], [], 0
         for i in range(self.slot_num):
             values, offsets = batch[i + 1]
             offsets = offsets.to(values.device)
-            if CONFIG["signid_mode"] == "modulo":
-                values = values % self.emb.num_embeddings  # 取模哈希（与训练一致时用）
-            else:
-                values = values.clamp(0, max_idx)  # 超出 vocab_size 的 sign id 截断，避免越界
-            sign_emb = self.emb(values).to(target_dtype)
-            res = torch.segment_reduce(sign_emb, reduce='sum', offsets=offsets, initial=0)
-            pooled_embs.append(res)
-        fused_embs = torch.cat(pooled_embs, dim=1)
-        norm_emb = self.input_norm(fused_embs)
-        rep_emb = self.linear(norm_emb)
-        return rep_emb
+            parts.append(values)
+            ends.append(offsets[1:] + base)  # 该 slot 各样本的段尾（平移 base）
+            base += values.numel()           # numel 读 shape，不触发同步
+        cat_values = self._signid(torch.cat(parts), max_idx)
+        seg = torch.cat([torch.zeros(1, dtype=torch.long, device=cat_values.device),
+                         torch.cat(ends)])   # [28*N + 1]
+        emb = self.emb(cat_values).to(target_dtype)
+        pooled = torch.segment_reduce(emb, reduce='sum', offsets=seg, initial=0)  # [28*N, emb]
+        pooled = pooled.view(self.slot_num, N, self.emb_dim).permute(1, 0, 2).reshape(
+            N, self.slot_num * self.emb_dim)
+        return self.linear(self.input_norm(pooled))
 
 
 def _varlen_attention(q, k, v, user_offsets):