feat: PCOC校准(logit_bias单调偏移,AUC不变,免费+0.34) + bench自动拟合建议bias

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

代码/code/infer.py

@@ -145,6 +145,7 @@ CONFIG = {
     "vectorize_moe": True,    # True=稠密向量化MoE(无同步点)；False=原逐expert循环(.nonzero同步)
     "moe_baddbmm": True,      # MoE FFN 用 baddbmm(cutlass GEMM+bias epilogue融合),省 bias add kernel
     "skip_moe_loss": True,    # 推理跳过 moe_loss(load-balance,推理无用),省 importance/std/mean kernel
+    "logit_bias": 0.0,        # PCOC 校准:输出 logit 加常数偏移使 PCOC→1.0(单调变换,AUC不变,免费+~0.34分)
     "fuse_embedding": True,   # True=28个slot的查表+池化融合为1次(减per-batch kernel启动)
     "syncfree_mask": True,    # True=用searchsorted构造因果mask(无同步)；False=repeat_interleave(同步)
     "emb_fp16": True,         # True=Embedding表转FP16(查表带宽减半,实测AUC 0.75932≈无损)
@@ -877,6 +878,9 @@ class CTRModel(nn.Module):
         encoder_output, moe_loss = self.seq_encoder(x=seq_input, extension=extension)
         encoder_output = encoder_output.squeeze(0)
         pred = self.linear(encoder_output)
+        bias = CONFIG.get("logit_bias", 0.0)
+        if bias != 0.0:
+            pred = pred + bias   # PCOC 校准（单调，不改 AUC）
         pred_logits = torch.clamp(pred, min=-15.0, max=15.0)
         return pred_logits, moe_loss