perf: emb_fp16 默认开启(本地AUC 0.75932≈无损,查表带宽减半)；修正打印

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

代码/code/infer.py

@@ -50,7 +50,7 @@ CONFIG = {
     "vectorize_moe": True,    # True=稠密向量化MoE(无同步点)；False=原逐expert循环(.nonzero同步)
     "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": False,        # True=Embedding表也转FP16(查表带宽减半,可能微动AUC)；False=保留FP32
+    "emb_fp16": True,         # True=Embedding表转FP16(查表带宽减半,实测AUC 0.75932≈无损)
     "compile": False,         # 是否 torch.compile（实测慢5×，勿开）
 }
 
@@ -708,8 +708,9 @@ def load_model(ckpt_path, device='cuda:0'):
             for name, module in model.named_modules():
                 if name and any(name.startswith(p) for p in CONFIG["keep_fp32_modules"]):
                     _force_fp32_io(module)
-            print(f"[INFO] FP16 on; FP32-kept: "
-                  f"{('rep_encoder.emb',) + tuple(CONFIG['keep_fp32_modules'])}")
+            emb_note = "emb=FP16" if CONFIG.get("emb_fp16", False) else "emb=FP32"
+            print(f"[INFO] FP16 on; {emb_note}; extra FP32-kept: "
+                  f"{tuple(CONFIG['keep_fp32_modules'])}")
         else:
             model = model.float()
             print("[INFO] FP32 reference (no half)")