feat: bench.py 加命令行参数，支持子进程方式跑（绕开内核torch限制）

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

代码/code/bench.py

@@ -115,5 +115,38 @@ def run_once(config_override=None, batch_size=50, max_batches=None, max_feasign_
     return res
 
 
+def _parse_args():
+    import argparse
+    ap = argparse.ArgumentParser(description="CTI 推理测量闭环（以子进程方式跑：!python bench.py ...）")
+    ap.add_argument("--smoke", type=int, default=None, help="只跑前 N 个 batch（冒烟）")
+    ap.add_argument("--bs", type=int, default=50, help="batch_size（本地参考）")
+    ap.add_argument("--fp32", action="store_true", help="FP32 天花板 = 关 fp16 + 关 expert 合并")
+    ap.add_argument("--no-fp16", action="store_true", help="关闭半精度")
+    ap.add_argument("--no-merge", action="store_true", help="关闭 expert 合并")
+    ap.add_argument("--signid", choices=["clamp", "modulo"], default=None, help="sign-id 处理方式")
+    ap.add_argument("--merge-th", type=float, default=None, help="expert 合并余弦阈值")
+    ap.add_argument("--keep", type=str, default=None,
+                    help="逗号分隔的 keep_fp32_modules，如 linear,rep_encoder.input_norm")
+    ap.add_argument("--feasign-none", action="store_true",
+                    help="不截断特征（max_feasign_per_slot=None）")
+    return ap.parse_args()
+
+
 if __name__ == "__main__":
-    run_once({})  # 默认配置基准
+    a = _parse_args()
+    cfg = {}
+    if a.fp32:
+        cfg["fp16"] = False
+        cfg["expert_merge"] = False
+    if a.no_fp16:
+        cfg["fp16"] = False
+    if a.no_merge:
+        cfg["expert_merge"] = False
+    if a.signid:
+        cfg["signid_mode"] = a.signid
+    if a.merge_th is not None:
+        cfg["merge_threshold"] = a.merge_th
+    if a.keep is not None:
+        cfg["keep_fp32_modules"] = tuple(x for x in a.keep.split(",") if x)
+    mf = None if a.feasign_none else {1: 2}
+    run_once(cfg, batch_size=a.bs, max_batches=a.smoke, max_feasign_per_slot=mf)