final: precompute_rep 默认关(评测端三连回退,无日志难诊断) — 锁定干净 ~68

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
OwnerSunshine530
2026-06-16 18:35:33 +08:00
parent 981b3aee11
commit ae7fce7d10
+2 -2
View File
@@ -55,8 +55,8 @@ CONFIG = {
"dedup_embedding": True, # True=查表前对sign去重(只查唯一值再展开),本地7.80->6.49s,AUC逐位等价
"sparse_pool": False, # True=用(段×唯一)稀疏矩阵乘做池化,避免materialize整个[M,512](段内高重复时省)
"compile": False, # 是否 torch.compile(实测慢5×,勿开)
# 预计算改为"捕获评测端 item_dict"(不猜路径/不重载/max_feasign必一致/gather必命中),根治回退
"precompute_rep": True, # True=load_model预计算RepEncoder向量跳过embedding层(灰区,评测真生效)
# 预计算三种实现在评测端均回退(无日志难诊断,推测评测调用顺序让load_model拿不到数据)。默认关
"precompute_rep": False, # True=load_model预计算RepEncoder向量(评测端三连回退,本地可跑见RISKS.md)
}