Serendipity/CTI-Inference-Opt
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

revert: collate_dedup默认关(评测33.44>33.00,per_sample_weights加权kernel更慢+评测重复率不够)。锁定71.34

Browse Source
This commit is contained in:
OwnerSunshine530
2026-06-20 15:34:48 +08:00
parent 850930d761
commit 5488ad02fd
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
代码/code/infer.py
+2 -1
View File
@@ -168,7 +168,8 @@ CONFIG = {
"syncfree_mask": True, # True=用searchsorted构造因果mask(无同步)；False=repeat_interleave(同步) "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) "emb_fp16": True, # True=Embedding表转FP16(查表带宽减半,实测AUC 0.75932≈无损)
"use_embedding_bag": True, # F.embedding_bag 融合查表+池化(单kernel,消dedup的unique同步,AUC≈无损) "use_embedding_bag": True, # F.embedding_bag 融合查表+池化(单kernel,消dedup的unique同步,AUC≈无损)
"collate_dedup": True, # collate(不计时)段内去重+计数→embedding_bag per_sample_weights,减查表带宽(本地4.10→3.98,AUC精确不变) # 评测净负33.44>33.00:per_sample_weights走更慢的加权kernel+评测重复率不够,盖过带宽节省。退回。
"collate_dedup": False, # True=collate段内去重+计数(本地快评测慢,勿开)
"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 "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) "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） "compile": False, # 是否 torch.compile（实测慢5×，勿开）