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

feat/auc-recovery-plan #1

Merged
Serendipity merged 20 commits from feat/auc-recovery-plan into main 2026-06-15 12:33:32 +08:00
Conversation 0 Commits 20 Files Changed 6
+4 -3
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit 0f359288a1 - Show all commits
代码/code/infer.py
+4 -3
View File
@@ -40,9 +40,10 @@ CONFIG = {
"signid_mode": "clamp", # "clamp" 或 "modulo"：处理超界 sign id 的方式 "signid_mode": "clamp", # "clamp" 或 "modulo"：处理超界 sign id 的方式
"sync_timing": False, # bench 里设 True，做 torch.cuda.synchronize 真实计时 "sync_timing": False, # bench 里设 True，做 torch.cuda.synchronize 真实计时
"filter_test_users": True, # 只处理含测试样本的用户（跳过会被丢弃的用户，省算力） "filter_test_users": True, # 只处理含测试样本的用户（跳过会被丢弃的用户，省算力）
# 实测(A800)：sdpa+loop=15.1s 最快；flex/dense/compile/小batch 都更慢。 # 实测(A800,本地5451用户)：sdpa=15.15s，varlen=10.28s(快32%,AUC不变)，
# attn: "sdpa"(稠密mask,默认/已验证) / "flex"(FlexAttention,慢) / "varlen"(嵌套张量变长flash) # flex/compile/小batch 都更慢。默认 varlen。
"attn": "sdpa", # attn: "varlen"(嵌套张量变长flash,默认) / "sdpa"(稠密mask) / "flex"(FlexAttention)
"attn": "varlen",
"vectorize_moe": False, # True=稠密向量化MoE；False=原逐expert循环（默认，已验证更快） "vectorize_moe": False, # True=稠密向量化MoE；False=原逐expert循环（默认，已验证更快）
"compile": False, # 是否 torch.compile（实测慢5×，勿开） "compile": False, # 是否 torch.compile（实测慢5×，勿开）
} }