From 2cf7f185fcf3960ee45cbe61f5b6c49ffd1adb22 Mon Sep 17 00:00:00 2001 From: OwnerSunshine530 Date: Wed, 17 Jun 2026 21:22:31 +0800 Subject: [PATCH] =?UTF-8?q?feat:=20=E9=BB=98=E8=AE=A4=E5=BC=80=E7=9C=9F?= =?UTF-8?q?=E7=A8=80=E7=96=8FMoE=20cap=3D2.0(=E6=9C=AC=E5=9C=B04.77->4.05s?= =?UTF-8?q?=20-15%,AUC=E5=BE=AE=E9=99=8D,PCOC1.105=E5=8C=BA=E9=97=B4?= =?UTF-8?q?=E5=86=85)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Opus 4.8 --- 代码/code/infer.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/代码/code/infer.py b/代码/code/infer.py index be23012..6f76508 100644 --- a/代码/code/infer.py +++ b/代码/code/infer.py @@ -144,8 +144,8 @@ CONFIG = { # 须靠提交验证。AUC中性、MoE仅占2%算力故风险极低。 "vectorize_moe": True, # True=稠密向量化MoE(无同步点)；False=原逐expert循环(.nonzero同步) "moe_baddbmm": True, # MoE FFN 用 baddbmm(cutlass GEMM+bias epilogue融合),省 bias add kernel - "moe_sparse": False, # True=真稀疏MoE(只算top-k,capacity分组),减GEMM~3x;风险:开销/容量丢弃AUC - "moe_capacity": 1.25, # 每expert容量 = ceil(Nk/E*factor);越大越不丢token但计算越多 + "moe_sparse": True, # 真稀疏MoE(只算top-k,capacity分组),本地4.77->4.05s(-15%),AUC微降无碍 + "moe_capacity": 2.0, # 每expert容量 = ceil(Nk/E*factor);cap=2.0 PCOC1.105在区间(1.25会炸到1.418) "skip_moe_loss": True, # 推理跳过 moe_loss(load-balance,推理无用),省 importance/std/mean kernel # PCOC 校准:本地拟合-0.1067(本地PCOC1.109),但评测PCOC稳定1.059,按斜率换算评测最优≈-0.059。 "logit_bias": -0.06, # logit 加常数偏移使评测 PCOC→~1.0(单调,AUC不变,免费+~0.33分)