From b991f9e78eea65af10b616fb688cba05e9bf30b5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E5=88=98=E8=88=AA=E5=AE=87?= <3364451258@qq.com> Date: Sat, 13 Jun 2026 13:29:21 +0800 Subject: [PATCH] =?UTF-8?q?docs:=20=E6=9B=B4=E6=96=B0=E6=8F=90=E4=BA=A4?= =?UTF-8?q?=E8=AE=B0=E5=BD=95=EF=BC=88=E6=B6=88=E9=99=A4=20GPU=20=E5=90=8C?= =?UTF-8?q?=E6=AD=A5=EF=BC=8C58.49=20=E5=88=86=EF=BC=8C88.1s=EF=BC=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- CLAUDE.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/CLAUDE.md b/CLAUDE.md index d50610d..0c38f98 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -168,6 +168,7 @@ Baseline 数据:推理 229s,AUC 0.759,PCOC 1.110,得分 25.85。 5. ❌ **torch.compile** — reduce-overhead 和 default 模式均因动态 batch 形状反效果,彻底放弃 6. ❌ **MoE Top-1 gating** — PCOC 从 1.059 炸到 2.075,已回退 7. ❌ **2:4 结构化稀疏** — PCOC 炸到 2.067,耗时反增 265s。to_sparse_semi_structured 与 nn.Linear 不兼容 +8. ✅ **SMoE 消除 GPU 同步** — 移除 mask.any()(64 次 GPU→CPU 同步/forward),88.1s CUDA Graph 已评估并放弃(batch 形状不固定,不适用)。 @@ -190,8 +191,9 @@ CUDA Graph 已评估并放弃(batch 形状不固定,不适用)。 | 日期 | 提交次数 | 得分 | AUC | PCOC | 耗时 | 优化手段 | 备注 | |------|----------|------|-----|------|------|----------|------| +| 06/13 | 12 | **58.49** | 0.7526 | 1.059 | 88.1s | + SMoE 消除 GPU 同步 | **当前最优** | | 06/13 | 11 | 0 | 0.748 | 2.067 | 265.5s | 2:4 sparse | ❌ 炸毁 | -| 06/13 | 10 | **57.45** | 0.7526 | 1.059 | 92.5s | + inference_mode | **当前最优** | +| 06/13 | 10 | 57.45 | 0.7526 | 1.059 | 92.5s | + inference_mode | | | 06/13 | 9 | 51.42 | 0.7525 | 1.059 | 118.4s | + compile(default) | 反效果 | | 06/12 | 8 | 0 | 0.736 | 2.075 | 119.6s | MoE k=1 + compile | PCOC 炸毁 | | 06/12 | 6 | 56.98 | 0.7526 | 1.059 | 94.5s | + Flash Attention | |