From c8a42b66ee4533354dcffdf57db3c0c8024360c3 Mon Sep 17 00:00:00 2001 From: Linz Date: Mon, 20 Jul 2026 16:53:14 +0000 Subject: [PATCH] Move trace shape sets to top-level shape/ and frame ab_fold_test.sh as A/B reference example --- .gitignore | 3 +-- README.md | 6 +++--- ab_fold_test.sh | 8 +++++--- {runs/shape => shape}/DeepSeek-V4-Flash-p1024d1024.txt | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p1024d1024.yaml | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p32768d1024.txt | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p32768d1024.yaml | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p4096d1024.txt | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p4096d1024.yaml | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p65536d1024.txt | 0 {runs/shape => shape}/DeepSeek-V4-Flash-p65536d1024.yaml | 0 11 files changed, 9 insertions(+), 8 deletions(-) rename {runs/shape => shape}/DeepSeek-V4-Flash-p1024d1024.txt (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p1024d1024.yaml (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p32768d1024.txt (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p32768d1024.yaml (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p4096d1024.txt (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p4096d1024.yaml (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p65536d1024.txt (100%) rename {runs/shape => shape}/DeepSeek-V4-Flash-p65536d1024.yaml (100%) diff --git a/.gitignore b/.gitignore index 67115f5..93cd52f 100644 --- a/.gitignore +++ b/.gitignore @@ -1,3 +1,2 @@ -runs/* -!runs/shape/ __pycache__/ +runs/ diff --git a/README.md b/README.md index d7eb9ac..8689dee 100644 --- a/README.md +++ b/README.md @@ -85,16 +85,16 @@ A/B 两侧的 `USE_FLAGTUNE` 必须取同值：该开关决定调优空间，两 replay 的兜底行为：B 侧遇到记录中没有的 key、或记录的 config 在新编译器下编译失败时，自动回退到现场 autotune 并在 `run.log` 打 `AUTOTUNE_REPLAY_FALLBACK reason=...` 标记——出现该标记的测量点不再满足"同 config"前提，解读时注意。另外 replay 模式的 run 目录不产生 `autotune_records/`，后续 run 的 `REPLAY_FROM` 应始终指向最初 record 的那次 A 侧目录，不要链式指向 replay 产物。 -### ab_fold_test.sh：FOLD 开关 A/B（现成示例） +### ab_fold_test.sh：A/B 测试参考示例 -`ab_fold_test.sh` 把上述流程封装成 FlagGems `FLAGGEMS_MXFP4_FOLDSCALE`（fold_scale 优化）开关的 A/B 验证：对 4 个 DeepSeek-V4-Flash 真实 trace shape 集（`runs/shape/DeepSeek-V4-Flash-p{1024,4096,32768,65536}d1024.yaml`，随仓库提供）各跑一对 FOLD=0（基线，record）→ FOLD=1（replay 同 config），共 8 轮： +`ab_fold_test.sh` 是上述 A/B 流程的现成参考：以 FlagGems `FLAGGEMS_MXFP4_FOLDSCALE`（fold_scale 优化）开关为对比对象，对 4 个 DeepSeek-V4-Flash 真实 trace shape 集（`shape/DeepSeek-V4-Flash-p{1024,4096,32768,65536}d1024.yaml`，随仓库提供）各跑一对 FOLD=0（基线，record）→ FOLD=1（replay 同 config），共 8 轮： ```bash bash ab_fold_test.sh # 汇总 log：runs/ab_fold_<时间戳>.log（已去色）；快速对账：grep -E '^#####|FAILED' ``` -对比其它开关 / 算子时可直接改脚本内的循环变量与环境变量，A/B 口径（同 config replay、cudagraph 计时）无需改动。 +要对比其它开关 / 编译器改动 / 算子，套用该脚本改循环变量与环境变量即可，A/B 口径（同 config replay、cudagraph 计时）无需改动。 ### cudagraph 计时的 warmup、回退与精度 diff --git a/ab_fold_test.sh b/ab_fold_test.sh index 7526cab..04708e4 100755 --- a/ab_fold_test.sh +++ b/ab_fold_test.sh @@ -1,7 +1,9 @@ #!/bin/bash -# FOLD 开关 A/B 验证脚本（4 个 DeepSeek-V4-Flash trace yaml × FOLD=0/1 共 8 轮）。 +# A/B 测试参考示例：FLAGGEMS_MXFP4_FOLDSCALE（fold_scale）开关验证 +# （4 个 DeepSeek-V4-Flash trace yaml × FOLD=0/1 共 8 轮）。 # 口径：USE_FLAGTUNE=0；FOLD=0（基线）先跑并 record autotune config，FOLD=1 用 -# REPLAY_FROM 回放同一套 config（两侧同配置比 wall-time，见实施记录 §5）。 +# REPLAY_FROM 回放同一套 config（两侧同配置比 wall-time）。 +# 对比其它开关/算子时，改循环变量与环境变量即可复用同一口径。 # 全部输出汇总到 runs/ab_fold_<时间戳>.log；各轮 runs/_<时间戳>/ 产物照常存档。 # 用法：bash ab_fold_test.sh set -uo pipefail @@ -17,7 +19,7 @@ fi LOG="runs/ab_fold_$(date +%Y%m%d-%H%M%S).log" { - for y in runs/shape/DeepSeek-V4-Flash-p{1024,4096,32768,65536}d1024.yaml; do + for y in shape/DeepSeek-V4-Flash-p{1024,4096,32768,65536}d1024.yaml; do echo "########## $y | FOLD=0 baseline (record) ##########" SHAPE_FILE="$y" USE_FLAGTUNE=0 FLAGGEMS_MXFP4_FOLDSCALE=0 bash run_pytest.sh \ || { echo "!!! FAILED fold=0: $y"; continue; } diff --git a/runs/shape/DeepSeek-V4-Flash-p1024d1024.txt b/shape/DeepSeek-V4-Flash-p1024d1024.txt similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p1024d1024.txt rename to shape/DeepSeek-V4-Flash-p1024d1024.txt diff --git a/runs/shape/DeepSeek-V4-Flash-p1024d1024.yaml b/shape/DeepSeek-V4-Flash-p1024d1024.yaml similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p1024d1024.yaml rename to shape/DeepSeek-V4-Flash-p1024d1024.yaml diff --git a/runs/shape/DeepSeek-V4-Flash-p32768d1024.txt b/shape/DeepSeek-V4-Flash-p32768d1024.txt similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p32768d1024.txt rename to shape/DeepSeek-V4-Flash-p32768d1024.txt diff --git a/runs/shape/DeepSeek-V4-Flash-p32768d1024.yaml b/shape/DeepSeek-V4-Flash-p32768d1024.yaml similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p32768d1024.yaml rename to shape/DeepSeek-V4-Flash-p32768d1024.yaml diff --git a/runs/shape/DeepSeek-V4-Flash-p4096d1024.txt b/shape/DeepSeek-V4-Flash-p4096d1024.txt similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p4096d1024.txt rename to shape/DeepSeek-V4-Flash-p4096d1024.txt diff --git a/runs/shape/DeepSeek-V4-Flash-p4096d1024.yaml b/shape/DeepSeek-V4-Flash-p4096d1024.yaml similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p4096d1024.yaml rename to shape/DeepSeek-V4-Flash-p4096d1024.yaml diff --git a/runs/shape/DeepSeek-V4-Flash-p65536d1024.txt b/shape/DeepSeek-V4-Flash-p65536d1024.txt similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p65536d1024.txt rename to shape/DeepSeek-V4-Flash-p65536d1024.txt diff --git a/runs/shape/DeepSeek-V4-Flash-p65536d1024.yaml b/shape/DeepSeek-V4-Flash-p65536d1024.yaml similarity index 100% rename from runs/shape/DeepSeek-V4-Flash-p65536d1024.yaml rename to shape/DeepSeek-V4-Flash-p65536d1024.yaml