zhoulin/zl_bench
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Default USE_FLAGTUNE to 0 and add fold_scale A/B script with trace shapes

Browse Source 
Plain autotune is the common case (fast, same-config A/B); FlagTune's
exhaustive first-run search is now opt-in via USE_FLAGTUNE=1.
ab_fold_test.sh automates the FLAGGEMS_MXFP4_FOLDSCALE=0/1 record-replay
comparison over the four DeepSeek-V4-Flash trace shape sets, which are
now tracked under runs/shape/.
This commit is contained in:
Linz
2026-07-20 16:50:59 +00:00
parent 26b071c6e1
commit 00902e6c67
12 changed files with 502 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
runs/shape/DeepSeek-V4-Flash-p1024d1024.yaml
+59
View File
@@ -0,0 +1,59 @@
fused_marlin_moe_mxfp4:
# 来源: DeepSeek-V4-Flash-p1024d1024.txt（55 个去重 shape，按 num_tokens 升序；行尾注释为 trace 中的调用次数）
shapes:
- [1, 256, 4096, 256, 6] # count: 172
- [2, 256, 4096, 256, 6] # count: 172
- [4, 256, 4096, 256, 6] # count: 172
- [8, 256, 4096, 256, 6] # count: 172
- [16, 256, 4096, 256, 6] # count: 172
- [24, 256, 4096, 256, 6] # count: 172
- [32, 256, 4096, 256, 6] # count: 172
- [40, 256, 4096, 256, 6] # count: 172
- [48, 256, 4096, 256, 6] # count: 172
- [56, 256, 4096, 256, 6] # count: 172
- [64, 256, 4096, 256, 6] # count: 172
- [72, 256, 4096, 256, 6] # count: 172
- [80, 256, 4096, 256, 6] # count: 172
- [88, 256, 4096, 256, 6] # count: 172
- [96, 256, 4096, 256, 6] # count: 172
- [104, 256, 4096, 256, 6] # count: 172
- [112, 256, 4096, 256, 6] # count: 172
- [120, 256, 4096, 256, 6] # count: 172
- [128, 256, 4096, 256, 6] # count: 172
- [136, 256, 4096, 256, 6] # count: 172
- [144, 256, 4096, 256, 6] # count: 172
- [152, 256, 4096, 256, 6] # count: 172
- [160, 256, 4096, 256, 6] # count: 172
- [168, 256, 4096, 256, 6] # count: 172
- [176, 256, 4096, 256, 6] # count: 172
- [184, 256, 4096, 256, 6] # count: 172
- [192, 256, 4096, 256, 6] # count: 172
- [200, 256, 4096, 256, 6] # count: 172
- [208, 256, 4096, 256, 6] # count: 172
- [216, 256, 4096, 256, 6] # count: 172
- [224, 256, 4096, 256, 6] # count: 172
- [232, 256, 4096, 256, 6] # count: 172
- [240, 256, 4096, 256, 6] # count: 172
- [248, 256, 4096, 256, 6] # count: 172
- [256, 256, 4096, 256, 6] # count: 172
- [272, 256, 4096, 256, 6] # count: 172
- [288, 256, 4096, 256, 6] # count: 172
- [304, 256, 4096, 256, 6] # count: 172
- [320, 256, 4096, 256, 6] # count: 172
- [336, 256, 4096, 256, 6] # count: 172
- [352, 256, 4096, 256, 6] # count: 172
- [368, 256, 4096, 256, 6] # count: 172
- [384, 256, 4096, 256, 6] # count: 172
- [400, 256, 4096, 256, 6] # count: 172
- [416, 256, 4096, 256, 6] # count: 172
- [432, 256, 4096, 256, 6] # count: 172
- [448, 256, 4096, 256, 6] # count: 172
- [464, 256, 4096, 256, 6] # count: 172
- [480, 256, 4096, 256, 6] # count: 172
- [496, 256, 4096, 256, 6] # count: 344
- [512, 256, 4096, 256, 6] # count: 344
- [1028, 256, 4096, 256, 6] # count: 215
- [2048, 256, 4096, 256, 6] # count: 43
- [3805, 256, 4096, 256, 6] # count: 215
- [16384, 256, 4096, 256, 6] # count: 1376
shape_desc: "num_tokens, num_experts, hidden_size, intermediate_size, topk"