Record successful 2048-token CUDA Graph retest and preserve failure samples

This commit is contained in:
2026-09-17 23:04:40 +08:00
parent 1e2f48d1a4
commit 48a1c9d7c4
19 changed files with 430 additions and 9 deletions
+5
View File
@@ -62,3 +62,8 @@
修正依据:
https://github.com/blazux/qwen3.8-Flash-DGX/blob/main/src/patch_mamba_block_size.py
## 后续 2048 预算复测
上述 CUDA Graph 未完成验收是首次实验的状态。后续 graph+前缀缓存组合已通过三轮复测,
短回答中位速度提升约 4.4%,默认仍保留 eager。详见 [复测报告](cuda-graph-retest.md)。