Skip to content

Commit 0e363a6

Browse files
committed
refactor: drop conc-512-tuned warmup grace override from glm5.2 b300 recipe / 移除 glm5.2 b300 recipe 中为 conc 512 调优的 warmup grace 覆盖值
AGENTIC_WARMUP_GRACE_PERIOD=3600 was tuned for the conc-512 DP8 point, which this PR removes (DEP arm now caps at conc 64). At conc 64 the warmup working set drains well within the default 1800s (warmup-grace-period is a ceiling, not a fixed sleep), so fall back to the default and delete the now-stale comment. No behavioral change at conc <= 64. 中文:AGENTIC_WARMUP_GRACE_PERIOD=3600 是为 conc 512 的 DP8 数据点调优的,而本 PR 已移除该点(DEP 分支并发上限现为 64)。在 conc 64 下,warmup 工作集会在默认 1800s 内完成 drain(warmup-grace-period 是上限而非固定等待),故回退到默认值并 删除已过时的注释。conc <= 64 下无行为变化。
1 parent 6d088b0 commit 0e363a6

1 file changed

Lines changed: 0 additions & 5 deletions

File tree

benchmarks/single_node/agentic/glm5.2_fp4_b300_sglang.sh

Lines changed: 0 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -106,11 +106,6 @@ if [ "$DP_ATTENTION" = "true" ]; then
106106
# while KV usage sat at ~0.01). Use the cookbook's own dp8 lever from
107107
# the B200 cells (32768 = ~4096/rank).
108108
CHUNKED_PREFILL_SIZE=32768
109-
# At conc 512 the saturation working set outlives the default 1800s
110-
# warmup drain grace: the drain converges healthily (~0.45 req/s, zero
111-
# errors) but needs ~2500s end to end. 3600 is a maximum wait, not a
112-
# fixed sleep — lower-conc DPA points still finish as fast as they drain.
113-
export AGENTIC_WARMUP_GRACE_PERIOD=3600
114109
PARALLEL_ARGS+=(
115110
--dp "$TP"
116111
--enable-dp-attention

0 commit comments

Comments
 (0)