From 589d3b079dfc12ed4e565a6576bbe1d2d774d3e7 Mon Sep 17 00:00:00 2001 From: SSKJ Dev Date: Thu, 16 Jul 2026 09:05:37 +0000 Subject: [PATCH] enable initial backoff for all add16 benchmarks --- .../run_adaptive_concurrency_add16.sh | 3 +++ .../run_adaptive_concurrency_add16.sh | 3 +++ .../run_adaptive_concurrency_add16.sh | 3 +++ .../run_adaptive_concurrency_add16.sh | 3 +++ .../run_adaptive_concurrency_add16.sh | 3 +++ .../run_adaptive_concurrency_add16.sh | 3 +++ 6 files changed, 18 insertions(+) diff --git a/experiments/h20/dsv4_h20_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/h20/dsv4_h20_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh index 2f95353..fe1c162 100755 --- a/experiments/h20/dsv4_h20_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh +++ b/experiments/h20/dsv4_h20_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh @@ -180,5 +180,8 @@ export BENCH_DATASET_NAME DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP_MAX_REQUE export SEARCH_START_CONCURRENCY=16 export SEARCH_ADDEND=16 +# If the initial concurrency violates the TTFT SLO, search downward. Stop at +# the first acceptable value (16 -> 8; only try 1 when 8 still violates it). +export SEARCH_INITIAL_BACKOFF_CONCURRENCIES="8 1" adaptive_main "$@" diff --git a/experiments/h200/dsv4_h200_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/h200/dsv4_h200_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh index 0e6801f..29a7eca 100755 --- a/experiments/h200/dsv4_h200_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh +++ b/experiments/h200/dsv4_h200_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh @@ -180,5 +180,8 @@ export BENCH_DATASET_NAME DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP_MAX_REQUE export SEARCH_START_CONCURRENCY=16 export SEARCH_ADDEND=16 +# If the initial concurrency violates the TTFT SLO, search downward. Stop at +# the first acceptable value (16 -> 8; only try 1 when 8 still violates it). +export SEARCH_INITIAL_BACKOFF_CONCURRENCIES="8 1" adaptive_main "$@" diff --git a/experiments/h200/dsv4_h200_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/h200/dsv4_h200_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh index 9acbe5a..a30729e 100755 --- a/experiments/h200/dsv4_h200_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh +++ b/experiments/h200/dsv4_h200_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh @@ -163,5 +163,8 @@ export BENCH_DATASET_NAME DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP_MAX_REQUE export SEARCH_START_CONCURRENCY=16 export SEARCH_ADDEND=16 +# If the initial concurrency violates the TTFT SLO, search downward. Stop at +# the first acceptable value (16 -> 8; only try 1 when 8 still violates it). +export SEARCH_INITIAL_BACKOFF_CONCURRENCIES="8 1" adaptive_main "$@" diff --git a/experiments/p800/dsv4_p800_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/p800/dsv4_p800_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh index 0e6f32f..4e8e30c 100755 --- a/experiments/p800/dsv4_p800_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh +++ b/experiments/p800/dsv4_p800_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh @@ -148,5 +148,8 @@ export BENCH_DATASET_NAME CONTAINER_DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP export SEARCH_START_CONCURRENCY=16 export SEARCH_ADDEND=16 +# If the initial concurrency violates the TTFT SLO, search downward. Stop at +# the first acceptable value (16 -> 8; only try 1 when 8 still violates it). +export SEARCH_INITIAL_BACKOFF_CONCURRENCIES="8 1" adaptive_main "$@" diff --git a/experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh index b9b28f3..7876ff9 100755 --- a/experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh +++ b/experiments/pro6000/dsv4_pro6000_sglang_tp_dp_matrix/run_adaptive_concurrency_add16.sh @@ -180,5 +180,8 @@ export BENCH_DATASET_NAME DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP_MAX_REQUE export SEARCH_START_CONCURRENCY=16 export SEARCH_ADDEND=16 +# If the initial concurrency violates the TTFT SLO, search downward. Stop at +# the first acceptable value (16 -> 8; only try 1 when 8 still violates it). +export SEARCH_INITIAL_BACKOFF_CONCURRENCIES="8 1" adaptive_main "$@" diff --git a/experiments/pro6000/dsv4_pro6000_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/pro6000/dsv4_pro6000_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh index 482886d..8912882 100755 --- a/experiments/pro6000/dsv4_pro6000_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh +++ b/experiments/pro6000/dsv4_pro6000_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh @@ -163,5 +163,8 @@ export BENCH_DATASET_NAME DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP_MAX_REQUE export SEARCH_START_CONCURRENCY=16 export SEARCH_ADDEND=16 +# If the initial concurrency violates the TTFT SLO, search downward. Stop at +# the first acceptable value (16 -> 8; only try 1 when 8 still violates it). +export SEARCH_INITIAL_BACKOFF_CONCURRENCIES="8 1" adaptive_main "$@"