update SEARCH_INITIAL_BACKOFF_CONCURRENCIES

experiments/h20/dsv4_h20_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 "$@"

scripts/common/adaptive_bench_lib.sh

@@ -77,6 +77,19 @@ adaptive_validate_config() {
     log "ERROR: ENABLE_TTFT_SLO_STOP must be 0 or 1, got ${ENABLE_TTFT_SLO_STOP}"
     return 1
   fi
+  local previous_backoff="$SEARCH_START_CONCURRENCY"
+  local backoff
+  for backoff in ${SEARCH_INITIAL_BACKOFF_CONCURRENCIES:-}; do
+    if ! [[ "$backoff" =~ ^[0-9]+$ ]] || (( backoff < 1 )); then
+      log "ERROR: SEARCH_INITIAL_BACKOFF_CONCURRENCIES must contain positive integers, got ${backoff}"
+      return 1
+    fi
+    if (( backoff >= previous_backoff )); then
+      log "ERROR: SEARCH_INITIAL_BACKOFF_CONCURRENCIES must be strictly descending below SEARCH_START_CONCURRENCY, got ${SEARCH_INITIAL_BACKOFF_CONCURRENCIES}"
+      return 1
+    fi
+    previous_backoff="$backoff"
+  done
 }
 
 adaptive_start_gpu_monitor() {
@@ -374,6 +387,7 @@ adaptive_run_shape() {
   local gain_pct=""
   local meaningful_gain=1
   local point_rc=0
+  local initial_backoff_active=0
 
   ADAPTIVE_RESTART_NEEDED=0
   ADAPTIVE_FATAL_WORKLOAD=0
@@ -423,6 +437,25 @@ adaptive_run_shape() {
             "$POINT_METRICS_FILE" "$tp" "$dp" "$mark" "$isl" "$osl" \
             "$concurrency" "$num_prompts" "$POINT_ATTEMPT" "$gain_pct" \
             "$plateau_streak" "$POINT_RAW_FILE" "$POINT_DETAIL_LOG"
+          local next_backoff=""
+          if (( concurrency == SEARCH_START_CONCURRENCY || initial_backoff_active == 1 )); then
+            local backoff
+            for backoff in ${SEARCH_INITIAL_BACKOFF_CONCURRENCIES:-}; do
+              if (( backoff < concurrency )); then
+                next_backoff="$backoff"
+                break
+              fi
+            done
+          fi
+          if [[ -n "$next_backoff" ]]; then
+            log "probe result c=${concurrency} total_tps=${current_tps} ttft_p95=${ttft_p95_ms}ms > SLO=${TTFT_SLO_MS}ms; retrying initial boundary at c=${next_backoff}"
+            initial_backoff_active=1
+            concurrency="$next_backoff"
+            previous_tps=""
+            plateau_streak=0
+            plateau_start_concurrency=""
+            continue
+          fi
           stop_reason="TTFT_SLO_EXCEEDED"
           shape_status="TTFT_BOUNDARY"
           log "probe result c=${concurrency} total_tps=${current_tps} ttft_p95=${ttft_p95_ms}ms > SLO=${TTFT_SLO_MS}ms, stopping shape"
@@ -436,6 +469,12 @@ adaptive_run_shape() {
         "$plateau_streak" "$POINT_RAW_FILE" "$POINT_DETAIL_LOG"
 
       log "probe result c=${concurrency} total_tps=${current_tps} gain_pct=${gain_pct:-first} plateau_streak=${plateau_streak}/${PLATEAU_PATIENCE}"
+      if (( initial_backoff_active == 1 )); then
+        stop_reason="TTFT_SLO_BACKOFF_FOUND"
+        shape_status="TTFT_BOUNDARY"
+        log "initial TTFT backoff found acceptable concurrency c=${concurrency}; stopping shape without probing lower values"
+        break
+      fi
       if (( plateau_streak >= PLATEAU_PATIENCE )); then
         saturation_concurrency="$plateau_start_concurrency"
         stop_reason="TPS_PLATEAU"
@@ -644,7 +683,7 @@ adaptive_main() {
   fi
 
   log "experiment=${EXPERIMENT_NAME} engine=${ENGINE} run_id=${RUN_ID}"
-  log "search_start=${SEARCH_START_CONCURRENCY} search_cap=${SEARCH_MAX_CONCURRENCY} multiplier=${SEARCH_MULTIPLIER} min_gain_pct=${TPS_MIN_GAIN_PCT} patience=${PLATEAU_PATIENCE} prompts_multiplier=${NUM_PROMPTS_MULTIPLIER} warmup_cap=${BENCH_WARMUP_MAX_REQUESTS}"
+  log "search_start=${SEARCH_START_CONCURRENCY} search_cap=${SEARCH_MAX_CONCURRENCY} multiplier=${SEARCH_MULTIPLIER} initial_backoff=[${SEARCH_INITIAL_BACKOFF_CONCURRENCIES:-none}] min_gain_pct=${TPS_MIN_GAIN_PCT} patience=${PLATEAU_PATIENCE} prompts_multiplier=${NUM_PROMPTS_MULTIPLIER} warmup_cap=${BENCH_WARMUP_MAX_REQUESTS}"
   log "dataset=${BENCH_DATASET_NAME} random_range_ratio=${RANDOM_RANGE_RATIO} tokenize_prompt=${tokenize_prompt_json} matrix=${MATRIX_FILE} mode=${MATRIX_MODE}"
 
   "$PYTHON" - \
@@ -660,6 +699,7 @@ adaptive_main() {
     "$SEARCH_START_CONCURRENCY" \
     "$SEARCH_MAX_CONCURRENCY" \
     "$SEARCH_MULTIPLIER" \
+    "${SEARCH_INITIAL_BACKOFF_CONCURRENCIES:-}" \
     "$NUM_PROMPTS_MULTIPLIER" \
     "$TPS_MIN_GAIN_PCT" \
     "$PLATEAU_PATIENCE" \
@@ -671,8 +711,9 @@ import json, sys
 
 (experiment, engine, run_id, model, hardware, matrix, dataset,
  tokenize_prompt, random_range_ratio, search_start, search_cap,
- search_multiplier, prompts_multiplier, min_gain_pct, plateau_patience,
+ search_multiplier, initial_backoff_concurrencies, prompts_multiplier,
- warmup_max_requests, ttft_slo_ms, enable_ttft_slo_stop) = sys.argv[1:19]
+ min_gain_pct, plateau_patience, warmup_max_requests, ttft_slo_ms,
+ enable_ttft_slo_stop) = sys.argv[1:20]
 
 print(json.dumps({
     "experiment": experiment,
@@ -688,6 +729,9 @@ print(json.dumps({
         "start_concurrency": int(search_start),
         "max_concurrency": int(search_cap),
         "multiplier": int(search_multiplier),
+        "initial_backoff_concurrencies": [
+            int(value) for value in initial_backoff_concurrencies.split()
+        ],
         "num_prompts_multiplier": int(prompts_multiplier),
         "min_tps_gain_pct": float(min_gain_pct),
         "plateau_patience": int(plateau_patience),