perf(amd): port MI355X Qwen3.5 FP8 disagg sweep to srt-slurm

benchmarks/benchmark_lib.sh

@@ -106,28 +106,6 @@ exit_after_background_process_cleanup() {
     exit "$final_status"
 }
 
-# Finish preflight on every allocated node before any server container starts its
-# peer-readiness deadline. A failed node prevents the entire serving step.
-run_amd_multinode_after_preflight() {
-    local nodelist="$1" node_count="$2" preflight_script="$3"
-    local container_filter="$4" skip_gpu_sanity="$5"
-    shift 5
-    local preflight_rc
-    if srun --nodelist="$nodelist" \
-        --nodes="$node_count" --ntasks="$node_count" --ntasks-per-node=1 \
-        --kill-on-bad-exit=1 --unbuffered \
-        bash "$preflight_script" "$container_filter" "$skip_gpu_sanity"; then
-        echo "[preflight] all nodes ready; launching server containers"
-    else
-        preflight_rc=$?
-        echo "[preflight][ERROR] node preflight failed; no server containers launched" >&2
-        return "$preflight_rc"
-    fi
-    srun --nodelist="$nodelist" \
-        --nodes="$node_count" --ntasks="$node_count" --ntasks-per-node=1 \
-        --kill-on-bad-exit=1 --signal=TERM@30 --unbuffered "$@"
-}
-
 # Launchers may load only input validation, without benchmark initialization.
 if [[ "${1-}" == "--validation-only" ]]; then
     return 0
@@ -2097,7 +2075,9 @@ run_lm_eval() {
         export INFERENCEX_LM_EVAL_RUNTIME_READY=true
     fi
 
-    local openai_server_base="http://0.0.0.0:${port}"
+    # A routed eval client may run on a different node from the API server.
+    local openai_server_host="${EVAL_SERVER_HOST:-0.0.0.0}"
+    local openai_server_base="http://${openai_server_host}:${port}"
     local openai_chat_base="${openai_server_base}/v1/chat/completions"
     export OPENAI_API_KEY=${OPENAI_API_KEY}
     MODEL_NAME=${MODEL_NAME:-$MODEL} # Prefer MODEL_NAME, else MODEL

benchmarks/multi_node/agentic_srt.sh

@@ -7,7 +7,7 @@ set -x
 # against the already-ready frontend on the head node.
 
 source "$(dirname "${BASH_SOURCE[0]}")/../benchmark_lib.sh" --validation-only
-check_env_vars INFMAX_CONTAINER_WORKSPACE RESULT_DIR EVAL_ONLY AIPERF_DRAIN_TIMEOUT_SECONDS AIPERF_DRAIN_POLL_SECONDS
+check_env_vars INFMAX_CONTAINER_WORKSPACE RESULT_DIR EVAL_ONLY FRAMEWORK AIPERF_DRAIN_TIMEOUT_SECONDS AIPERF_DRAIN_POLL_SECONDS
 source "$INFMAX_CONTAINER_WORKSPACE/benchmarks/benchmark_lib.sh"
 
 if [[ -n "${SRT_FRONTEND_HOST:-}" ]]; then
@@ -62,6 +62,61 @@ if [[ "${EVAL_ONLY}" == "true" ]]; then
     _wait_for_openai_chat_route --port "$PORT"
 fi
 
+# Preserve the legacy DP-attention replay contract. The SGLang router uses this
+# header to keep every request in one AgentX correlation tree on a stable DP
+# route, which is important for both session continuity and prefix-cache reuse.
+if [[ "$FRAMEWORK" == sglang || "$FRAMEWORK" == sglang-disagg ]] && [[ "${PREFILL_DP_ATTN:-false}" == "true" ]]; then
+    export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
+fi
+
+# Reset every advertised SGLang worker before each concurrency point, matching
+# the retired amd_utils trace replay. /flush_cache covers GPU radix + host
+# HiCache; the storage-backend endpoint is best-effort because L3 is optional.
+clear_agentic_worker_caches() {
+    check_env_vars FLUSH_DRAIN_TIMEOUT
+    local timeout_seconds="$FLUSH_DRAIN_TIMEOUT"
+    local metrics_csv="${AIPERF_SERVER_METRICS_URLS:-}"
+    if [[ -z "$metrics_csv" ]]; then
+        echo "[clear_caches] WARN: AIPERF_SERVER_METRICS_URLS unset; skipping cache flush" >&2
+        return 0
+    fi
+
+    local -a metrics_urls
+    IFS=',' read -r -a metrics_urls <<< "$metrics_csv"
+    local metrics_url base_url start response code flushed
+    for metrics_url in "${metrics_urls[@]}"; do
+        [[ -n "$metrics_url" ]] || continue
+        base_url="${metrics_url%/metrics}"
+        start=$(date +%s)
+        flushed=0
+        response=""
+        while :; do
+            response=$(curl -sf -m 10 -X POST "${base_url}/flush_cache" 2>/dev/null || true)
+            if grep -qi "Cache flushed" <<< "$response"; then
+                flushed=1
+                break
+            fi
+            if (( $(date +%s) - start >= timeout_seconds )); then
+                break
+            fi
+            sleep 3
+        done
+        if (( flushed )); then
+            echo "[clear_caches] ${base_url}: L1+L2 flushed"
+        else
+            echo "[clear_caches] WARN ${base_url}: L1+L2 flush not confirmed after ${timeout_seconds}s" >&2
+        fi
+
+        code=$(curl -s -m 60 -o /dev/null -w '%{http_code}' \
+            -X POST "${base_url}/hicache/storage-backend/clear" 2>/dev/null || true)
+        if [[ "$code" == "200" ]]; then
+            echo "[clear_caches] ${base_url}: L3 store cleared"
+        else
+            echo "[clear_caches] ${base_url}: L3 clear http=${code:-000} (optional backend unavailable)"
+        fi
+    done
+}
+
 wait_for_agentic_servers_idle() {
     local timeout_seconds="${AIPERF_DRAIN_TIMEOUT_SECONDS}"
     local poll_seconds="${AIPERF_DRAIN_POLL_SECONDS}"
@@ -144,6 +199,9 @@ for index in "${!CONCURRENCIES[@]}"; do
     mkdir -p "$RESULT_DIR"
 
     echo "Running agentic concurrency $concurrency of: ${CONCURRENCIES[*]}"
+    if [[ "${CLEAR_CACHE_BETWEEN_CONC:-}" == "1" ]]; then
+        clear_agentic_worker_caches
+    fi
     build_replay_cmd "$RESULT_DIR"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"