revert(agentx): restore legacy AIPerf integration (#2398)

Revert PR #2395 after upstream AIPerf 0.12 emitted empty message content during Weka warmup replay, causing Kimi K3 vLLM requests to fail with HTTP 400.

中文：回退 PR #2395，恢复旧版 AIPerf 集成。上游 AIPerf 0.12 在 Weka 预热回放中生成了空消息内容，导致 Kimi K3 vLLM 请求返回 HTTP 400。

Author: cquil11

.gitmodules

@@ -1,4 +1,4 @@
 [submodule "utils/aiperf"]
 	path = utils/aiperf
 	url = https://github.com/SemiAnalysisAI/aiperf.git
-	branch = main
+	branch = cquil11/aiperf-agentx-v1.0

benchmarks/benchmark_lib.sh

@@ -1729,14 +1729,19 @@ resolve_trace_source() {
 build_replay_cmd() {
     # aiperf invocation for the inferencex-agentx-mvp scenario.
     #
-    # Weka replay uses recorded assistant responses when constructing future
-    # prompts, preserving the captured hash-id and KV-prefix structure.
+    # Pre-canned assistant replay is the default: recorded assistant responses
+    # are used for future prompt construction, and live server responses are
+    # discarded. Set AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES=1 explicitly
+    # to use live-assistant mode, where the loader emits user-only deltas and
+    # the worker threads the server's live assistant response back into the
+    # session.
     #
     # The scenario plugin locks: --cache-bust first_turn_prefix and
-    # --system-idle-gap-cap-seconds 10. The global cap advances all pending
-    # timers only when the whole replay is idle; per-trace think times remain
-    # unchanged. The scenario forbids the old per-trace and per-turn delay
-    # caps. See utils/aiperf/docs/tutorials/agentx-mvp.md.
+    # --trace-idle-gap-cap-seconds 10 (per-trace idle-gap compression
+    # against parent + subagent request-start timestamps; supersedes the
+    # legacy --use-think-time-only / --inter-turn-delay-cap-seconds path),
+    # and auto-injects them — so we do not pass them. See
+    # utils/aiperf/docs/tutorials/agentx-mvp.md.
     local result_dir="$1"
     local duration="$DURATION"
     local cache_warmup_duration="${AIPERF_AGENTIC_CACHE_WARMUP_DURATION:-600}"
@@ -1749,6 +1754,7 @@ build_replay_cmd() {
         cache_warmup_duration=300
     fi
 
+    export AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES="${AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES:-0}"
     # Dataset configuration (load + reconstruct + inputs.json + mmap)
     # routinely takes 4-5 min for the Weka corpus on fast /tmp
     # (B300) but can stretch to 14 min on slower /tmp + parallel contention
@@ -1778,14 +1784,6 @@ build_replay_cmd() {
     # least one profile turn after warmup.
     REPLAY_CMD+=" --trajectory-start-min-ratio 0.25"
     REPLAY_CMD+=" --trajectory-start-max-ratio 0.75"
-    # Preserve the existing InferenceX traffic shape when moving to upstream
-    # aiperf: start all configured warmup/profiling lanes immediately instead
-    # of spreading their first requests by the recorded offset around t*.
-    REPLAY_CMD+=" --burst-phase-starts"
-    # Filtering can leave fewer Weka roots than requested concurrency. Reuse
-    # roots with independent replay points and cache-bust identities, matching
-    # the previous fork behavior instead of failing oversubscription.
-    REPLAY_CMD+=" --allow-dataset-wrap"
     # After the normal t* snapshot warmup, continue those exact trajectories
     # with one-token outputs and no idle delays. Profiling begins only after
     # those requests drain and resumes from the resulting live state.
@@ -1803,13 +1801,22 @@ build_replay_cmd() {
     # CPU on minimax-m2.5 at high concurrency. Lossless for vLLM (server
     # usage is authoritative).
     REPLAY_CMD+=" --use-server-token-count"
-    # Current Dynamo routers use session-affinity headers rather than the
-    # retired nvext.session_control request body. Stamp the stable correlation
-    # ID onto every root/child request; router-side TTLs live in the recipes.
-    # Individual recipes can explicitly disable the header.
-    if [[ "${FRAMEWORK:-}" == dynamo-* &&
-          -z "${AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID+x}" ]]; then
-        export AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID=true
+    # Dynamo's KV router needs an explicit conversation session binding to
+    # keep later turns on the prefill worker that owns their prefix blocks.
+    # X-Correlation-ID is useful tracing metadata but does not establish that
+    # binding by itself. AIPerf emits nvext.session_control bind/close actions
+    # keyed by the stable conversation correlation ID when this flag is set.
+    # Opt-out: recipes set AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING=0 to skip this.
+    # aiperf's conv-aware routing emits nvext.session_control, a removed POC field
+    # (dynamo #9920 / v1.3.0-dev) that current dynamo builds reject with a 400
+    # (they moved to router/routing_constraints/agent_context). Default stays on.
+    if [[ "${FRAMEWORK:-}" == dynamo-* && "${AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING:-1}" != "0" ]]; then
+        REPLAY_CMD+=" --use-dynamo-conv-aware-routing"
+        # The upstream 300s affinity TTL is shorter than an overloaded
+        # high-concurrency agentic request. Keep bindings alive across long
+        # prefills, generation, and capped inter-turn delay. This controls the
+        # router's inactivity lease; it does not relax HTTP/request failures.
+        REPLAY_CMD+=" --dynamo-session-timeout-seconds ${AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS:-3600}"
     fi
     # Disable DCGM GPU telemetry collection. aiperf's GpuMetricTimeSeries
     # freezes its metric schema on the first DCGM scrape, then KeyErrors when

benchmarks/multi_node/amd_utils/server_sglang.sh

@@ -988,7 +988,7 @@ if [ "$NODE_RANK" -eq 0 ]; then
                       WEKA_LOADER_OVERRIDE AIPERF_FAILED_REQUEST_THRESHOLD \
                       AIPERF_AGENTIC_CACHE_WARMUP_DURATION AIPERF_UNSAFE_OVERRIDE \
                       AIPERF_TRAJECTORY_START_MIN_RATIO AIPERF_TRAJECTORY_START_MAX_RATIO \
-                      ROUTER_PORT TQDM_MININTERVAL; do
+                      AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES ROUTER_PORT TQDM_MININTERVAL; do
                 if [[ -n "${!_v+x}" ]]; then
                     _val="${!_v}"
                     # docker --env-file requires one KEY=VALUE per line with no

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml

@@ -139,13 +139,16 @@ benchmark:
     # in process_agentic_result.py instead. TP must match sglang_config tp-size.
     IS_MULTINODE: "false"
     TP: "4"
-    # Enable header-based Dynamo session affinity so multi-turn requests retain
-    # KV locality. The retired nvext.session_control body is not used.
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    # Enable dynamo conv-aware routing (per-session affinity so multi-turn
+    # KV cache is reused across turns). Previously opted out because the
+    # frontend 400'd on aiperf's nvext.session_control actions; re-enabled
+    # to test with the current build.
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
     # Container-side path of the aiperf mmap dataset cache; host-side mount
     # wired via launch_gb300-*.sh srtslurm.yaml default_mounts. Without it,
     # aiperf re-tokenizes + re-writes the dataset mmap on every run.
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     # Persistent HF hub cache (also via default_mounts) so the trace dataset
     # isn't re-downloaded each run; overrides the workflow-level HF_HUB_CACHE.
     HF_HUB_CACHE: "/hf_hub_cache"
+

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1024-mtp-kvoffload.yaml

@@ -206,6 +206,7 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
+

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c8-mtp-kvoffload.yaml

@@ -205,6 +205,7 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
+

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml

@@ -206,6 +206,7 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
+

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c256-mtp-kvoffload.yaml

@@ -206,6 +206,7 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
+

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c512-mtp-kvoffload.yaml

@@ -206,6 +206,7 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
+

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml

@@ -140,7 +140,7 @@ benchmark:
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "false"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
     AIPERF_AGENTIC_CACHE_WARMUP_DURATION: "600"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"