chore(agentx): move InferenceX to upstream-based AIPerf (#2395)

Point the AIPerf submodule at SemiAnalysisAI/aiperf main after syncing with upstream and applying the global-idle barrier fix.

Preserve the existing AgentX phase-start and dataset-wrap behavior explicitly, and migrate Dynamo affinity from retired nvext request controls to supported session headers and router-side TTLs.

中文：将 AIPerf 子模块切换到基于上游同步的 SemiAnalysisAI/aiperf main，并包含全局空闲上限在屏障延迟后的修复。显式保留 AgentX 阶段启动和数据集循环复用行为，同时将 Dynamo 会话亲和从已弃用的 nvext 请求控制迁移到受支持的会话请求头和路由器侧 TTL。

Signed-off-by: Cam Quilici <cjquilici@gmail.com>

Author: cquil11

.gitmodules

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

benchmarks/benchmark_lib.sh

@@ -1729,19 +1729,14 @@ resolve_trace_source() {
 build_replay_cmd() {
     # aiperf invocation for the inferencex-agentx-mvp scenario.
     #
-    # 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.
+    # Weka replay uses recorded assistant responses when constructing future
+    # prompts, preserving the captured hash-id and KV-prefix structure.
     #
     # The scenario plugin locks: --cache-bust first_turn_prefix and
-    # --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.
+    # --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.
     local result_dir="$1"
     local duration="$DURATION"
     local cache_warmup_duration="${AIPERF_AGENTIC_CACHE_WARMUP_DURATION:-600}"
@@ -1754,7 +1749,6 @@ 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
@@ -1784,6 +1778,14 @@ 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.
@@ -1801,22 +1803,13 @@ build_replay_cmd() {
     # CPU on minimax-m2.5 at high concurrency. Lossless for vLLM (server
     # usage is authoritative).
     REPLAY_CMD+=" --use-server-token-count"
-    # 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 ${AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS:-3600}"
+    # Current Dynamo routers use session-affinity headers rather than the
+    # retired nvext.session_control request body. Stamp the stable correlation
+    # ID onto every root/child request; router-side TTLs live in the recipes.
+    # Individual recipes can explicitly disable the header.
+    if [[ "${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
     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 [ "$NODE_RANK" -eq 0 ]; then
                       WEKA_LOADER_OVERRIDE AIPERF_FAILED_REQUEST_THRESHOLD \
                       AIPERF_AGENTIC_CACHE_WARMUP_DURATION AIPERF_UNSAFE_OVERRIDE \
                       AIPERF_TRAJECTORY_START_MIN_RATIO AIPERF_TRAJECTORY_START_MAX_RATIO \
-                      AIPERF_DATASET_WEKA_LIVE_ASSISTANT_RESPONSES ROUTER_PORT TQDM_MININTERVAL; do
+                      ROUTER_PORT TQDM_MININTERVAL; do
                 if [[ -n "${!_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

@@ -139,16 +139,13 @@ benchmark:
     # in process_agentic_result.py instead. TP must match sglang_config tp-size.
     IS_MULTINODE: "false"
     TP: "4"
-    # 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"
+    # 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"
     # 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

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

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

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

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

@@ -206,7 +206,6 @@ benchmark:
     RESULT_DIR: /logs/agentic
     PORT: "8000"
     IS_MULTINODE: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
     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_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "false"
     AIPERF_AGENTIC_CACHE_WARMUP_DURATION: "600"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"