FrankD412 · Oct 20, 2026
@@ -94,8 +94,8 @@ PRs are welcome! See [CONTRIBUTING.md](CONTRIBUTING.md) for more details on the
 For maintainer and agent instructions, start with [`AGENTS.md`](AGENTS.md). It links the configuration, workflow, eval, and recovery references.
 
 To benchmark an existing server without CI or Slurm, see
-[Run AgentX-Harness Standalone](inferencex-e2e/docs/agentx-standalone.md)
-for client installation and a direct `agentx profile` command.
+[Run AgentX with AIPerf Standalone](inferencex-e2e/docs/agentx-standalone.md)
+for client installation and a direct `aiperf profile` command.
 
 ## Acknowledgements & Supporters
 Thank you to Lisa Su and Anush Elangovan for providing the MI355X and CDNA3 GPUs for this free and open-source project. We want to recognize the many AMD contributors for their responsiveness and for debugging, optimizing, and validating performance across AMD GPUs. 

@@ -93,8 +93,8 @@ SGLang、vLLM、TensorRT-LLM、CUDA、ROCm 等 AI 软件通过核函式優化、
 维护者与 Agent 请从 [`AGENTS.md`](AGENTS.md) 开始阅读，其中链接了配置、Workflow、评估与恢复操作参考。
 
 无需 CI 或 Slurm 即可对已有服务进行基准测试，请参阅
-[独立运行 AgentX-Harness](inferencex-e2e/docs/agentx-standalone.md)，
-其中包含客户端安装步骤和直接执行的 `agentx profile` 命令。
+[使用 AIPerf 独立运行 AgentX](inferencex-e2e/docs/agentx-standalone.md)，
+其中包含客户端安装步骤和直接执行的 `aiperf profile` 命令。
 
 ## 致谢与支持者
 感谢 Lisa Su 与 Anush Elangovan 为这一免费开源项目提供 MI355X 与 CDNA3 GPU。我们也要感谢众多 AMD 贡献者的积极响应，以及他们在各类 AMD GPU 上进行调试、优化与性能验证所付出的努力。

@@ -1,19 +1,18 @@
-# Run AgentX-Harness Standalone
+# Run AgentX with AIPerf Standalone
 
 Run the AgentX client against an already-running OpenAI-compatible server.
 
 ## Install
 
-Install the pinned PyPI release in a Python 3.11–3.13 virtual environment:
+Install the pinned upstream AIPerf commit in a Python 3.11–3.13 virtual environment:
 
 ```bash
 python3 -m venv .venv-agentx
 source .venv-agentx/bin/activate
-python -m pip install 'agentx-harness==1.0.12'
+python -m pip install 'aiperf @ https://github.com/ai-dynamo/aiperf/archive/8aa7039f872cd029f05a4c313fb8759e9a482cd3.tar.gz'
 ```
 
-The package version matches the `agentx-v1.0.12` GitHub release. The `agentx`
-command selects the `agentx` scenario by default.
+The commit matches `inferencex-e2e/infx/bench/agentic/requirements.txt`.
 
 ## Run
 
@@ -26,20 +25,32 @@ DATASET="semianalysis_cc_traces_weka_062126"
 CONC=8
 OUTPUT_DIR="$PWD/results/agentx-c${CONC}-$(date -u +%Y%m%dT%H%M%SZ)"
 
+export AIPERF_DATASET_CONFIGURATION_TIMEOUT=1800
+export AIPERF_SERVICE_PROFILE_CONFIGURE_TIMEOUT=1800
+export AIPERF_UI_REALTIME_METRICS_ENABLED=true
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+
 mkdir -p "$OUTPUT_DIR"
-agentx profile \
+aiperf profile --scenario inferencex-agentx-mvp \
   --url "$SERVER_URL" --endpoint /v1/chat/completions \
   --model "$SERVED_MODEL_NAME" --tokenizer "$TOKENIZER" \
   --tokenizer-trust-remote-code \
-  --concurrency "$CONC" \
+  --concurrency "$CONC" --benchmark-duration 3600 \
+  --endpoint-type chat --streaming --stats-interval 30 --random-seed 42 \
+  --trajectory-start-min-ratio 0.25 --trajectory-start-max-ratio 0.75 \
+  --trace-idle-gap-cap-seconds 300 --use-server-token-count --no-gpu-telemetry \
+  --num-dataset-entries 393 --slice-duration 1.0 \
+  --failed-request-threshold 0.10 --warmup-requests-per-lane 10 --warmup-grace-period 1800 \
   --public-dataset "$DATASET" \
   --output-artifact-dir "$OUTPUT_DIR/aiperf_artifacts" \
   2>&1 | tee "$OUTPUT_DIR/aiperf.log"
 ```
 
-The `agentx` preset supplies a one-hour profile, ten warmup requests per lane,
-and shared reporting and runtime settings. Explicit CLI arguments and `AIPERF_*`
-environment variables override preset defaults. Model, tokenizer, dataset,
+The `inferencex-agentx-mvp` scenario locks the replay invariants: agentic replay
+timing, streaming, `ignore_eos`, no input truncation, first-turn cache busting,
+the accepted trace loaders and a 900-second minimum profile. The remaining flags
+and `AIPERF_*` variables are the AgentX run settings InferenceX passes for every
+point; keep them when reproducing a result. Model, tokenizer, dataset,
 concurrency, and output remain caller inputs; `--tokenizer-trust-remote-code` is
 needed only for tokenizers with custom code.
 
@@ -49,9 +60,9 @@ with `dsr1`, `qwen3.5`, or `qwen3.8next` explicitly select the
 256k-capped corpus (`semianalysis_cc_traces_weka_062126_256k`). New model families
 use the full corpus unless added to `CAPPED_FAMILIES` in
 `infx/bench/agentic/traces.py`. `WEKA_LOADER_OVERRIDE` can pin either loader
-regardless of model family; it does not configure the standalone `agentx` CLI.
+regardless of model family; it does not configure the standalone `aiperf` CLI.
 
-The default random seed is **42**. Keep this seed when reproducing AgentX results,
+The random seed is **42**. Keep this seed when reproducing AgentX results,
 or pass `--random-seed 123` to use a different seed.
 
 Allow additional time for dataset preparation, warmup, and draining requests.
@@ -64,7 +75,7 @@ Add `--server-metrics "${SERVER_URL}/metrics"` if the server exposes metrics.
 ## Optional session-affinity flags
 
 For DP-attention runs requiring router affinity, export these before
-`agentx profile`. They add headers without replacing `X-Correlation-ID`.
+`aiperf profile`. They add headers without replacing `X-Correlation-ID`.
 
 ```bash
 # Adds X-Dynamo-Session-ID; subagents also get X-Dynamo-Parent-Session-ID.
@@ -73,3 +84,6 @@ export AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID=true
 # Also sends X-Session-ID, if a front-end router needs it.
 export AIPERF_HTTP_X_SESSION_ID_FROM_CORRELATION_ID=true
 ```
+
+AIPerf sends `X-Session-Affinity` with the correlation ID by default; set
+`AIPERF_HTTP_X_SESSION_AFFINITY_FROM_CORRELATION_ID=false` to omit it.
@@ -21,6 +21,21 @@
     "AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING",
 )
 
+SCENARIO = "inferencex-agentx-mvp"
+# AIPerf's scenario locks invariants but leaves these to the caller.
+PRESET_FLAGS = (
+    "--endpoint-type", "chat", "--streaming", "--stats-interval", "30", "--random-seed", "42",
+    "--trajectory-start-min-ratio", "0.25", "--trajectory-start-max-ratio", "0.75",
+    "--trace-idle-gap-cap-seconds", "300", "--use-server-token-count", "--no-gpu-telemetry",
+    "--num-dataset-entries", "393", "--slice-duration", "1.0",
+)  # fmt: skip
+PRESET_ENVIRONMENT = {
+    "AIPERF_DATASET_CONFIGURATION_TIMEOUT": "1800",
+    "AIPERF_SERVICE_PROFILE_CONFIGURE_TIMEOUT": "1800",
+    "AIPERF_UI_REALTIME_METRICS_ENABLED": "true",
+    "AIPERF_HTTP_TCP_USER_TIMEOUT": "900000",
+}
+
 MIN_SCENARIO_DURATION_S = 900
 FAST_DURATION_S = 1200
 FAST_WARMUP_REQUESTS_PER_LANE = "1"
@@ -135,17 +150,20 @@ def _server_metrics_urls(env: Mapping[str, str]) -> tuple[str, ...]:
     return tuple(parts)
 
 
+def replay_environ(environ: Mapping[str, str]) -> dict[str, str]:
+    """``environ`` over the preset's AIPerf runtime settings; the caller's values win."""
+    return {**PRESET_ENVIRONMENT, **environ}
+
+
 def replay_argv(cfg: ReplayConfig, cli: str | Path) -> list[str]:
-    """The ``agentx profile`` argv for ``cfg``; ``cli`` is the ``agentx`` executable."""
-    argv = [str(cli), "profile", "--url", cfg.url]
+    """The ``aiperf profile`` argv for ``cfg``; ``cli`` is the ``aiperf`` executable."""
+    argv = [str(cli), "profile", "--scenario", SCENARIO, "--url", cfg.url]
     argv += ["--model", cfg.model, "--tokenizer", cfg.tokenizer]
     argv += ["--concurrency", str(cfg.concurrency), "--benchmark-duration", str(cfg.duration)]
-    if cfg.live_failed_request_threshold is not None:
-        argv += ["--failed-request-threshold", cfg.live_failed_request_threshold]
-    if cfg.warmup_requests_per_lane is not None:
-        argv += ["--warmup-requests-per-lane", cfg.warmup_requests_per_lane]
-    if cfg.warmup_grace_period is not None:
-        argv += ["--warmup-grace-period", cfg.warmup_grace_period]
+    argv += PRESET_FLAGS
+    argv += ["--failed-request-threshold", cfg.live_failed_request_threshold or "0.10"]
+    argv += ["--warmup-requests-per-lane", cfg.warmup_requests_per_lane or "10"]
+    argv += ["--warmup-grace-period", cfg.warmup_grace_period or "1800"]
     if cfg.extra_inputs:
         argv += ["--extra-inputs", *cfg.extra_inputs]
     if cfg.dynamo_session_timeout is not None:

@@ -1,2 +1,2 @@
-agentx-harness==1.0.12
+aiperf @ https://github.com/ai-dynamo/aiperf/archive/8aa7039f872cd029f05a4c313fb8759e9a482cd3.tar.gz
 datasets>=4.7.0
@@ -17,6 +17,7 @@
     REQUIRED as REPLAY_REQUIRED,
     ReplayConfig,
     replay_argv,
+    replay_environ,
 )
 from infx.bench.agentic.venv import Runtime, bootstrap
 
@@ -133,11 +134,11 @@ def execute(plan: Plan, runtime: Runtime, environ: Mapping[str, str]) -> int:
     rc = _open_power_window(plan, python, env)
     if rc:
         return rc
-    argv = replay_argv(cfg, runtime.agentx)
+    argv = replay_argv(cfg, runtime.aiperf)
     (plan.result_dir / "benchmark_command.txt").write_text(f"{shlex.join(argv)}\n")
     log = plan.result_dir / "benchmark.log"
     with proc.DeferSignals() as signals:
-        replay_rc = 0 if signals.received else proc.tee(argv, log, env)
+        replay_rc = 0 if signals.received else proc.tee(argv, log, replay_environ(env))
     if signals.received:
         return 128 + signals.received
     return _score(plan, python, env, replay_rc)
@@ -247,7 +248,7 @@ def main(argv: list[str]) -> int:
     plan = Plan.from_env(os.environ)
     runtime = Runtime.for_job(os.environ)
     if not runtime.active():
-        print(f"Preparing the AgentX runtime in {runtime.venv}", flush=True)
+        print(f"Preparing the AIPerf runtime in {runtime.venv}", flush=True)
         rc = bootstrap(runtime)
         if rc:
             return rc

@@ -1,4 +1,4 @@
-"""The isolated AgentX runtime: a per-job uv venv with the pinned PyPI release."""
+"""The isolated AIPerf runtime: a per-job uv venv with the pinned upstream commit."""
 
 from __future__ import annotations
 
@@ -20,7 +20,7 @@
 
 @dataclass(frozen=True)
 class Runtime:
-    """Paths of one job's AgentX runtime directory."""
+    """Paths of one job's AIPerf runtime directory."""
 
     root: Path
 
@@ -42,8 +42,8 @@ def python(self) -> Path:
         return self.venv / "bin" / "python"
 
     @property
-    def agentx(self) -> Path:
-        return self.venv / "bin" / "agentx"
+    def aiperf(self) -> Path:
+        return self.venv / "bin" / "aiperf"
 
     @property
     def hf(self) -> Path:
@@ -77,8 +77,8 @@ def bootstrap(runtime: Runtime) -> int:
             "benchmark client dependency bootstrap failed; inspect network/package "
             "resolution before recipe repairs"
         )
-    if not (_executable(runtime.agentx) and _executable(runtime.hf)):
-        raise BenchError(f"isolated AgentX environment is incomplete at {runtime.venv}")
+    if not (_executable(runtime.aiperf) and _executable(runtime.hf)):
+        raise BenchError(f"isolated AIPerf environment is incomplete at {runtime.venv}")
     return 0
 
 

@@ -3,7 +3,7 @@
 Reads <in-dir>/<session_id>.jsonl produced by `sample_proxy_traces.py`
 and writes <in-dir>/../<out-dir>/<session_id>.json in the v1 weka trace
 format consumed by the kv-cache-tester replayer (see
-agentx-harness: src/aiperf/dataset/loader/weka_trace_models.py).
+aiperf: src/aiperf/dataset/loader/weka_trace_models.py).
 
 Subagent grouping mirrors the conversation-view algorithm from the
 SemiAnalysis claude-code-proxy:

@@ -28,7 +28,7 @@
 AGENTIC = "agentic-coding"
 _SHA = re.compile(r"[0-9a-f]{40}(?:[0-9a-f]{24})?")
 _PINNED = re.compile(r"@(sha256:[0-9a-f]{64})$")
-_HARNESS = re.compile(r"^agentx-harness==(\S+)", re.MULTILINE)
+_AIPERF = re.compile(r"^aiperf @ \S+/archive/([0-9a-f]{40})\.tar\.gz$", re.MULTILINE)
 _RESULT_GPUS = re.compile(r"_gpus_([0-9]+)")
 
 
@@ -71,6 +71,7 @@ class Versions(_Record):
     srt_slurm_sha: str | None = None
     srtctl_version: str | None = None
     agentx_harness: str | None = None
+    aiperf_sha: str | None = None
     image_digest: str | None = None
 
 
@@ -513,15 +514,15 @@ def _versions(env: Mapping[str, str], *, agentx: bool) -> Versions:
         head = _git_head(repository_root())
     workflow = _text(env, "GITHUB_SHA")
     pinned = _PINNED.search(env.get("IMAGE", ""))
-    harness = None
+    aiperf = None
     if agentx:
         with contextlib.suppress(OSError):
-            match = _HARNESS.search(REQUIREMENTS.read_text())
-            harness = match.group(1) if match else None
+            match = _AIPERF.search(REQUIREMENTS.read_text())
+            aiperf = match.group(1) if match else None
     return Versions(
         inferencex_sha=head,
         workflow_sha=workflow if workflow != head else None,
-        agentx_harness=harness,
+        aiperf_sha=aiperf,
         image_digest=pinned.group(1) if pinned else None,
     )
 

@@ -108,7 +108,7 @@ def _point(tmp_path: Path, **overrides: str | None) -> dict[str, str]:
 def _runtime(tmp_path: Path) -> Runtime:
     runtime = Runtime(tmp_path / "runtime")
     runtime.python.parent.mkdir(parents=True)
-    stubs = {runtime.python: FAKE_PYTHON, runtime.agentx: FAKE_AIPERF, runtime.hf: FAKE_HF}
+    stubs = {runtime.python: FAKE_PYTHON, runtime.aiperf: FAKE_AIPERF, runtime.hf: FAKE_HF}
     for path, text in stubs.items():
         executable(path, text)
     return runtime
@@ -350,6 +350,7 @@ def test_signal_during_the_replay_skips_scoring_and_exits_128_plus_n(tmp_path):
 out = pathlib.Path(argv[argv.index("--output-artifact-dir") + 1])
 out.mkdir(parents=True, exist_ok=True)
 pathlib.Path(os.environ["REPLAY_ARGV"]).write_text(shlex.join([sys.argv[0], *argv]))
+print("fake aiperf: configure timeout", os.environ.get("AIPERF_DATASET_CONFIGURATION_TIMEOUT"))
 with open(os.environ["EVENTS"], "a") as events:
     events.write("replay\n")
 record = {
@@ -379,10 +380,10 @@ def test_signal_during_the_replay_skips_scoring_and_exits_128_plus_n(tmp_path):
     venv) for target; do :; done; exec "$REAL_PYTHON" -m venv --without-pip "$target" ;;
     pip)
         bin="$(dirname "$4")"
-        printf '#!%s\n' "$bin/python" > "$bin/agentx"
-        cat "$FAKE_AIPERF_SCRIPT" >> "$bin/agentx"
+        printf '#!%s\n' "$bin/python" > "$bin/aiperf"
+        cat "$FAKE_AIPERF_SCRIPT" >> "$bin/aiperf"
         printf '#!/bin/sh\necho "hf $*" >> "$EVENTS"\n' > "$bin/hf"
-        chmod +x "$bin/agentx" "$bin/hf" ;;
+        chmod +x "$bin/aiperf" "$bin/hf" ;;
 esac
 """
 
@@ -433,8 +434,9 @@ def frontend(method: str, path: str) -> tuple[int, object]:
     results = tmp_path / "results"
     command = (results / "benchmark_command.txt").read_text()
     assert command == (tmp_path / "replay-argv").read_text() + "\n"
-    assert shlex.split(command)[0] == str(tmp_path / "runtime/venv/bin/agentx")
+    assert shlex.split(command)[0] == str(tmp_path / "runtime/venv/bin/aiperf")
     assert "fake aiperf: profiled 1 request" in (results / "benchmark.log").read_text()
+    assert "fake aiperf: configure timeout 1800" in result.stdout
     assert "fake aiperf: profiled 1 request" in result.stdout
     aggregate = json.loads((tmp_path / "agentx.json").read_text())
     assert (aggregate["conc"], aggregate["num_requests_total"]) == (8, 1)