Abench

English | 简体中文

Agentic inference benchmark: closed-loop replay of recorded agent trajectories (DeepSWE v1.1, LMCache) — or a locally generated workload — against any OpenAI-compatible inference server (SGLang, vLLM, ...).

It answers: how fast does a single agent advance? — per-turn TTFT / TPOT / E2E, cache behavior, and throughput at a fixed concurrency, on a fixed, reproducible agent workload.

Install

pip install agentic-bench          # from PyPI (after the first release)

Or install the wheel attached to a GitHub Release:

pip install https://github.com/AlibabaResearch/abench/releases/download/v0.2.0/agentic_bench-0.2.0-py3-none-any.whl

From a checkout:

pip install .

The PyPI distribution is named agentic-bench; the Python module and the abench command keep the project name. An unrelated PyPI package named abench (ActivitySim) also installs an abench module and command, so do not install both in the same environment. If you already have it, run pip uninstall abench first.

Quick start

Two subcommands.

# 1) one-time data prep: downloads deepswe/mixed (~42 MB, SHA-256
#    verified, resumable) from the ABench ModelScope mirror; data lands
#    under ~/abench_data/deepswe/mixed/
python3 -m abench data deepswe/mixed

# 2) replay at one concurrency; every session runs to completion
#    (or under --duration) and full metrics are reported. --output-file
#    appends one JSON line per run; per-turn records are included by
#    default (--no-output-details keeps lines slim).
python3 -m abench run \
    --model default \
    --dataset-name deepswe/mixed --max-concurrency 16 \
    --output-file results.jsonl
# no --host/--port needed: localhost :30000 (sglang) / :8000 (vllm) is
# auto-detected; pass --port/--base-url only for non-default deployments

Smoke test with no download (locally generated random workload):

python3 -m abench run \
    --dataset-name random \
    --turns 3 --output-len 64 --max-concurrency 4

python3 -m abench ... is the primary invocation (matching how bench_serving is called). The wheel also installs an equivalent abench console script. Results always land on disk: without --output-file the run writes {backend}_{date}_{dataset}.jsonl in the working directory.

Backends & auth

Any OpenAI-compatible server works; --backend auto|sglang|vllm|openai (default auto, probed from the server) adapts the observation layer — SGLang-only extras (per-layer cache breakdown, /flush_cache, /metrics gauge names) are gated per backend, and unsupported metrics report null rather than 0. Request shape is standard OpenAI either way.

Leave --backend and --port unset: auto probes the engine and resolves the port (sglang/openai 30000, vLLM 8000). A wrong explicit --backend does not error — it silently drops that backend's metrics (cache breakdown / KV pool / server_info); abench warns when a declared backend does not match the server.

• Auth: export OPENAI_API_KEY=... (or API_KEY= for custom gateway schemes) — attached to every request, no CLI flag needed.
• Escape hatches: --header Key=Value (repeatable) for custom HTTP headers; --extra-request-body '{"seed": 123}' merges into every request body and can override built-in fields.
• --ready-check-timeout-sec polls /v1/models before the run so a still-loading server fails fast instead of timing out every request.

CLI

command	purpose
python3 -m abench data <DATASET>/<TIER>	the single data-prep entry: download a preset dataset tier (SHA-256 verified, resumable) from the ABench ModelScope mirror to <data-path>/<dataset>/<tier>/ (--list shows all; a bare dataset name without a tier is rejected)
python3 -m abench run	the single bench entry: closed-loop replay at one concurrency; data located by --dataset-name (identity, looked up under ~/abench_data) and/or --dataset-path (file / tier dir / root)

--help shows the core flags; --help-advanced shows everything. The tokenizer resolves automatically (--tokenizer, else a local --model dir, else the server's /model_info, else AutoTokenizer.from_pretrained for non-local names, else tiktoken, else chars/4) — the Hub fetch matches bench_serving; set HF_HUB_OFFLINE=1 on air-gapped machines.

Docs

• docs/README.md — reference: closed-loop protocol, dataset tiers, every flag, metric definitions
• docs/DESIGN.md — design rationale, measurement decisions
• CONTRIBUTING.md — how to contribute, CLA, and review expectations
• CODE_OF_CONDUCT.md

Development

src/abench/
  cli.py          argument parsing and dispatch (python3 -m abench)
  runner.py       what `data` and `run` actually do
  core/           domain: session data, workload planning, result records
  serve/          the server under test: endpoints, telemetry, report
  replay/         closed-loop engine and turn transport
docs/             reference and design notes

pip install -e .
python -m build --wheel   # ships the package only

License

Copyright (c) 2026, Alibaba Group and its affiliates. Abench is licensed under the Apache License, Version 2.0 — see LICENSE. The upstream conventions this CLI follows, and where the datasets come from, are recorded in NOTICE; the repository itself ships no trajectory data.