| --- |
| license: apache-2.0 |
| task_categories: |
| - tabular-regression |
| language: |
| - en |
| tags: |
| - llm-inference |
| - benchmarking |
| - gpu-profiling |
| - vllm |
| - sglang |
| - agentic-workloads |
| size_categories: |
| - 1K<n<10K |
| pretty_name: AgentPerfBench |
| version: "1.0" |
| configs: |
| - config_name: trace_replay |
| data_files: |
| - split: summary |
| path: trace_replay/summary.parquet |
| - config_name: distributional |
| data_files: |
| - split: summary |
| path: distributional/summary.parquet |
| dataset_info: |
| - config_name: trace_replay |
| features: |
| - name: run_id |
| dtype: string |
| - name: model |
| dtype: string |
| - name: model_family |
| dtype: string |
| - name: hardware |
| dtype: string |
| - name: engine |
| dtype: string |
| - name: tensor_parallelism |
| dtype: int64 |
| - name: profile |
| dtype: string |
| - name: concurrency |
| dtype: int64 |
| - name: num_requests |
| dtype: int64 |
| - name: duration_s |
| dtype: float64 |
| - name: successful_requests |
| dtype: int64 |
| - name: failed_requests |
| dtype: int64 |
| - name: request_throughput |
| dtype: float64 |
| - name: input_token_throughput |
| dtype: float64 |
| - name: output_token_throughput |
| dtype: float64 |
| - name: total_token_throughput |
| dtype: float64 |
| - name: mean_ttft_ms |
| dtype: float64 |
| - name: median_ttft_ms |
| dtype: float64 |
| - name: p90_ttft_ms |
| dtype: float64 |
| - name: p99_ttft_ms |
| dtype: float64 |
| - name: mean_tpot_ms |
| dtype: float64 |
| - name: median_tpot_ms |
| dtype: float64 |
| - name: p90_tpot_ms |
| dtype: float64 |
| - name: p99_tpot_ms |
| dtype: float64 |
| - name: mean_itl_ms |
| dtype: float64 |
| - name: median_itl_ms |
| dtype: float64 |
| - name: p90_itl_ms |
| dtype: float64 |
| - name: p99_itl_ms |
| dtype: float64 |
| - name: mean_e2el_ms |
| dtype: float64 |
| - name: median_e2el_ms |
| dtype: float64 |
| - name: p90_e2el_ms |
| dtype: float64 |
| - name: p99_e2el_ms |
| dtype: float64 |
| splits: |
| - name: summary |
| num_rows: 3147 |
| num_bytes: 694254 |
| - config_name: distributional |
| features: |
| - name: run_id |
| dtype: string |
| - name: model |
| dtype: string |
| - name: model_family |
| dtype: string |
| - name: hardware |
| dtype: string |
| - name: engine |
| dtype: string |
| - name: tensor_parallelism |
| dtype: int64 |
| - name: profile |
| dtype: string |
| - name: concurrency |
| dtype: int64 |
| - name: num_requests |
| dtype: int64 |
| - name: duration_s |
| dtype: float64 |
| - name: successful_requests |
| dtype: int64 |
| - name: failed_requests |
| dtype: int64 |
| - name: request_throughput |
| dtype: float64 |
| - name: input_token_throughput |
| dtype: float64 |
| - name: output_token_throughput |
| dtype: float64 |
| - name: total_token_throughput |
| dtype: float64 |
| - name: mean_ttft_ms |
| dtype: float64 |
| - name: median_ttft_ms |
| dtype: float64 |
| - name: p90_ttft_ms |
| dtype: float64 |
| - name: p99_ttft_ms |
| dtype: float64 |
| - name: mean_tpot_ms |
| dtype: float64 |
| - name: median_tpot_ms |
| dtype: float64 |
| - name: p90_tpot_ms |
| dtype: float64 |
| - name: p99_tpot_ms |
| dtype: float64 |
| - name: mean_itl_ms |
| dtype: float64 |
| - name: median_itl_ms |
| dtype: float64 |
| - name: p90_itl_ms |
| dtype: float64 |
| - name: p99_itl_ms |
| dtype: float64 |
| - name: mean_e2el_ms |
| dtype: float64 |
| - name: median_e2el_ms |
| dtype: float64 |
| - name: p90_e2el_ms |
| dtype: float64 |
| - name: p99_e2el_ms |
| dtype: float64 |
| splits: |
| - name: summary |
| num_rows: 245 |
| num_bytes: 70836 |
| --- |
| |
| # AgentPerfBench |
|
|
| LLM inference benchmark: 3,392 runs measuring TTFT, TPOT, ITL, and throughput across 9 models, up to 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights. |
|
|
| ## Dataset configurations |
|
|
| The dataset provides two configurations. *trace_replay* replays exact input/output sequences from recorded agent sessions. *distributional* samples from statistical distributions fitted to those same workloads, trading fidelity for faster sweeps across the hardware matrix. |
|
|
| ### trace_replay (3,147 rows) |
| |
| Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). Covers 77 unique (model, hardware, engine) combinations across 17 profiles and 6 concurrency levels. The full 5-dimensional matrix is 12.2% filled; not all models run on all hardware. |
| |
| 17 profiles: `chat-medium`, `chat-multiturn-long`, `chat-multiturn-medium`, `chat-multiturn-short`, `chat-short`, `chat-singleturn`, `coding-singleturn`, `decode-heavy`, `osworld-multiturn-long`, `osworld-multiturn-medium`, `osworld-multiturn-short`, `prefill-heavy`, `random-1k`, `swebench-multiturn-medium`, `swebench-multiturn-short`, `terminalbench-multiturn-medium`, `terminalbench-multiturn-short` |
| |
| ### distributional (245 rows) |
| |
| Samples ISL/OSL from lognormal distributions fitted to real workload statistics. Covers 42 unique (model, hardware, engine) combinations across 6 profiles and 7 concurrency levels (3.0% matrix fill). `gpt-oss-120b`, `3090x8`, and `A100-40GBx8` are excluded from this configuration. |
| |
| 6 profiles: `chat-multiturn`, `chat-singleturn`, `coding-singleturn`, `osworld-multiturn`, `swebench-multiturn`, `terminalbench-multiturn` |
| |
| ### Concurrency filtering |
| |
| The benchmark harness capped actual concurrent connections at the session pool size. Rows where declared concurrency exceeded the pool were excluded: |
| |
| - trace_replay: concurrency > 100 removed (session pool was 100). Remaining values: {1, 5, 10, 20, 40, 80}. |
| - distributional (pre-fix): concurrency > 10 removed (session pool was 10). Post-fix data has no cap. Remaining values: {1, 5, 10, 40, 80, 200, 320}. |
|
|
| | Config | Rows | |
| |--------|------| |
| | trace_replay | 3,147 | |
| | distributional | 245 | |
| | **Total** | **3,392** | |
| |
| ### Failed requests |
| |
| Some runs produce request failures, typically at high concurrency where the engine hits memory or timeout limits. 30.8% of trace_replay rows and 42% of distributional rows have `failed_requests > 0`. Summary metrics (TTFT, TPOT, throughput) are computed from successful requests only. |
|
|
| ## Coverage |
|
|
| ### Hardware |
|
|
| All benchmarks collected on PyTorch 2.10.0, CUDA 12.8. |
|
|
| | GPU | VRAM | HBM bandwidth | Peak half-precision TFLOPS | |
| |-----|------|---------------|---------------------------| |
| | NVIDIA H100 SXM | 80 GB | 3.35 TB/s | 989 | |
| | NVIDIA A100 SXM4 | 40 GB | 1.56 TB/s | 312 | |
| | NVIDIA RTX 3090 | 24 GB | 936 GB/s | 71 | |
| | NVIDIA RTX 2080 Ti | 11 GB | 616 GB/s | 27 | |
|
|
| Multi-GPU configurations: 1, 2, 4, or 8 GPUs with tensor parallelism. TP degree depends on model size and available GPUs. |
|
|
| ### Models |
|
|
| All models served in BF16 unless noted. |
|
|
| | Model | Family | Parameters | Architecture | Notes | |
| |-------|--------|-----------|--------------|-------| |
| | Llama-3.1-8B | Llama | 8B | Dense | | |
| | Llama-3.1-70B | Llama | 70B | Dense | | |
| | Llama-3.3-70B | Llama | 70B | Dense | | |
| | Qwen2.5-72B | Qwen | 72B | Dense | | |
| | Qwen3.5-9B | Qwen | 9B | Dense | | |
| | Qwen3.5-27B | Qwen | 27B | Dense | | |
| | Mixtral-8x7B | Mixtral | 46.7B (12.9B active) | MoE | | |
| | gpt-oss-20b | GPT-OSS | 21B (3.6B active) | MoE | mxfp4 projections | |
| | gpt-oss-120b | GPT-OSS | 117B (5.1B active) | MoE | mxfp4 projections | |
|
|
| Model names in this table match the `model` column in the parquet files. |
|
|
| ### Engines |
|
|
| - vLLM 0.19.0 |
| - SGLang 0.5.9 |
|
|
| ## Schema |
|
|
| Each row in `summary.parquet` (both configs): |
|
|
| | Column | Type | Description | |
| |--------|------|-------------| |
| | run_id | string | Deterministic hash of run parameters | |
| | model | string | Model short name | |
| | model_family | string | Model family (llama, qwen, gpt-oss, mixtral) | |
| | hardware | string | GPU configuration (e.g., H100x4) | |
| | engine | string | Serving engine (vllm, sglang) | |
| | tensor_parallelism | int | TP degree | |
| | profile | string | Workload profile name | |
| | concurrency | int | Concurrent request level | |
| | num_requests | int | Total requests in run | |
| | duration_s | float | Total run duration | |
| | successful_requests | int | Completed requests | |
| | failed_requests | int | Failed requests | |
| | request_throughput | float | Requests/second | |
| | input_token_throughput | float | Input tokens/second | |
| | output_token_throughput | float | Output tokens/second | |
| | total_token_throughput | float | Total tokens/second | |
| | mean/median/p90/p99_ttft_ms | float | Time to first token | |
| | mean/median/p90/p99_tpot_ms | float | Time per output token | |
| | mean/median/p90/p99_itl_ms | float | Inter-token latency | |
| | mean/median/p90/p99_e2el_ms | float | End-to-end latency | |
|
|
| ## Loading |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay") |
| # or "distributional" |
| ``` |
|
|
| ## Benchmark methodology |
|
|
| - Closed-loop concurrency with semaphore control. |
| - Concurrency levels: {1, 5, 10, 20, 40, 80} (trace_replay), {1, 5, 10, 40, 80, 200, 320} (distributional). |
| - 3-request warmup before each configuration. |
| - Metrics: TTFT, TPOT, ITL, E2EL, request throughput, token throughput. |
| - Summary statistics: mean, median, p90, p99. |
| - Collection period: March 2026 onwards. |
| - PyTorch 2.10.0, CUDA 12.8 on all machines. All models served in BF16 (gpt-oss: mxfp4 projection weights). |
| |
| ## Future releases |
| |
| - Per-request and multi-turn granularity data (pending raw JSON availability from collection infrastructure). |
| - Per-kernel CUDA roofline profiles (PyTorch profiler, 2-layer forward passes, batch sizes 1/4/8/32/64). |
| - This is version 1.0. Updates will be tagged with semantic versions. |
| |
| ## Intended uses |
| |
| - Inference engine comparison under controlled conditions. |
| - Capacity planning for LLM deployments. |
| - TTFT scaling with context length in multi-turn sessions. |
| |
| ## Limitations |
| |
| - Results are specific to tested hardware and software versions (vLLM 0.19.0, SGLang 0.5.9, PyTorch 2.10.0, CUDA 12.8). |
| - Distributional profiles approximate but do not replicate production traffic patterns. |
| - No consumer GPUs beyond RTX 3090; no non-NVIDIA accelerators. |
| - Closed-loop concurrency only; no open-loop (Poisson) arrivals. |
| - The model-hardware-concurrency matrix is sparse (12.2% fill for trace_replay, 3.0% for distributional). Not all model-hardware combinations are represented. |
| - No model quality metrics. This is a systems benchmark. |
|
|
| ## Ethical considerations |
|
|
| No PII. Trace-replay profiles derive from open benchmarks (SWE-Bench MIT, TerminalBench, OSWorld). Synthetic profiles use random tokens. |
|
|
| ## License |
|
|
| Benchmark data released under Apache-2.0. Source datasets retain their original licenses. |
|
|
| ## Source datasets |
|
|
| - [SWE-Bench](https://github.com/princeton-nlp/SWE-bench) (MIT) |
| - [TerminalBench](https://github.com/TerminalBench/TerminalBench) |
| - [ShareGPT (Aeala/ShareGPT_Vicuna_unfiltered)](https://huggingface.co/datasets/Aeala/ShareGPT_Vicuna_unfiltered) |
| - [OSWorld](https://github.com/xlang-ai/OSWorld) |
|
|
| ## Citation |
|
|
| ```bibtex |
| @inproceedings{agentperfbench2026, |
| title={AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance of Agentic LLMs}, |
| author={Anonymous}, |
| booktitle={NeurIPS 2026 Evaluations and Datasets Track}, |
| year={2026} |
| } |
| ``` |
|
|
