Aiperf

AIPerf — NVIDIA generative-AI inference benchmarking

Target audience: operators producing defensible latency/throughput/goodput numbers against any OpenAI-compatible inference server (vLLM, SGLang, TensorRT-LLM, NVIDIA Dynamo, NIM, Triton, HF TGI, Ollama), and developers extending AIPerf with custom endpoints, datasets, exporters, or metrics.

Why this matters

aiperf is the open-source successor to genai-perf, written by NVIDIA's AI-Dynamo team. It is the vendor-neutral way to:

1. Replay production traces (Mooncake / Bailian / BurstGPT) at exact timestamps — synthetic load lies about cache reuse and tail behavior.
2. Measure goodput, not just throughput — the percentage of requests that meet all SLOs simultaneously. A system at 1000 req/s throughput and 28% goodput is mis-provisioned by ~3.5×.
3. Account for reasoning tokens correctly. GPT-OSS / DeepSeek-R1 / Qwen3 emit reasoning_content before the answer. genai-perf ignored those; aiperf splits TTFT (any first token) from TTFO (first non-reasoning token). Numbers are not directly comparable across the two tools — see migration notes.
4. Collect server + GPU telemetry alongside client-side timings in one run — DCGM / pynvml GPU metrics + Prometheus /metrics scrape, all aligned in the same artifact dir.
5. Extend cleanly — 25 plugin categories with a YAML manifest + entry-point system. New endpoints, dataset formats, exporters, accuracy graders, plot types all go through the same registry.

If the target is exclusively vLLM and the operator wants the in-tree vllm bench toolchain (sweep, latency-only, startup, mm-processor), defer to the vllm-benchmarking skill. AIPerf is the right answer when (a) the target is non-vLLM or (b) the workload is trace-driven, multi-modal, multi-server, or needs goodput.

What landed in v0.9.0 – v0.11.0

Three minors shipped between 2026-05-30 and 2026-07-08. The ones that change how you drive AIPerf:

• Adaptive sweep orchestrator + YAML-native v2 config (v0.10.0, PR #912) — Bayesian Optimization and reusable search recipes, instead of hand-rolled concurrency sweeps.
• Multi-tier SLO search (v0.11.0, PR #1035) — resolve N tier boundaries in a single job rather than one job per tier.
• Accuracy suite went from a curiosity to a platform — AIME (v0.9.0), then HellaSwag, BigBench-Hard, AIME 2024/2025 (v0.10.0, DeepEval/lighteval-backed), then MATH-500, GPQA-Diamond, LCB CodeGen (with a code_execution grader) and GSM8K (v0.11.0).
• Exporters: live OpenTelemetry streaming + MLflow (v0.9.0, PR #900); Weights & Biases (v0.11.0, PR #1049).
• Telemetry beyond NVIDIA: AMD ROCm collector via amdsmi (v0.9.0, PR #908).
• network_adjusted_* latency metrics (v0.11.0, PR #1066) — subtract network RTT so numbers reflect the server, not the path to it.
• Power metrics — initial implementation (v0.10.0, PR #803).
• Warmup fix (v0.11.0) preventing prefix-cache reuse from skewing results — if you have baselines from ≤v0.10.x with warmup enabled, they may not be comparable.

Versions

• Stable on PyPI: v0.11.0 (2026-07-08), requires-python >=3.10,<3.14. pip install aiperf. A nightly wheel is published alongside as aiperf-nightly (v0.10.0+).
• Repo main at https://github.com/ai-dynamo/aiperf: post-v0.11.0.
• Python: ≥3.10, <3.14. Uses uvloop on Linux/macOS, falls back to default asyncio on Windows — Windows is a first-class port with blocking CI since v0.11.0 (PR #1007).
• Source of truth for flags: aiperf profile --help. CLI options doc is auto-generated via make generate-cli-docs. If the doc disagrees with --help on a flag spelling, trust --help.

Decision tree — which subcommand

The four scheduling modes — pick exactly one

aiperf profile schedules requests in one of four mutually exclusive modes. Picking the wrong one is the single most common source of misleading numbers. See references/timing-modes.md for the full compatibility matrix and validation errors.

A stop condition is required: --request-count, --num-sessions, or --benchmark-duration. Duration mode also reads --benchmark-grace-period (default 30s; user-centric defaults to ∞).

Two methodologies operators actually need

A — Health check / SLO validation

Question: Does this deployment meet the SLO under realistic load?

aiperf profile \
  --model Qwen/Qwen3-0.6B \
  --url http://endpoint:8000 \
  --endpoint-type chat \
  --streaming \
  --tokenizer Qwen/Qwen3-0.6B \
  --input-file prod-trace.jsonl --custom-dataset-type mooncake_trace --fixed-schedule \
  --goodput "time_to_first_token:250 inter_token_latency:10 request_latency:2000" \
  --warmup-duration 30 \
  --benchmark-duration 600 --benchmark-grace-period 60 \
  --artifact-dir artifacts/health-$(date -Iseconds)

The operator reports goodput-req/s and the offending tail percentiles. Goodput=0 with high throughput means the SLO is failing across the board.

B — A/B comparison / before-after

Question: Did config change X help, and at what cost?

Run a concurrency or request-rate sweep, not a single point. The knee of the throughput-vs-tail-latency curve is the usable operating point. A change that moves the knee right is a win.

for c in 10 50 100 200 500; do
  aiperf profile \
    --model Qwen/Qwen3-0.6B --url http://endpoint:8000 \
    --endpoint-type chat --streaming \
    --tokenizer Qwen/Qwen3-0.6B \
    --concurrency $c --request-count 1000 \
    --isl 1000 --osl 500 \
    --random-seed 42 \
    --artifact-dir artifacts/pareto-c$c
done
aiperf plot --paths artifacts/pareto-c* --dashboard

Same seed, same dataset, back-to-back runs, same hardware. Multi-run confidence (--num-profile-runs 3 --confidence-level 0.95) yields CI bands when each run is short enough that re-running 3× is feasible.

Critical pitfalls

1. Tokenizer mismatch. Default tokenizer is whatever HF resolves from --model. If the served model and the HF ID differ (custom path, fine-tune, gated repo), every token-count metric is fiction. Always pass --tokenizer <hf-id> explicitly. For zero-network use --tokenizer builtin (tiktoken o200k_base, GPT-4o-class). Set --tokenizer-trust-remote-code for tokenizers that ship custom Python (Kimi K2.5, some DeepSeek). Pin via --tokenizer-revision <sha> to avoid silent drift on model-card updates.

2. Reasoning-model TTFT/TTFO/OSL semantics. AIPerf's TTFT counts reasoning tokens (reasoning_content SSE field); genai-perf's TTFT did not. To compare apples-to-apples to historical genai-perf numbers, read TTFO (time-to-first-output-token) instead. Same trap on OSL: aiperf OSL = output + reasoning, aiperf output_token_count = output only. Prerequisite: the server must surface reasoning_content as a separate SSE field (vLLM --reasoning-parser deepseek_r1 / qwen3 / gpt_oss etc.; SGLang --reasoning-parser deepseek-r1) — without it, aiperf can't separate streams and TTFO collapses to TTFT, reasoning_token_count stays 0. Verify with curl ... | jq '.choices[0].delta.reasoning_content' before trusting numbers. See references/migration-from-genai-perf.md.

3. --request-rate alone, no concurrency cap. The benchmark will queue indefinitely on slow servers and the timing manager will fall behind, producing a curve that looks like a system collapse but is really client-side queue buildup. Always pair with a sane --concurrency ceiling (e.g. 2–3× expected steady-state inflight).

4. --fixed-schedule with synthetic data. Fixed-schedule needs timestamped traces (mooncake / bailian / burst_gpt). It is auto-enabled for trace dataset types. Combining it with --request-rate or --user-centric-rate raises a validation error.

5. --isl-block-size and the Mooncake/server block-size relationship — read this carefully. Mooncake traces encode hash_ids at 512-token granularity by design — the trace was generated with that block size and the IDs only make sense at 512. Do not lower --isl-block-size to "match the server" (e.g. 64 for SGLang --page-size 64); doing so makes aiperf reconstruct prompts ~8× shorter than the trace intends and corrupts the ISL distribution. What matters is whether 512 is a multiple of the server's block size — if yes (e.g. SGLang 64 → 512/64=8, vLLM 16 → 512/16=32), each Mooncake block aligns with N server pages and prefix-cache reuse works fine. If the server block size doesn't divide 512, that's the broken case. Run aiperf analyze-trace --input-file trace.jsonl --block-size 512 to confirm what the trace expects.

6. Skipping warmup on cold pods. First N requests hit cold CUDA graphs / torch.compile caches / un-pinned KV pages. Warmup is opt-in: --warmup-duration 30 or --warmup-request-count 100. Multi-run confidence has --profile-run-disable-warmup-after-first to amortize.

7. Ignoring --use-server-token-count for OpenAI-compat servers. When set, AIPerf reads token counts from the server's usage field instead of re-tokenizing client-side. Auto-enables stream_options.include_usage for streaming. This avoids client/server tokenizer drift entirely. Use this whenever the server is trustworthy. The --tokenizer flag is still required (for input-shape generation) but tokenizer.encode() is not called for metrics.

8. --export-level raw on a long run. Writes every request/response payload to disk. Fine for 1k requests, catastrophic for 1M. Default is records (per-request metrics, no payloads); only escalate when actually debugging.

9. CSV export prefix conflict. --profile-export-prefix /abs/path/run1 was broken pre-v0.7.0 (PR #801 fixed it). On v0.7.0+ this works; on older builds use a relative prefix and --artifact-dir.

10. Async video generation polling. endpoint-type video_generation polls /v1/videos/{job_id} until the job completes. Tune AIPERF_HTTP_VIDEO_POLL_INTERVAL (default 0.1 s) — faster polling burns server resources, slower polling inflates request latency. With --download-video-content, request latency includes the bytes download.

11. --goodput is per-request, not per-percentile. The flag asks "is this individual request below the threshold?" and reports good_request_count / benchmark_duration as goodput. There is no built-in p99_time_to_first_token:400 syntax. To express "p99 TTFT must be ≤ 400 ms", set --goodput "time_to_first_token:400" and check goodput / request_throughput ≥ 0.99 post-run — that ratio is the fraction of requests meeting all SLOs, which for a single threshold is exactly the "X% under the budget" measure operators usually want. Combine multiple thresholds and the same ratio still works as joint compliance.

What to read next — and when

The upstream repo at https://github.com/ai-dynamo/aiperf is the most authoritative reference — docs/cli-options.md, docs/metrics-reference.md, docs/environment-variables.md, and docs/tutorials/*.md are all auto-generated or hand-curated from the same code. When this skill disagrees with the repo, trust the repo (and update this skill).

Quick recipes

Smoke test a vLLM endpoint

aiperf profile --model Qwen/Qwen3-0.6B --url http://localhost:8000 \
  --endpoint-type chat --streaming --tokenizer Qwen/Qwen3-0.6B \
  --concurrency 10 --request-count 100 --isl 1000 --osl 500

ShareGPT against an OpenAI-compatible server

aiperf profile --model my-model --url http://endpoint:8000 \
  --endpoint-type chat --streaming --tokenizer my-tokenizer \
  --public-dataset sharegpt --num-sessions 200 --concurrency 50

Mooncake trace replay with goodput

curl -O https://raw.githubusercontent.com/kvcache-ai/Mooncake/refs/heads/main/FAST25-release/arxiv-trace/mooncake_trace.jsonl
aiperf profile --model my-model --url http://endpoint:8000 \
  --endpoint-type chat --streaming --tokenizer my-tokenizer \
  --input-file mooncake_trace.jsonl --custom-dataset-type mooncake_trace --fixed-schedule \
  --goodput "time_to_first_token:400 request_latency:2000"

NIM embeddings benchmark

aiperf profile --model nvidia/nv-embedqa-e5-v5 --url http://nim:8000 \
  --endpoint-type nim_embeddings --tokenizer intfloat/e5-large-v2 \
  --concurrency 32 --request-count 5000 --prompt-batch-size 16

Custom multi-turn JSONL with shared system prompt

aiperf profile --model my-model --url http://endpoint:8000 \
  --endpoint-type chat --streaming --tokenizer my-tokenizer \
  --input-file conversations.jsonl --custom-dataset-type multi_turn \
  --shared-system-prompt-length 1000 --user-context-prompt-length 200 \
  --num-sessions 50 --concurrency 25

Goodput per concurrency point with multi-run CI

aiperf profile --model my-model --url http://endpoint:8000 \
  --endpoint-type chat --streaming --tokenizer my-tokenizer \
  --concurrency 100 --request-count 2000 --isl 1000 --osl 500 \
  --num-profile-runs 5 --confidence-level 0.95 \
  --profile-run-cooldown-seconds 30 \
  --goodput "time_to_first_token:300 inter_token_latency:8"

MMLU accuracy benchmark

aiperf profile --model my-model --url http://endpoint:8000 \
  --endpoint-type chat --streaming --tokenizer my-tokenizer \
  --accuracy-benchmark mmlu --accuracy-n-shots 5 --accuracy-grader multiple_choice

For the full flag catalogue, validation rules, and every endpoint/dataset/metric option, drill into the references/ files.