rohitjain28/vllm-benchmark-payloads

vLLM Benchmark Payloads

Synthetic OpenAI-format chat-completion payloads for latency / throughput benchmarking of a model served with vLLM's OpenAI-compatible server.

File

File	Records	Notes
payloads_1k_generic.jsonl	1,000	The benchmark dataset — one request body per line. Each has a ~12k-token system prompt + a dynamic per-record customer context.
sample_payload.json	1	One pretty-printed record, to inspect the format quickly.

All data is 100% synthetic (generated with Faker) using a fictional company; nothing real. Records are reproducible (fixed seed). Every payload sets benchmark-friendly flags: temperature: 0, ignore_eos: true (runs the full max_tokens for comparable latency), and stream: true (enables time-to-first-token). max_tokens varies across 128 / 256 / 512.

Each record is a generic market-standard customer-support system prompt (~12,000 tokens, shared across records — useful for prefix-cache testing) wrapped around a dynamic account block that changes every record: customer name, language (en-IN/hi-IN), plan, account/phone last-4, amount due, due date, status, city, and more.

1. Serve the model

pip install vllm
vllm serve <ORG>/<MODEL_NAME>     # starts on http://localhost:8000

Set "model" in the payloads to match the served name. Check GET /v1/models to confirm it.

2. Quick smoke test (first record)

head -n 1 payloads_1k_generic.jsonl \
  | curl http://localhost:8000/v1/chat/completions \
      -H "Content-Type: application/json" -d @-

3. Run all payloads and time each

while IFS= read -r line; do
  curl -s -o /dev/null -w 'latency: %{time_total}s\n' \
    http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" -d "$line"
done < payloads_1k_generic.jsonl

4. Full benchmark (recommended metrics)

vLLM's built-in benchmark reports latency, TTFT, TPOT, and throughput:

vllm bench serve \
  --model <ORG>/<MODEL_NAME> \
  --dataset-name random \
  --random-input-len 12000 \
  --random-output-len 256 \
  --num-prompts 200 \
  --request-rate 10