vllm-bench-serve（vllm 替补发球）

vLLM Bench Serve

Benchmark vLLM or any OpenAI-compatible serving endpoint using the vllm bench serve CLI. Measures throughput, latency (TTFT, TPOT), and goodput against configurable request load.

Reference: vLLM Bench Serve Documentation

Prerequisites

• vLLM installed (or any OpenAI-compatible server running)
• A vLLM server or API endpoint already serving a model
• Python environment with vLLM for the benchmark client

Quick Start

Basic benchmark against local vLLM server (default random dataset, 1000 prompts):

vllm bench serve \
  --backend openai-chat \
  --host 127.0.0.1 \
  --port 8000 \
  --model Qwen/Qwen2.5-1.5B-Instruct \
  --endpoint /v1/chat/completions

Save results to JSON:

vllm bench serve \
  --backend openai-chat \
  --host 127.0.0.1 \
  --port 8000 \
  --model Qwen/Qwen2.5-1.5B-Instruct \
  --endpoint /v1/chat/completions \
  --save-result \
  --result-dir ./bench-results \
  --metadata "version=0.6.0" "tp=1"

Note: When using --backend openai-chat, you must specify --endpoint /v1/chat/completions (default is /v1/completions).

Core Arguments

Datasets

Dataset-specific options (examples):

# Random: control input/output length
--dataset-name random --random-input-len 1024 --random-output-len 128

# Sonnet defaults: input 550, output 150, prefix 200
--dataset-name sonnet --sonnet-input-len 550 --sonnet-output-len 150

# HuggingFace dataset
--dataset-name hf --dataset-path "lmarena-ai/VisionArena-Chat" --hf-split test

# General overrides (map to dataset-specific args)
--input-len 512 --output-len 256

Load Control

# Fixed request rate (Poisson process)
--request-rate 10

# More bursty arrivals (gamma distribution, burstiness < 1)
--request-rate 10 --burstiness 0.5

# Ramp-up from low to high RPS
--ramp-up-strategy linear --ramp-up-start-rps 1 --ramp-up-end-rps 50

# Limit concurrency (useful for rate-limited APIs)
--max-concurrency 32

Results and Metrics

Sampling Parameters (OpenAI-compatible backends)

--temperature 0.7 --top-p 0.95 --top-k 50
--frequency-penalty 0 --presence-penalty 0 --repetition-penalty 1.0

Common Workflows

1. Throughput test with random dataset (burst):

vllm bench serve --backend openai-chat --host 127.0.0.1 --port 8000 \
  --model Qwen/Qwen2.5-1.5B-Instruct \
  --endpoint /v1/chat/completions \
  --dataset-name random \
  --num-prompts 500 --random-input-len 512 --random-output-len 128

2. Latency test with fixed QPS:

vllm bench serve --backend openai-chat --host 127.0.0.1 --port 8000 \
  --model Qwen/Qwen2.5-1.5B-Instruct \
  --endpoint /v1/chat/completions \
  --request-rate 5 --num-prompts 200 \
  --save-result --percentile-metrics ttft,tpot --metric-percentiles 50,99

3. Benchmark against remote API (base-url):

vllm bench serve --backend openai-chat \
  --base-url "https://api.example.com/v1" \
  --model my-model \
  --header "Authorization=Bearer $API_KEY"

4. Run inside Docker (when vLLM client not on host):

docker exec <container-name> vllm bench serve \
  --backend openai-chat --host 127.0.0.1 --port 8000 \
  --model Qwen/Qwen2.5-1.5B-Instruct \
  --endpoint /v1/chat/completions \
  --dataset-name random --num-prompts 100

Troubleshooting

• Connection refused: Ensure the server is running and --host/--port or --base-url are correct.
• Model not found: Pass --model explicitly or ensure /v1/models returns the model.
• URL must end with chat/completions: Use --endpoint /v1/chat/completions when --backend openai-chat.
• Rate limit / 429: Reduce --request-rate or --max-concurrency.
• Ready check: Use --ready-check-timeout-sec 60 to wait for the endpoint before benchmarking.
• SSL: Use --insecure for self-signed certificates.

Notes

• For embeddings/rerank benchmarks, use --backend openai-embeddings, vllm-pooling, or vllm-rerank.
• --profile requires --profiler-config on the server for vLLM profiling.
• Goodput SLOs are useful for SLA-style analysis; see DistServe paper for details.