This guide explains how to run benchmarks on inference engines using Inference Engine Arena. There are three approaches to running benchmarks:
  1. Manual Mode: Run benchmarks on already-running engines
  2. Lifecycle Mode: Start engines, run benchmarks, and stop engines in a single command
  3. Batch Mode(YAML Configuration): Define complex benchmark scenarios in a YAML file

​
Approach 1: Manual Mode

In this approach, you run benchmarks on engines that are already running.

​
Prerequisites

Before running benchmarks in manual mode, you need:
  1. One or more inference engines running (see Starting Engines)

​
Basic Usage

The basic syntax for running benchmarks in manual mode is: 
arena run --engines <engine_list> --benchmark <benchmark_list>
Where:
  • <engine_list> is a space-separated list of engine types (e.g., vllm sglang)
  • <benchmark_list> is a space-separated list of benchmark types (e.g., conversational_short summarization)

​
Running a Simple Benchmark on a Single Engine

Here’s how to run a simple conversational benchmark on a running vLLM engine: 
arena run --engines vllm --benchmark conversational_short
This will:
  1. Connect to the running vLLM engine
  2. Run the conversational benchmark with default settings
  3. Collect metrics and display the results

​
Running Multiple Benchmarks on a Single Engine

You can run multiple benchmarks on a single engine: 
arena run --engines vllm --benchmark conversational_short summarization

​
Running Multiple Benchmarks on Multiple Engines

To compare multiple benchmarks on multiple engines: 
arena run --engines vllm sglang --benchmark conversational_short summarization conversational_medium

​
Approach 2: Lifecycle Mode in one command line

TO BE IMPLEMENTED In lifecycle mode, the tool manages the entire process: starting engines, running benchmarks, and stopping engines.

​
Basic Usage

The basic syntax for lifecycle mode is:

​
Approach 3: Batch Mode (YAML Configuration Mode)

For complex benchmark scenarios, you can define the entire configuration in a YAML file.

​
Running from YAML

To run the benchmarks defined in a YAML file: 
arena runyaml <path_to_yaml_file>
You may refer to the /example_yaml directory for more examples.

​
Example YAML Configuration

Use your imagination to define your own benchmark scenarios to share with friends. Here’s one example: 
runs:
  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 1"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10

  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 2"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10

  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 4"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10

  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 8"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10

  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 16"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10

  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 32"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10

  - engine:
      - type: vllm
        model: NousResearch/Meta-Llama-3.1-8B
        args: "--max-num-batched-tokens 163840 --max-num-seqs 256"
        env: {}
    benchmarks:
      - type: conversational_short
        dataset-name: random
        random-input-len: 100
        random-output-len: 100
        random-prefix-len: 0
        num-prompts: 50
        request-rate: 10
Inference Engine Arena provides a comprehensive set of benchmarks to evaluate the performance of different inference engines. Each benchmark is designed to test specific aspects of engine performance.

​
Preset Benchmarks

We find that the following benchmarks are useful. But feel free to define your own benchmarks.
Benchmark NameWorkload TypeInput LengthOutput LengthPrefix LengthQPSUse Case
summarizationLISO1200010002Long document summarization, Meeting notes
rewrite_essayLILO12000300002Essay rewriting and editing
write_essaySILO100300002Essay generation from short prompts
conversational_shortSISO100100010Short chat interactions
conversational_mediumMISO100010020005Medium-length chat with context
conversational_longLISO500010070002Long conversations with extensive history
  • SISO: Short Input, Short Output
  • MISO: Medium Input, Short Output
  • LISO: Long Input, Short Output
  • SILO: Short Input, Long Output
  • LILO: Long Input, Long Output
For detailed information about metrics, see Metrics.

​
Custom Benchmarks

Adding custom benchmarks is easy : just create a new yaml file in the src/benchmarks/benchmark_configs folder