joffref / ai-benchmarks

Benchmarking suite for popular AI APIs

ai-benchmarks

This repo contains a handful of utilities for benchmarking the response latency of popular AI services, including:

Large Language Models (LLMs):

• OpenAI GPT-3.5, GPT-4 (from OpenAI or Azure OpenAI service)
• Anthropic Claude 3, Claude 2, Claude Instant
• Google Gemini Pro and PaLM 2 Bison
• Llama2 and 3 from several different providers, including
  • Anyscale
  • Azure
  • Cloudflare
  • Groq
  • OctoAI
  • Perplexity
  • Together
• Mixtral 8x7B from several different providers, including
  • Anyscale
  • Azure
  • Groq
  • OctoAI
  • Perplexity

Embedding Models:

• Ada-002
• Cohere

Text-to-Speech Models (TTS):

• ElevenLabs
• PlayHT

Leaderboard

Snapshot below, click it to jump to the latest spreadsheet.

Test methodology

• Tests are run from a Google Cloud console in us-west1.
• Input requests are short, typically a single message (~20 tokens), and typically ask for a brief output response.
• Max output tokens is set to 100, to avoid distortion of TPS values from long outputs.
• A warmup connection is made to remove any connection setup latency.
• The TTFT clock starts when the HTTP request is made and stops when the first token result is received in the response stream.
• For each provider, three separate inferences are done, and the best result is kept (to remove any outliers due to queuing etc).
• A best result is selected on 3 different days, and the median of these values is displayed.

Initial setup

This repo uses Poetry for dependency management. To install the dependencies, run:

pip install poetry
poetry install --sync

Running benchmarks

To run a benchmark, first set the appropriate environment variable (e.g., OPENAI_API_KEY, ELEVEN_API_KEY) etc, and then run the appropriate benchmark script.

LLM benchmarks

To generate LLM benchmarks, use the llm_benchmark.py script. For most providers, you can just pass the model name and the script will figure out what API endpoint to invoke. e.g.,

poetry run python llm_benchmark.py -m gpt-3.5-turbo "Write me a haiku."

However, when invoking generic models like Llama2, you'll need to pass in the base_url and api_key via the -b and -k parameters, e.g.,

poetry run python llm_benchmark.py -k $OCTOML_API_KEY -b https://text.octoai.run/v1 \
       -m llama-2-70b-chat-fp16 "Write me a haiku."

Similarly, when invoking Azure OpenAI, you'll need to specify your Azure API key and the base URL of your Azure deployment, e.g.,

poetry run python llm_benchmark.py -b https://fixie-westus.openai.azure.com \
       -m gpt-4-1106-preview "Write me a haiku."

See this script for more examples of how to invoke various providers.

Options

usage: llm_benchmark.py [-h] [--model MODEL] [--temperature TEMPERATURE] [--max-tokens MAX_TOKENS] [--base-url BASE_URL]
                        [--api-key API_KEY] [--no-warmup] [--num-requests NUM_REQUESTS] [--print] [--verbose]
                        [prompt]

positional arguments:
  prompt                                       Prompt to send to the API

optional arguments:
  -h, --help                                   show this help message and exit
  --model MODEL, -m MODEL                      Model to benchmark
  --temperature TEMPERATURE, -t TEMPERATURE    Temperature for the response
  --max-tokens, -T MAX_TOKEN                   Max tokens for the response
  --base-url BASE_URL, -b BASE_URL             Base URL for the LLM API endpoint
  --api-key API_KEY, -k API_KEY                API key for the LLM API endpoint
  --no-warmup                                  Don't do a warmup call to the API
  --num-requests NUM_REQUESTS, -n NUM_REQUESTS Number of requests to make
  --print, -p                                  Print the response
  --verbose, -v                                Print verbose output

Output

By default a summary of the requests is printed:

Latency saved: 0.01 seconds                <---- Difference between first response time and fastest reponse time
Optimized response time: 0.14 seconds      <---- fastest(http_response_time - http_start_time) of N requests
Median response time: 0.15 seconds         <---- median(http_response_time - http_start_time) of N requests
Time to first token: 0.34 seconds          <---- first_token_time - http_start_time
Tokens: 147 (211 tokens/sec)               <---- num_generated_tokens / (last_token_time - first_token_time)
Total time: 1.03 seconds                   <---- last_token_time - http_start_time

You can specify -p to print the output of the LLM, or -v to see detailed timing for each request.

TTS benchmarks

To generate TTS benchmarks, there are various scripts for the individual providers, e.g.,

python elevenlabs_stream_benchmark.py "Haikus I find tricky, With a 5-7-5 count, But I'll give it a go"

Playing audio

By default, only timing information for TTS is emitted. Follow the steps below to actually play out the received audio.

First, install mpv via

brew install mpv

Then, just pass the -p argument when generating text, e.g.,

python playht_benchmark.py -p "Well, basically I have intuition."

You can use the -v parameter to select which voice to use for generation.