LLMVal is a tool for validating and benchmarking LLMs.
Validation: we send a simple query to the LLM and ensure the returned data is valid. In particular it checks for inter-request cross-over (request A gets the responses for request B).
Benchmarking: LLMVal measures time to first token (TTFT), inter-token latency (ITL) and requests that take longer than 3 seconds to start returning data.
Variation in input and output token lengths is a design parameter since this is intended to be representative. This is because there are some optimizations (e.g. continuous batching) that we know work better with varying input and output length.
Currently supported endpoints include:
- Any OpenAI compatible endpoints, including Anyscale Endpoints, Anyscale Private Endpoints, OpenAI, Fireworks, Perplexity etc
- Together
- Vertex AI
- SageMaker
Please see requirments.txt for more details on dependency requirments.
This is prototype code. We are currently refactoring the code to be more extensible (including a pluggable endpoints, varying traffic load etc).
In addition we plan to:
- Make running the benchmark not only possible from command line, but also possible to integrate easily into CI/CD or job scheduling systems.
- Control where the generated files and information go.
- Automate report generation.
We expect this refactor to be complete some time in November 2023.
Many LLM providers have extremely low rate limits by default (e.g. Perplexity 3 requests per 90 seconds).
You can use the sleep parameter to overcome these difficulties, but it does affect the representativeness of the results.
Other systems do not have rate limits, but we consider that if the TTFT exceeds 3 second for more than 5% of queries that the system is overloaded.
Default values are the ones that we use for testing Anyscale Endpoints. The distribution of inputs and outputs roughly mirrors the input and output patterns we see there.
We recommend setting the seed (or using the provided seed) to reduce variable but still have randomization.
Do a python llmval.py --help to see all options.
- Provide API base and key in .env file. Check out env_sample.txt
- Test out Anyscale Endpoint with following command by sending 20 requests
python llmval.py -r 20 -m "meta-llama/Llama-2-70b-chat-hf" - Control input token numbers by setting min/max lines, and control output token number by setting req-lines and max_tokens
python llmval.py -r 20 -f openai -m "gpt-3.5-turbo" --min-lines 8 --max-lines 10
python llmval.py -r 20 -f openai -m "gpt-3.5-turbo" --req-lines 3 --max-tokens 128 - Control sleep between rounds to avoid hitting rate limit
python llmval.py -r 20 -f fireworks -m "accounts/fireworks/models/llama-v2-70b-chat" --sleep 10 - Output will be saved at framework-timestamp.json and framework-timestamp_raw.json
- Use Jupyter with analyze-raw.ipynb to visualize and/or interact with the raw data.
