GenAIEval¶

Evaluation, benchmark, and scorecard, targeting for performance on throughput and latency, accuracy on popular evaluation harness, safety, and hallucination

Installation¶

Install from Pypi

pip install -r requirements.txt
pip install opea-eval

notes: We have to install requirements.txt at first, cause Pypi can’t have direct dependency with specific commit.

Build from Source

git clone https://github.com/opea-project/GenAIEval
cd GenAIEval
pip install -r requirements.txt
pip install -e .

Evaluation¶

lm-evaluation-harness¶

For evaluating the models on text-generation tasks, we follow the lm-evaluation-harness and provide the command line usage and function call usage. Over 60 standard academic benchmarks for LLMs, with hundreds of subtasks and variants implemented, such as ARC, HellaSwag, MMLU, TruthfulQA, Winogrande, GSM8K and so on.

command line usage¶

Gaudi2¶

# pip install --upgrade-strategy eager optimum[habana]
cd evals/evaluation/lm_evaluation_harness/examples
python main.py \
    --model gaudi-hf \
    --model_args pretrained=EleutherAI/gpt-j-6B \
    --tasks hellaswag \
    --device hpu \
    --batch_size 8

CPU¶

cd evals/evaluation/lm_evaluation_harness/examples
python main.py \
    --model hf \
    --model_args pretrained=EleutherAI/gpt-j-6B \
    --tasks hellaswag \
    --device cpu \
    --batch_size 8

function call usage¶

from evals.evaluation.lm_evaluation_harness import LMEvalParser, evaluate

args = LMevalParser(
    model="hf",
    user_model=user_model,
    tokenizer=tokenizer,
    tasks="hellaswag",
    device="cpu",
    batch_size=8,
)
results = evaluate(args)

remote service usage¶

setup a separate server with GenAIComps

# build cpu docker
docker build -f Dockerfile.cpu -t opea/lm-eval:latest .

# start the server
docker run -p 9006:9006 --ipc=host  -e MODEL="hf" -e MODEL_ARGS="pretrained=Intel/neural-chat-7b-v3-3" -e DEVICE="cpu" opea/lm-eval:latest

evaluate the model

set base_url, tokenizer and --model genai-hf

cd evals/evaluation/lm_evaluation_harness/examples

python main.py \
    --model genai-hf \
    --model_args "base_url=http://{your_ip}:9006,tokenizer=Intel/neural-chat-7b-v3-3" \
    --tasks  "lambada_openai" \
    --batch_size 2

bigcode-evaluation-harness¶

For evaluating the models on coding tasks or specifically coding LLMs, we follow the bigcode-evaluation-harness and provide the command line usage and function call usage. HumanEval, HumanEval+, InstructHumanEval, APPS, MBPP, MBPP+, and DS-1000 for both completion (left-to-right) and insertion (FIM) mode are available.

command line usage¶

cd evals/evaluation/bigcode_evaluation_harness/examples
python main.py \
    --model "codeparrot/codeparrot-small" \
    --tasks "humaneval" \
    --n_samples 100 \
    --batch_size 10 \
    --allow_code_execution

function call usage¶

from evals.evaluation.bigcode_evaluation_harness import BigcodeEvalParser, evaluate

args = BigcodeEvalParser(
    user_model=user_model,
    tokenizer=tokenizer,
    tasks="humaneval",
    n_samples=100,
    batch_size=10,
    allow_code_execution=True,
)
results = evaluate(args)

Kubernetes platform optimization¶

Node resource management helps optimizing AI container performance and isolation on Kubernetes nodes. See Platform optimization.

Benchmark¶

We provide a OPEA microservice benchmarking tool which is designed for microservice performance testing and benchmarking. It allows you to define test cases for various services based on YAML configurations, run load tests using stresscli, built on top of locust, and analyze the results for performance insights.

Features¶

Services load testing: Simulates high concurrency levels to test services like LLM, reranking, ASR, E2E and more.
YAML-based configuration: Easily define test cases, service endpoints, and parameters.
Service metrics collection: Optionally collect service metrics to analyze performance bottlenecks.
Flexible testing: Supports a variety of tests like chatqna, codegen, codetrans, faqgen, audioqna, and visualqna.
Data analysis and visualization: Visualize test results to uncover performance trends and bottlenecks.

How to use¶

Define Test Cases: Configure your tests in the benchmark.yaml file.

Increase File Descriptor Limit (if running large-scale tests):

ulimit -n 100000

This ensures the system can handle high concurrency by allowing more open files and connections.

Run the benchmark script:

python evals/benchmark/benchmark.py

Results will be saved in the directory specified by test_output_dir in the configuration.

For more details on configuring test cases, refer to the README.

Grafana Dashboards¶

Prometheus metrics collected during the tests can be used to create Grafana dashboards for visualizing performance trends and monitoring bottlenecks. For more information, refer to the Grafana README

tgi microservice dashboard