Evaluation, benchmark, and scorecard, targeting for performance on throughput and latency, accuracy on popular evaluation harness, safety, and hallucination
- Install from Pypi
pip install -r requirements.txt
pip install opea-evalnotes: 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 .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.
# 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 8cd evals/evaluation/lm_evaluation_harness/examples
python main.py \
--model hf \
--model_args pretrained=EleutherAI/gpt-j-6B \
--tasks hellaswag \
--device cpu \
--batch_size 8from 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)-
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,tokenizerand--model genai-hfcd 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
-
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.
cd evals/evaluation/bigcode_evaluation_harness/examples
python main.py \
--model "codeparrot/codeparrot-small" \
--tasks "humaneval" \
--n_samples 100 \
--batch_size 10 \
--allow_code_executionfrom 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)Node resource management helps optimizing AI container performance and isolation on Kubernetes nodes. See Platform optimization.
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.
- 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.
Define Test Cases: Configure your tests in the benchmark.yaml file.
Increase File Descriptor Limit (if running large-scale tests):
ulimit -n 100000This ensures the system can handle high concurrency by allowing more open files and connections.
Run the benchmark script:
cd evals/benchmark/
python benchmark.pyNOTE: benchmark.py will take benchmark.yaml file as test case data.
By giving a --yaml argument with a custom yaml file, benchmark.py could take a custom test case data.
e.g. Test data for the examples starting by docker compose on HPU.
cd evals/benchmark/
python benchmark.py --yaml docker.hpu.benchmark.yamlResults 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.
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
