lita 0.0.7

LLM Integrated Testing & Analysis

LITA

LLM Integrated Testing & Analysis Framework

Lita is a comprehensive testing and analysis framework for Large Language Models (LLMs), designed to provide an integrated environment for efficient execution, benchmarking, and profiling.

Key Features:

• Multi-Framework Support: Run models on various execution backends, including vLLM, Hugging Face (HF), and ONNX Runtime (ORT).
• Performance Profiling: Collect detailed execution metrics using built-in profilers like vllmprof.

Lita is designed to help researchers and developers evaluate and optimize LLM workloads across different execution environments, enabling seamless integration into existing machine learning workflows.

Installation (Development)

To set up the development environment, use the following commands:

pip install -e .
pip install torch==2.6.0, torchvision
pip uninstall onnxruntime onnxruntime-gpu
pip install onnxruntime-gpu==1.19

Ensure that PyTorch version 2.6.0 or later is installed.

Setting Cache Directory

Before running the Lita framework, you can configure the cache directory by setting environment variables. Add the following lines to your .bashrc (or .bash_profile for macOS):

export LITA_CACHE="your_cache_path"
export HF_HOME=$LITA_CACHE

Apply the changes by running:

source ~/.bashrc  # or source ~/.bash_profile

Now, Lita will use the specified cache directory.

Usage

1. Running Models with Different Frameworks

Lita supports executing models on various frameworks such as vLLM, Hugging Face (HF), and ONNX Runtime (ORT).

from lita import Lita
import time

model_name = "meta-llama/Llama-3.2-3B-Instruct"

for mode in ["vllm", "hf", "ort"]:
    mm = Lita(model_name, mode)
    output_str = mm.generate('hello')
    print(f"Generation output: {output_str}")

2. Performance Measurement

Lita provides built-in performance profiling using vLLM Profiler (vllmprof).

from lita import Lita
import time

model_name = "meta-llama/Llama-3.2-3B-Instruct"
mode = "hf"

mm = Lita(model_name, mode, perf='vllmprof')
output_str = mm.generate('hello')

print(output_str)
print(mm.metric.summary())

This script runs the model in hf mode while enabling performance profiling with vllmprof. After text generation, it prints the output along with detailed performance metrics.

3. Benchmark Test

Lita provides built-in benchmarking capabilities for evaluating model performance on standardized datasets. The following script demonstrates running a benchmark using the MMLU dataset.

from lita import Lita
from lita.benchmark.mmlu import MMLUDataLoader
from lita.benchmark.commons import run_benchmark, extract_choice
from lita.utils import get_system_info
import json

mmlu_ = MMLUDataLoader(n_shots=5)

model_name = "meta-llama/Llama-3.2-3B-Instruct"
model = Lita(model_name, "hf", perf="time")
        
log = run_benchmark(mmlu_, model, 1, extract_fn=extract_choice)

json_log_data = {
    "system_info": get_system_info(),
    "model_info": model.get_configs(),
    "benchmark_data": log
}
with open("examples/mmlu_log.json", "w", encoding="utf-8") as f:
    json.dump(json_log_data, f, indent=4, ensure_ascii=False)

This script performs an MMLU benchmark test by loading the dataset, running the model using transformers, and measuring performance at the simple time. The results, including system specifications and model configuration, are saved in examples/mmlu_log.json for further analysis.