Model-Optimizer/examples/windows/accuracy_benchmark

Table of Contents

• Overview
• Prerequisites
• Accuracy Benchmarks
  • MMLU (Massive Multitask Language Understanding)
    • Setup
    • Evaluation Methods
• Additional Metrics
• API changes in ONNX Runtime GenAI v0.6
• Troubleshoot

Overview

This repository provides scripts, popular third-party benchmarks, and instructions for evaluating the accuracy of Large Language Models (LLMs). It demonstrates how to use a ModelOpt quantized LLM with various established benchmarks, including deployment options using DirectML and TensorRT-LLM in a Windows environment.

Prerequisites

Category	Details
Operating System	Windows 10 or later
Python	- For ORT-DML GenAI, use Python 3.11. - For TensorRT-LLM, use Python 3.10. - All other backends are compatible with both Python 3.10 and 3.11.
Package Manager	pip
Compatible Hardware and Drivers	- Ensure necessary hardware (e.g., CUDA-compatible GPU) and drivers are installed, depending on the evaluation method: - DirectML for DirectML-based evaluation - CUDA for TensorRT
Additional Tools	- cmd: Recommended for running the provided commands. - Tar Utility: Included in Windows 10 and later via PowerShell. - Curl: Included in Windows 10 and later via PowerShell.

Accuracy Benchmarks

MMLU (Massive Multitask Language Understanding)

The MMLU benchmark assesses LLM performance across a wide range of tasks, producing a score between 0 and 1, where a higher score indicates better accuracy. Please refer the MMLU Paper for more details on this.

Setup

The table below lists the setup steps to prepare your environment for evaluating LLMs using the MMLU benchmark.

Step	Command or Description
Open PowerShell as Administrator	-
Create and Activate a Virtual Environment (Optional but Recommended)	python -m venv llm_env .\llm_env\Scripts\Activate.ps1
Install PyTorch and Related Packages	pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128
Install ONNX Runtime Packages	pip install onnxruntime-directml==1.21.1 pip install onnxruntime-genai-directml==0.6.0
Install Benchmark Requirements	pip install -r requirements.txt
Download MMLU Data	mkdir data curl.exe -L -o .\data\mmlu.tar https://huggingface.co/datasets/cais/mmlu/resolve/c30699e8356da336a370243923dbaf21066bb9fe/data.tar tar -xf .\data\mmlu.tar -C .\data Move-Item .\data\data .\data\mmlu

Evaluation Methods

Once the MMLU benchmark is set up, you can use the mmlu_benchmark.py script to evaluate LLMs deployed with various backends. Please refer examples below.

MMLU Benchmark with GenAI APIs for ORT-DML Deployment

To run the model with ORT-DML using GenAI, use the --ep genai_dml argument.

• Test Suite

  python mmlu_benchmark.py `
      --model_name causal `
      --model_path <ONNX_model_folder> `
      --ep genai_dml `
      --output_file <output_log_file.json> `
      --ntrain 5
  

• Specific Subjects

  python mmlu_benchmark.py `
      --model_name causal `
      --model_path <ONNX_model_folder> `
      --ep genai_dml `
      --output_file <output_log_file.json> `
      --subject abstract_algebra,anatomy,college_mathematics `
      --ntrain 5
  

MMLU Benchmark with ONNX Runtime APIs for DML, CUDA, or CPU Deployment

To run the model with ORT-DML, ORT-CUDA or ORT-CPU execution providers, use --ep ort_dml, --ep ort_cuda, or --ep ort_cpu respectively.

• Test Suite

  python mmlu_benchmark.py `
      --model_name causal `
      --model_path <ONNX_model_folder> `
      --ep ort_dml `
      --output_file <output_log_file.json> `
      --ntrain 5
  

• Specific Subjects

  python mmlu_benchmark.py `
      --model_name causal `
      --model_path <ONNX_model_folder> `
      --ep ort_dml `
      --output_file <output_log_file.json> `
      --subject abstract_algebra,anatomy,college_mathematics `
      --ntrain 5
  

MMLU Benchmark with Transformer APIs for PyTorch Hugging Face Models

To evaluate the PyTorch Hugging Face (HF) model, use the --ep pt argument.

• Test Suite

  python mmlu_benchmark.py `
      --model_name causal `
      --model_path <ONNX_model_folder> `
      --ep pt `
      --output_file <output_log_file.json> `
      --ntrain 5 `
      --dtype <torch_dtype in model's config.json {float16|bfloat16}>
  

• Specific Subjects

  python mmlu_benchmark.py `
      --model_name causal `
      --model_path <ONNX_model_folder> `
      --ep pt `
      --output_file <output_log_file.json> `
      --subject abstract_algebra,anatomy,college_mathematics `
      --ntrain 5 `
      --dtype <torch_dtype in model's config.json {float16|bfloat16}>
  

MMLU Benchmark with TensorRT-LLM APIs for TensorRT-LLM Deployment

1. Install TensorRT-LLM and Compatible PyTorch

   pip install torch==2.4.0+cu121 --index-url https://download.pytorch.org/whl
   pip install tensorrt_llm==0.12.0 `
       --extra-index-url https://pypi.nvidia.com `
       --extra-index-url https://download.pytorch.org/whl/cu121/torch/
   

2. Run the Benchmark

   • Test Suite

     python mmlu_benchmark.py `
         --model_name causal `
         --hf_model_dir <hf_model_path> `
         --engine_dir <engine_path> `
         --ep trt-llm `
         --ntrain 5 `
         --output_file result.json
     

   • Specific Subjects

     python mmlu_benchmark.py `
         --model_name causal `
         --hf_model_dir <hf_model_path> `
         --engine_dir <engine_path> `
         --ep trt-llm `
         --ntrain 5 `
         --output_file result.json `
         --subject abstract_algebra,anatomy,college_mathematics
     

Additional Metrics

Metric	Directory	Description
KL Divergence	kl_divergence_metrics/	Measures output similarity between two models using KL divergence
Perplexity	perplexity_metrics/	Evaluates language model quality using WikiText-2 perplexity
FVD	fvd_metrics/	Computes Fréchet Video Distance between two sets of videos using I3D features

Each sub-directory contains its own README.md with detailed setup and usage instructions.

API changes in ONNX Runtime GenAI v0.6

In onnxruntime-genai (GenAI) v0.6, generator.compute_logits() and generator_params.input_ids are deprecated and new API generator.append_tokens(List: token_ids) is added (see GenAI PR-867 for details).

So, this MMLU script has been updated accordingly - refer following change-snippet from this MMLU script (left works with GenAI < 0.6, right works with GenAI 0.6+). Make sure to update the MMLU script accordingly (left part) for trying it with GenAI < 0.6.

Troubleshoot

1. In case of any model specific issue (e.g. in tokenizer or in onnxruntime-genai package etc.), one can try using older GenAI e.g. export the ONNX model with onnxruntime-genai-directml 0.4 and transformers 4.44.

2. In case of trying out MMLU run of ONNX model through GenAI, make sure that the input model is running fine with GenAI. Onnxruntime-genai has example inference scripts (e.g. see phi3 example script).