Run Phi-3 language models with the ONNX Runtime generate() API

Introduction

Phi-3 and Phi 3.5 ONNX models are hosted on HuggingFace and you can run them with the ONNX Runtime generate() API.

The mini (3.3B) and medium (14B) versions available now, with support. Both mini and medium have a short (4k) context version and a long (128k) context version. The long context version can accept much longer prompts and produce longer output text, but it does consume more memory.

Available models are:

This tutorial demonstrates how to download and run the short context (4k) mini (3B) model variant pf Phi 3 model. See the model reference for download commands for the other variants.

This tutorial downloads and runs the short context (4k) mini (3B) model variant. See the model reference for download commands for the other variants.

Setup

  1. Install the git large file system extension

    HuggingFace uses git for version control. To download the ONNX models you need git lfs to be installed, if you do not already have it.

    • Windows: winget install -e --id GitHub.GitLFS (If you don’t have winget, download and run the exe from the official source)
    • Linux: apt-get install git-lfs
    • MacOS: brew install git-lfs

    Then run git lfs install

  2. Install the HuggingFace CLI

    pip install huggingface-hub[cli]

Choose your platform

Are you on a Windows machine with GPU?

  • I don’t know → Review this guide to see whether you have a GPU in your Windows machine and confirm that your GPU is DirectML-supported
  • Yes → Follow the instructions for DirectML.
  • No → Do you have an NVIDIA GPU?
    • I don’t know → Review this guide to see whether you have a CUDA-capable GPU.
    • Yes → Follow the instructions for NVIDIA CUDA GPU.
    • No → Follow the instructions for CPU.

Note: Only one package and model is required based on your hardware. That is, only execute the steps for one of the following sections

Run with DirectML

  1. Download the model

    huggingface-cli download microsoft/Phi-3-mini-4k-instruct-onnx --include directml/* --local-dir .

    This command downloads the model into a folder called directml.

  2. Install the generate() API

    pip install numpy
pip install --pre onnxruntime-genai-directml

    You should now see onnxruntime-genai-directml in your pip list.

  3. Run the model

    Run the model with phi3-qa.py.

    curl https://raw.githubusercontent.com/microsoft/onnxruntime-genai/main/examples/python/phi3-qa.py -o phi3-qa.py
python phi3-qa.py -m directml\directml-int4-awq-block-128

    Once the script has loaded the model, it will ask you for input in a loop, streaming the output as it is produced the model. For example:

    Input: Tell me a joke about GPUs

Certainly! Here\'s a light-hearted joke about GPUs:


Why did the GPU go to school? Because it wanted to improve its "processing power"!


This joke plays on the double meaning of "processing power," referring both to the computational abilities of a GPU and the idea of a student wanting to improve their academic skills.

Run with NVIDIA CUDA

  1. Download the model

    huggingface-cli download microsoft/Phi-3-mini-4k-instruct-onnx --include cuda/cuda-int4-rtn-block-32/* --local-dir .

    This command downloads the model into a folder called cuda.

  2. Install the generate() API

    pip install numpy
pip install --pre onnxruntime-genai-cuda --index-url=https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-genai/pypi/simple/

  3. Run the model

    Run the model with phi3-qa.py.

    curl https://raw.githubusercontent.com/microsoft/onnxruntime-genai/main/examples/python/phi3-qa.py -o phi3-qa.py
python phi3-qa.py -m cuda/cuda-int4-rtn-block-32

    Once the script has loaded the model, it will ask you for input in a loop, streaming the output as it is produced the model. For example:

    Input: Tell me a joke about creative writing
 
Output:  Why don't writers ever get lost? Because they always follow the plot!

Run on CPU

  1. Download the model

    huggingface-cli download microsoft/Phi-3-mini-4k-instruct-onnx --include cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/* --local-dir .

    This command downloads the model into a folder called cpu_and_mobile

  2. Install the generate() API for CPU

    pip install numpy
pip install --pre onnxruntime-genai

  3. Run the model

    Run the model with phi3-qa.py.

    curl https://raw.githubusercontent.com/microsoft/onnxruntime-genai/main/examples/python/phi3-qa.py -o phi3-qa.py
python phi3-qa.py -m cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4

    Once the script has loaded the model, it will ask you for input in a loop, streaming the output as it is produced the model. For example:

    Input: Tell me a joke about generative AI

Output:  Why did the generative AI go to school?

To improve its "creativity" algorithm!

Phi-3 ONNX model reference

Phi-3 mini 4k context CPU 

huggingface-cli download microsoft/Phi-3-mini-4k-instruct-onnx --include cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/* --local-dir .
python phi3-qa.py -m cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4

Phi-3 mini 4k context CUDA 

huggingface-cli download microsoft/Phi-3-mini-4k-instruct-onnx --include cuda/cuda-int4-rtn-block-32/* --local-dir .
python phi3-qa.py -m cuda/cuda-int4-rtn-block-32

Phi-3 mini 4k context DirectML 

huggingface-cli download microsoft/Phi-3-mini-4k-instruct-onnx --include directml/* --local-dir .
python phi3-qa.py -m directml\directml-int4-awq-block-128

Phi-3 mini 128k context CPU 

huggingface-cli download microsoft/Phi-3-mini-128k-instruct-onnx --include cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/* --local-dir .
python phi3-qa.py -m cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4

Phi-3 mini 128k context CUDA 

huggingface-cli download microsoft/Phi-3-mini-128k-instruct-onnx --include cuda/cuda-int4-rtn-block-32/* --local-dir .
python phi3-qa.py -m cuda/cuda-int4-rtn-block-32

Phi-3 mini 128k context DirectML 

huggingface-cli download microsoft/Phi-3-mini-128k-instruct-onnx --include directml/* --local-dir .
python phi3-qa.py -m directml\directml-int4-awq-block-128

Phi-3 medium 4k context CPU 

git clone https://huggingface.co/microsoft/Phi-3-medium-4k-instruct-onnx-cpu
python phi3-qa.py -m Phi-3-medium-4k-instruct-onnx-cpu/cpu-int4-rtn-block-32-acc-level-4

Phi-3 medium 4k context CUDA 

git clone https://huggingface.co/microsoft/Phi-3-medium-4k-instruct-onnx-cuda
python phi3-qa.py -m Phi-3-medium-4k-instruct-onnx-cuda/cuda-int4-rtn-block-32

Phi-3 medium 4k context DirectML 

git clone https://huggingface.co/microsoft/Phi-3-medium-4k-instruct-onnx-directml
python phi3-qa.py -m Phi-3-medium-4k-instruct-onnx-directml/directml-int4-awq-block-128

Phi-3 medium 128k context CPU 

git clone https://huggingface.co/microsoft/Phi-3-medium-128k-instruct-onnx-cpu
python phi3-qa.py -m Phi-3-medium-128k-instruct-onnx-cpu/cpu-int4-rtn-block-32-acc-level-4

Phi-3 medium 128k context CUDA 

git clone https://huggingface.co/microsoft/Phi-3-medium-128k-instruct-onnx-cuda
python phi3-qa.py -m Phi-3-medium-128k-instruct-onnx-cuda/cuda-int4-rtn-block-32

Phi-3 medium 128k context DirectML 

git clone https://huggingface.co/microsoft/Phi-3-medium-128k-instruct-onnx-directml
python phi3-qa.py -m Phi-3-medium-128k-instruct-onnx-directml/directml-int4-awq-block-128

Phi-3.5 mini 128k context CUDA 

huggingface-cli download microsoft/Phi-3.5-mini-instruct-onnx --include cuda/cuda-int4-awq-block-128/* --local-dir .
python phi3-qa.py -m cuda/cuda-int4-awq-block-128

Phi-3.5 mini 128k context CPU 

huggingface-cli download microsoft/Phi-3.5-mini-instruct-onnx --include cpu_and_mobile/cpu-int4-awq-block-128-acc-level-4/* --local-dir .
python phi3-qa.py -m cpu_and_mobile/cpu-int4-awq-block-128-acc-level-4