Generate and run fine-tuned models with LoRA adapters

Learn how to generate models and adapters in formats suitable for executing with ONNX Runtime.

LoRA stands for Low Rank Adaptation. It is a popular method of fine-tuning that freezes some layers in a graph and provides the values of the weights of the variable layers in an artifact called an adapter.

Multi LoRA uses multiple adapters at runtime to run different fine-tunings of the same model. The adapter could be per-scenario, per-tenant/customer, or per-user i.e. there could be just a few adapters to many hundreds or thousands.

Olive generates models and adapters in ONNX format. These models and adapters can then be run with ONNX Runtime.

Setup

  1. Install Olive

    This installs Olive from main. Replace with version 0.8.0 when it is released.

    pip install git+https://github.com/microsoft/olive

  2. Install ONNX Runtime generate()

    pip install onnxruntime-genai

  3. Install other dependencies

    pip install optimum peft

  4. Downgrade torch and transformers

    TODO: There is an export bug with torch 2.5.0 and an incompatibility with transformers>=4.45.0

    pip uninstall torch
pip install torch==2.4
pip uninstall transformers
pip install transformers==4.44

  5. Choose a model

    You can use a model from HuggingFace, or your own model. The model must be a PyTorch model.

  6. Decide whether you are fine-tuning your model, or using a pre-existing adapter

    There are many pre-existing adapters on HuggingFace. If you are using multiple different adapters, these must all use the same fine-tuned layers of the original model.

Generate model and adapters in ONNX format

  1. If fine-tuning, run Olive to fine-tune your model

    Note: this operations requires a system with an NVIDIA GPU, with CUDA installed

    Use the olive fine-tune command: https://microsoft.github.io/Olive/how-to/cli/cli-finetune.html

    Here is an example usage of the command:

    olive finetune --method qlora -m meta-llama/Meta-Llama-3-8B -d nampdn-ai/tiny-codes --train_split "train[:4096]" --eval_split "train[4096:4224]" --text_template "### Language: {programming_language} \n### Question: {prompt} \n### Answer: {response}" --per_device_train_batch_size 16 --per_device_eval_batch_size 16 --max_steps 150 --logging_steps 50 -o adapters\tiny-codes

  2. Optionally, quantize your model

    Use the olive quantize command: https://microsoft.github.io/Olive/how-to/cli/cli-quantize.html

  3. Generate the ONNX model and adapter using the quantized model

    Use the olive auto-opt command for this step: https://microsoft.github.io/Olive/how-to/cli/cli-auto-opt.html

    The --adapter path can either be a HuggingFace adapter reference, or a path to the adapter you fine-tuned above.

    The --provider argument can be an ONNX Runtime execution provider.

    olive auto-opt -m <path to your model folder> --adapter_path <path to your adapter> -o <output model folder> --device cpu\|gpu --provider <provider>

  4. Convert adapters to .onnx_adapter format

    Run this step once for each adapter that you have generated.

    olive convert-adapters --adapter_path <path to your fine-tuned adapter --output_path <path to .onnx_adapter location --dtype float32

Write your application

This example is shown in Python, but you can also use the C/C++ API, the C# API, and the Java API (coming soon!)

import onnxruntime_genai as og
import numpy as np
import argparse

parser = argparse.ArgumentParser(description='Application to load and switch ONNX LoRA adapters')
parser.add_argument('-m', '--model', type=str, help='The ONNX base model')
parser.add_argument('-a', '--adapters', nargs='+', type=str, help='List of adapters in .onnx_adapters format')
parser.add_argument('-t', '--template', type=str, help='The template with which to format the prompt')
parser.add_argument('-s', '--system', type=str, help='The system prompt to pass to the model')
parser.add_argument('-p', '--prompt', type=str, help='The user prompt to pass to the model')
args = parser.parse_args()

model = og.Model(args.model)
if args.adapters:
    adapters = og.Adapters(model)
    for adapter in args.adapters:
        adapters.load(adapter, adapter)

tokenizer = og.Tokenizer(model)
tokenizer_stream = tokenizer.create_stream()

prompt = args.template.format(system=args.system, input=args.prompt)

params = og.GeneratorParams(model)
params.set_search_options(max_length=2048, past_present_share_buffer=False)
# This input is generated for transformers versions > 4.45
#params.set_model_input("onnx::Neg_67", np.array(0, dtype=np.int64))
params.input_ids = tokenizer.encode(prompt)

generator = og.Generator(model, params)

if args.adapters:
   for adapter in args.adapters:
      print(f"[{adapter}]: {prompt}")
      generator.set_active_adapter(adapters, adapter)

      while not generator.is_done():
        generator.compute_logits()
        generator.generate_next_token()

        new_token = generator.get_next_tokens()[0]
        print(tokenizer_stream.decode(new_token), end='', flush=True)
else:
    print(f"[Base]: {prompt}")

    while not generator.is_done():
       generator.compute_logits()
       generator.generate_next_token()

Call the application 

python app.py -m <model folder> -a <.onnx_adapter files> -t <prompt template> -s <systemm prompt> -p <prompt>

References