How to Finetune, Quantize and Inference Phi-2

In this notebook & tutorial, we’ll explore the process of fine-tuning Phi-2, a language model with 2.7 billion parameters released by Microsoft Research in December 2023. Phi-2 surpasses the performance of Mistral and Llama-2 models at 7B and 13B parameters on various aggregated benchmarks. You can also access the tutorial directly through the provided colab notebook. For this tutorial, we have used an RLHF-like technique known as Direct Preference Optimization (DPO) and done the fine-tuning on Azure Cloud Platform. We will use the argilla/distilabel-intel-orca-dpo-pairs dataset from Argilla, which is an improvised version of the Intel/orca_dpo_pairs dataset from Intel. For model quantization, we will load the model in a 4-bit format using bitsandbytes. Finally, when deploying the model on Inferless, you can anticipate the following outcomes.

Inference Time	Cold Start Time	Token/Sec	Latency/Token	VRAM Required
11.96 secs	7.82 secs	21.34	46.85 ms	1.72 GB

Why finetuning?

Fine-tuning base language models, especially with techniques like Reinforcement Learning from Human Feedback (RLHF), is crucial for several reasons. RLHF allows the incorporation of human feedback to enhance the model’s performance, creating custom, task-specific, and expert models. The process involves setting up a good starting point, collecting human feedback, and iteratively improving the model. For this tutorial, we will use an RLHF-like technique known as Direct Preference Optimization (DPO). This technique aligns with human preferences better than existing methods, and it offers a promising alternative to RLHF for fine-tuning language models to meet specific human preferences.

Let’s get started:

Installing the Required Libraries

You need the following libraries for fine-tuning Phi-2 model using DPO.

!pip install -U bitsandbytes
!pip install -U transformers
!pip install -U accelerate
!pip install -U peft
!pip install -U trl
!pip install -U datasets
!pip install -U sentencepiece
!pip install -U wandb
!pip install -U pynvml

Dataset for DPO

The DPO trainer required a very specific type of dataset comprising instances of preferred and rejected responses in relation to a prompts. The preference dataset follows a defined format:

Prompt: It is the context prompt provided to the model during inference. It serves as the input for the text generation process.
Chosen: The “chosen” key holds the information about the preferred generated response corresponding to the given prompt.
Rejected: The “rejected” key contains information about a response that is not preferred or should be avoided when generating text in response to the provided prompt.

Now, we want our model to follow a specific chat template (ChatML), and we will format our dataset according to the requirements of DPOTrainer.

def format_data(example):

    system = tokenizer.apply_chat_template([{"role": "system", "content": example['system']}], tokenize=False) if example['system'] else ""
    prompt = tokenizer.apply_chat_template([{"role": "user", "content": example['input']}], tokenize=False, add_generation_prompt=True)

    return {
        "prompt": system + prompt,
        "chosen": example['chosen'] + "\n",
        "rejected": example['rejected'] + "\n",
    }

#Spliting the data in 95%(train) and 5%(eval)
train_ds = load_dataset('argilla/distilabel-intel-orca-dpo-pairs', split='train[:95%]')
eval_ds = load_dataset('argilla/distilabel-intel-orca-dpo-pairs', split='train[95%:]')

# Save columns
train_original_columns = train_ds.column_names
eval_original_columns = eval_ds.column_names

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"

# Format dataset
train_ds = train_ds.map(
    format_data,
    remove_columns=train_original_columns
)
eval_ds = eval_ds.map(
    format_data,
    remove_columns=eval_original_columns
)

Finetuning the Phi-2 model with DPO

Once you are done with the formatting of the dataset, you are now ready for the finetuning. DPO requires two models, the model that you want to finetune (Phi-2) and a reference model. Now load the tokenizer and the model then quantize and prepare the model for finetuning in 4bit using bitsandbytes.

#Load the Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=True, trust_remote_code=True)
tokenizer.pad_token = tokenizer.unk_token
tokenizer.pad_token_id =  tokenizer.unk_token_id
tokenizer.padding_side = 'left'

#Load the model
model = AutoModelForCausalLM.from_pretrained(model_name,torch_dtype=getattr(torch, "float16"), load_in_4bit=True, device_map={"": 0}, trust_remote_code=True)
model = prepare_model_for_kbit_training(model)

#Configure the pad token in the model
model.config.pad_token_id = tokenizer.pad_token_id
model.config.use_cache = False

#Load the Reference Model
model_ref = AutoModelForCausalLM.from_pretrained(model_name,load_in_4bit=True, torch_dtype=getattr(torch, "float16"), trust_remote_code=True, device_map={"": 0})

Define the LoRA configuration and the Training arguments required for finetuning the model. For the training hyperparameters, I have been following the Hugging Face settings of Zephyr 7B. Now you can define the DPOTrainer with the LoRA configuration and Training arguments, and then start the training process. We have used a single A100(80GB) GPU for the training.

# LoRA configuration
peft_config = LoraConfig(
        lora_alpha=16,
        lora_dropout=0.05,
        r=16,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules= ["q_proj","k_proj","v_proj", "dense"]
)

# Training arguments
training_arguments = TrainingArguments(
        output_dir="./results",
        evaluation_strategy="steps",
        do_eval=True,
        optim="paged_adamw_32bit",
        per_device_train_batch_size=4,
        gradient_accumulation_steps=4,
        per_device_eval_batch_size=2,
        log_level="debug",
        save_steps=10,
        logging_steps=1,
        learning_rate=5e-5,
        eval_steps=20,
        #num_train_epochs=1,
        max_steps=500,
        warmup_steps=100,
        bf16=True,
        lr_scheduler_type="cosine",
        report_to="wandb",
)

# Create DPO trainer
trainer = DPOTrainer(
    model,
    model_ref,
    args=training_arguments,
    train_dataset=train_ds,
    eval_dataset=eval_ds,
    tokenizer=tokenizer,
    peft_config=peft_config,
    beta=0.1,
    max_prompt_length=1024,
    max_length=1536,
)

# START DPO fine-tuning
trainer.train()

After finishing the training, combine the adapter with the original model.

# Save the adapter
trainer.model.save_pretrained("final_checkpoint")
tokenizer.save_pretrained("final_checkpoint")

#Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")

#Load the base model
base_model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2", torch_dtype="auto", trust_remote_code=True,device_map={"": 0})

#Merge base model with the adapter
model = PeftModel.from_pretrained(base_model, "final_checkpoint")
model = model.merge_and_unload()

#Save the model and the tokenizer
model.save_pretrained(new_model)
tokenizer.save_pretrained(new_model)

Quantize and Inference

Loading the finetuned model required 5.19 GB of VRAM. So, we will quantize and load the model to further reduce the memory requirement. Bitsandbytes enable you to load the model in 4 bits and further reduce the memory requirements to 1.72 GB. Here are our observations:

Inference Time	Cold Start Time	Token/Sec	Latency/Token	VRAM Required
11.96 secs	7.82 secs	21.34	46.85 ms	1.72 GB

We are using the bitsandbytes library, which enables you to run LLM on low memory. Using bitsandbytes only required 1.72 GB of GPU memory.

Last step, Tutorial to deploy on Inferless

Constructing the GitHub/GitLab Template

Now quickly construct the GitHub/GitLab template, this process is mandatory and make sure you don’t add any file named model.py

Phi-2/
├── app.py
├── inferless-runtime-config.yaml
├── inferless.yaml
└── input_schema.py

You can also add other files to this directory.

Create the class for inference

In the app.py we will define the class and import all the required functions

def initialize: In this function, you will initialize your model and define any variable that you want to use during inference.
def infer: This function gets called for every request that you send. Here you can define all the steps that are required for the inference. You can also pass custom values for inference and pass it through inputs(dict) parameters.
def finalize: This function cleans up all the allocated memory.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig, pipeline
import time

class InferlessPythonModel:
    def initialize(self):
        model_id = "Inferless/inferless-phi-2-DPO"
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True)
        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
        model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, quantization_config=bnb_config, device_map="cuda")
        self.pipe = pipeline("text-generation", model=model, tokenizer=self.tokenizer)

    def infer(self, inputs):
        prompt = inputs["prompt"]
        messages = [{"role": "system", "content":prompt}]

        prompt = self.pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        out = self.pipe(prompt, max_new_tokens=256, do_sample=True, top_p=0.9,temperature=0.9)
        generated_text = out[0]["generated_text"][len(prompt):]
        return {'generated_result': generated_text}

    def finalize(self):
        pass

Creating the Custom Runtime

This is a mandatory step where we allow the users to upload their custom runtime through inferless-runtime-config.yaml.

build:
  cuda_version: "12.1.1"
  system_packages:
    - "libssl-dev"
  python_packages:
    - "bitsandbytes==0.41.3"
    - "transformers==4.36.2"
    - "accelerate==0.25.0"
    - "scipy==1.11.4"

Method A: Deploying the model on Inferless Platform

Inferless supports multiple ways of importing your model. For this tutorial, we will use GitHub. Navigate to your desired workspace in Inferless and Click on Add a custom model button that you see on the top right. An import wizard will open up.

Step 2: Follow the UI to complete the model Import

Select the GitHub/GitLab Integration option to connect your source code repository with the deployment environment.
Navigate to the specific GitHub repository that contains your model’s code. Here, you will need to identify and enter the name of the model you wish to import.
Choose the appropriate type of machine that suits your model’s requirements. Additionally, specify the minimum and maximum number of replicas to define the scalability range for deploying your model.
Optionally, you have the option to enable automatic build and deployment. This feature triggers a new deployment automatically whenever there is a new code push to your repository.
If your model requires additional software packages, configure the Custom Runtime settings by including necessary pip or apt packages. Also, set up environment variables such as Inference Timeout, Container Concurrency, and Scale Down Timeout to tailor the runtime environment according to your needs.
Wait for the validation process to complete, ensuring that all settings are correct and functional. Once validation is successful, click on the “Import” button to finalize the import of your model.

Step 3: Wait for the model build to complete usually takes ~5-10 minutes

Step 4: Use the APIs to call the model

Once the model is in ‘Active’ status you can click on the ‘API’ page to call the model

Here is the Demo:

Method B: Deploying the model on Inferless CLI

Inferless allows you to deploy your model using Inferless-CLI. Follow the steps to deploy using Inferless CLI.

Clone the repository of the model

Let’s begin by cloning the model repository:

git clone https://github.com/inferless/Phi-2.git

Deploy the Model

To deploy the model using Inferless CLI, execute the following command:

inferless deploy --gpu A100 --runtime inferless-runtime-config.yaml

Explanation of the Command:

--gpu A100: Specifies the GPU type for deployment. Available options include A10, A100, and T4.
--runtime inferless-runtime-config.yaml: Defines the runtime configuration file. If not specified, the default Inferless runtime is used.

How-to Guides

​Why finetuning?

​Let’s get started:

​Installing the Required Libraries

​Dataset for DPO

​Finetuning the Phi-2 model with DPO

​Quantize and Inference

​Last step, Tutorial to deploy on Inferless

​Constructing the GitHub/GitLab Template

​Create the class for inference

​Creating the Custom Runtime

​Method A: Deploying the model on Inferless Platform

​Step 1: Login to the inferless dashboard can click on Import model button

​Step 2: Follow the UI to complete the model Import

​Step 3: Wait for the model build to complete usually takes ~5-10 minutes

​Step 4: Use the APIs to call the model

​Here is the Demo:

​Method B: Deploying the model on Inferless CLI

​Clone the repository of the model

​Deploy the Model