When training machine learning models, finding the perfect combination of hyper parameters can feel overwhelming however if you do it well, it can turn a good model into a great one! Hyper parameter sweeps help you find the best performing model for the least amount of compute or time spent training - think of them as your systematic approach to testing every variation to uncover the best result.

In this tutorial, we’ll walk through training Llama 3.2, using Wandb (Weights and Biases) to run hyper parameter sweeps to optimize its performance and we’ll leverage Cerebrium to scale our experiments across serverless GPUs, allowing us to find the best-performing model faster than ever.

If you would like to see the final version of this tutorial, you can view it on Github here.

Read this section if you’re unfamiliar with sweeps.

​
Analogy: Pizza Topping Sweep

Forget about ML for a second. Imagine you’re making pizzas, and you want to discover the most delicious combination of toppings. You can change three things about your pizza:

• Type of Cheese (mozzarella, cheddar, parmesan)

• Type of Sauce (tomato, pesto)

• Extra Topping (pepperoni, mushrooms, olives)

There are 12 possible combinations of pizzas you can make. One of them will taste the best!

To find out which pizza is the tastiest, you need to try all the combinations and rate them. This process is called a hyperparameter sweep. Your three hyperparameters are the cheese, sauce, and extra topping.

If you do it one pizza at a time, it could take hours. But if you had 12 ovens, you could bake all the pizzas at once and find the best one in just a few minutes!

If a kitchen is a GPU, then you need 12 GPUs to run each experiment to see which cookie is the best. The power of Cerebrium is the ability to run sweeps like this on 12 different GPUs (or 1,000 GPUs if you’d like) to get you the best version of a model fast.

​
Setup Cerebrium

If you don’t have a Cerebrium account, you can run the following in your cli:

pip install cerebrium --upgrade
cerebrium login
cerebrium init wandb-sweep

This creates a folder with two files:

  • main.py - Our entrypoint file where our code lives.
  • cerebrium.toml - A configuration file that contains all our build and environment settings Add the following pip packages near the bottom of your cerebrium.toml. This will be used in creating our deployment environment.

We will return back to these files later but for now, we will continue the rest of this tutorial in this folder.

​
Setup Wandb

Weights & Biases (Wandb) is a powerful tool for tracking, visualizing, and managing machine learning experiments in real-time. It helps you log hyperparameters, metrics, and results, making it easy to compare models, optimize performance, and collaborate effectively with your team.

  1. Sign up for a free account and then log in to your wandb account by running the following in your CLI.
pip install wandb
wandb login

You should see a link printed in your terminal - click it and copy the API from the webpage back into your terminal.

Add your W&B API key to Cerebrium secrets for use in your code. You can go to your Cerebrium Dashboard and navigate to the “secrets” tab in the left side bar. Add the following:

  • Key: WANDB_API_KEY
  • Value: The value you copied from the Wandb website.

Click the “Save All Changes” button to save the changes!

You should then be authenticated with Wandb and ready to go!

​
Training Script

To train with Llama 3.2, you’ll need:

  1. Model access permission:

  2. Hugging Face token:

    • Click your profile image (top right)
    • Select “Access token”
    • Create a new token if needed
    • Add to Cerebrium Secrets:
      • Key: HF_TOKEN
      • Value: Your Hugging Face token
    • Click “Save All Changes”

Our training script adapts this Kaggle notebook.

Create a requirements.txt file with these dependencies:

transformers
datasets
accelerate
peft
trl
bitsandbytes
wandb

These packages are required both locally and on Cerebrium. Update your cerebrium.toml to include:

  • The requirements.txt path
  • Hardware requirements for training
  • A 1-hour max timeout using response_grace_period

Add this configuration:

#existing configuration

[cerebrium.hardware]
cpu = 6
memory = 30.0
compute = "ADA_L40"

[cerebrium.scaling]
#existing configuration
response_grace_period = 3600

[cerebrium.dependencies.paths]
pip = "requirements.txt"

Install the dependencies locally:

pip install -r requirements.txt

Add this code to main.py:

import os
from typing import Dict
import wandb
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    TrainingArguments,
)
from peft import LoraConfig, get_peft_model
from datasets import load_dataset
from trl import SFTTrainer, setup_chat_format
import torch
import bitsandbytes as bnb
from huggingface_hub import login

login(token=os.getenv("HF_TOKEN"))

def find_all_linear_names(model):
    cls = bnb.nn.Linear4bit
    lora_module_names = set()
    for name, module in model.named_modules():
        if isinstance(module, cls):
            names = name.split('.')
            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
    if 'lm_head' in lora_module_names:
        lora_module_names.remove('lm_head')
    return list(lora_module_names)

def train_model(params: Dict):
    """
    Training function that receives parameters from the Cerebrium endpoint
    """
    # Initialize wandb
    wandb.login(key=os.getenv("WANDB_API_KEY"))
    wandb.init(
        project=params.get("wandb_project", "Llama-3.2-Customer-Support"),
        name=params.get("run_name", None),
        config=params,
    )

    # Model configuration
    base_model = params.get("base_model", "meta-llama/Llama-3.2-3B-Instruct")
    new_model = params.get("output_model_name", f"/persistent-storage/llama-3.2-3b-it-Customer-Support-{wandb.run.id}")

    # Set torch dtype and attention implementation
    torch_dtype = torch.bfloat16 if torch.cuda.get_device_capability()[0] >= 8 else torch.float16
    attn_implementation = "eager"

    # QLoRA config
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch_dtype,
        bnb_4bit_use_double_quant=True,
    )

    # Load model and tokenizer
    model = AutoModelForCausalLM.from_pretrained(
        base_model,
        quantization_config=bnb_config,
        device_map="auto",
        attn_implementation=attn_implementation
    )
    tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)

    # Find linear modules
    modules = find_all_linear_names(model)

    # LoRA config
    peft_config = LoraConfig(
        r=params.get("lora_r", 16),
        lora_alpha=params.get("lora_alpha", 32),
        lora_dropout=params.get("lora_dropout", 0.05),
        bias="none",
        task_type="CAUSAL_LM",
        target_modules=modules
    )

    # Setup model
    tokenizer.chat_template = None  # Clear existing chat template
    model, tokenizer = setup_chat_format(model, tokenizer)
    model = get_peft_model(model, peft_config)

    # Load and prepare dataset
    dataset = load_dataset(
        params.get("dataset_name", "bitext/Bitext-customer-support-llm-chatbot-training-dataset"),
        split="train"
    )
    dataset = dataset.shuffle(seed=params.get("seed", 65))
    if params.get("max_samples"):
        dataset = dataset.select(range(params.get("max_samples")))

    instruction = params.get("instruction",
        """You are a top-rated customer service agent named John.
        Be polite to customers and answer all their questions."""
    )

    def format_chat_template(row):
        row_json = [
            {"role": "system", "content": instruction},
            {"role": "user", "content": row["instruction"]},
            {"role": "assistant", "content": row["response"]}
        ]
        row["text"] = tokenizer.apply_chat_template(row_json, tokenize=False)
        return row

    dataset = dataset.map(format_chat_template, num_proc=4)
    dataset = dataset.train_test_split(test_size=params.get("test_size", 0.1))

    # Training arguments
    training_arguments = TrainingArguments(
        output_dir=new_model,
        per_device_train_batch_size=params.get("batch_size", 1),
        per_device_eval_batch_size=params.get("batch_size", 1),
        gradient_accumulation_steps=params.get("gradient_accumulation_steps", 2),
        optim="paged_adamw_32bit",
        num_train_epochs=params.get("epochs", 1),
        eval_strategy="steps",
        eval_steps=params.get("eval_steps", 0.2),
        logging_steps=params.get("logging_steps", 1),
        warmup_steps=params.get("warmup_steps", 10),
        learning_rate=params.get("learning_rate", 2e-4),
        fp16=params.get("fp16", False),
        bf16=params.get("bf16", False),
        group_by_length=params.get("group_by_length", True),
        report_to="wandb"
    )

    # Initialize trainer
    trainer = SFTTrainer(
        model=model,
        train_dataset=dataset["train"],
        eval_dataset=dataset["test"],
        peft_config=peft_config,
        args=training_arguments,
    )

    # Train and save
    model.config.use_cache = False
    trainer.train()
    model.config.use_cache = True

    # Save model
    trainer.model.save_pretrained(new_model)

    wandb.finish()
    return {"status": "success", "model_path": new_model}

You can read a deeper explanation of the training script here but here’s a high-level explanation of the code in bullet points:

  • This code sets up a fine-tuning pipeline for a Large Language Model (specifically Llama 3.2) using several modern training techniques:
  • The function takes a dictionary of parameters for flexibility in training configurations - this is our hyper parameter sweep.
  • We load a customer support dataset from Hugging Face and format the data into a chat template format
  • We implement QLoRA (Quantized Low-Rank Adaptation) for efficient fine-tuning.
  • We use Weights & Biases (wandb) for experiment tracking logging results to our Wandb dashboard as they are available.
  • At the end, we saves the final model to our Cerebrium volume and return a “success” message to show that the training was successful.

Deploy the training endpoint:

cerebrium deploy

This command:

  1. Sets up the environment with required packages
  2. Deploys the training script as an endpoint
  3. Returns a POST URL (save this for later)

Cerebrium requires no special decorators or syntax - just wrap your training code in a function. The endpoint automatically scales based on request volume, making it perfect for hyperparameter sweeps.

​
Hyperparameter Sweep

Let us create a run.py file that we will use to run locally. Put the following code in there:

import wandb
import requests
import os
from typing import Dict, Any
from dotenv import load_dotenv
load_dotenv()

CEREBRIUM_API_KEY = os.getenv("CEREBRIUM_API_KEY")
ENDPOINT_URL = "https://api.cortex.cerebrium.ai/v4/p-xxxxx/wandb-sweep/train_model?async=true"

def train_with_params(params: Dict[str, Any]):
    """
    Send training parameters to Cerebrium endpoint
    """
    headers = {
        "Authorization": f"Bearer {CEREBRIUM_API_KEY}",
        "Content-Type": "application/json"
    }

    response = requests.post(ENDPOINT_URL, json={"params": params}, headers=headers)
    if response.status_code != 202:
        raise Exception(f"Training failed: {response.text}")

    return response.json()

# Define the sweep configuration
sweep_config = {
    "method": "bayes",  # Bayesian optimization
    "metric": {
        "name": "eval/loss",
        "goal": "minimize"
    },
    "parameters": {
        "learning_rate": {
            "distribution": "log_uniform",
            "min": -10,  # exp(-10) ≈ 4.54e-5
            "max": -7,   # exp(-7) ≈ 9.12e-4
        },
        "batch_size": {
            "values": [1, 2, 4]
        },
        "gradient_accumulation_steps": {
            "values": [2, 4, 8]
        },
        "lora_r": {
            "values": [8, 16, 32]
        },
        "lora_alpha": {
            "values": [16, 32, 64]
        },
        "lora_dropout": {
            "distribution": "uniform",
            "min": 0.01,
            "max": 0.1
        },
        "max_seq_length": {
            "values": [512, 1024]
        }
    }
}

def main():
    # Initialize the sweep
    sweep_id = wandb.sweep(sweep_config, project="Llama-3.2-Customer-Support")

    def sweep_iteration():
        # Initialize a new W&B run
        with wandb.init() as run:
            # Get the parameters for this run
            params = wandb.config.as_dict()

            # Add any fixed parameters
            params.update({
                "wandb_project": "Llama-3.2-Customer-Support",
                "base_model": "meta-llama/Llama-3.2-3B-Instruct",
                "dataset_name": "bitext/Bitext-customer-support-llm-chatbot-training-dataset",
                "run_name": f"sweep-{run.id}",
                "epochs": 1,
                "test_size": 0.1,
            })

            # Call the Cerebrium endpoint with these parameters
            try:
                result = train_with_params(params)
                print(f"Training completed: {result}")
            except Exception as e:
                print(f"Training failed: {str(e)}")
                run.finish(exit_code=1)

    # Run the sweep
    wandb.agent(sweep_id, function=sweep_iteration, count=10)  # Run 10 experiments

if __name__ == "__main__":
    main()

This code implements a hyperparameter sweep system using Weights & Biases (wandb) sweeps to train a Llama 3.2 model for customer support. Here’s what it does:

  • Create a .env file and add your Inference API key from your Cerebrium Dashboard.
CEREBRIUM_API_KEY=eyJ....
  • Update the Cerebrium endpoint based on your project ID and the function name your wish to call. You will see we append this url with “?async=true”. This means its a fire-and-forgot request that can run up to 12 hours. You can read more here.
  • We then define a Bayesian optimization sweep configuration that will search through different hyperparameters including:
    • Learning rate (log uniform distribution between ~4.54e-5 and ~9.12e-4)
    • Batch size (1, 2, or 4)
    • Gradient accumulation steps (2, 4, or 8)
    • LoRA parameters (r, alpha, and dropout)
    • Maximum sequence length (512 or 1024)
  • We create this sweep in the “Llama-3.2-Customer-Support” W&B project
  • For each sweep iteration:
    • We initialize a new W&B run
    • Combines the sweep’s hyperparameters with fixed parameters (like model name and dataset)
    • Sends the parameters to a Cerebrium endpoint for training that happens asynchronously.
    • Logs the results back to W&B
  • Run these combinations across 10 experiments (10 concurrent GPU’s is the limit on Cerebrium’s Hobby plan)

Run the script:

python run.py

Monitor training progress in your W&B dashboard:

​
Next Steps

  1. Export model:

    • Copy to AWS S3 using Boto3
    • Download locally using Cerebrium Python package

  2. Quality assurance:

  3. Deployment:

    • Create inference endpoint
    • Load model directly from Cerebrium volume

​
Conclusion

Hyperparameter optimization becomes manageable with the right tools. This tutorial showed how combining W&B for tracking and Cerebrium for serverless compute enables efficient hyperparameter sweeps for Llama 3.2, optimizing model performance with minimal effort.

View the complete code in the GitHub repository

Gradio Chat Interface