Hyperparameter Sweep training Llama 3.2 with WandB

Hyperparameter sweeps systematically test parameter combinations to find the best-performing model for the least compute or training time. This tutorial covers training Llama 3.2, using Wandb (Weights and Biases) to run hyperparameter sweeps and Cerebrium to scale experiments across serverless GPUs. View the final version on GitHub. Read this section if you’re unfamiliar with sweeps.

Analogy: Pizza Topping Sweep

Forget about ML for a second. Imagine making pizzas to discover the best combination of toppings. Three variables are available: • Type of Cheese (mozzarella, cheddar, parmesan) • Type of Sauce (tomato, pesto) • Extra Topping (pepperoni, mushrooms, olives) There are 12 possible combinations. One of them will taste the best. To find the tastiest pizza, try all combinations and rate them. This process is a hyperparameter sweep. The three hyperparameters are cheese, sauce, and extra topping. One pizza at a time takes hours. With 12 ovens, all pizzas bake at once and the best one emerges in minutes. If a kitchen is a GPU, 12 GPUs run all experiments in parallel. Cerebrium enables sweeps across 12 GPUs (or 1,000) to find the best model version fast.

Setup Cerebrium

If you don’t have a Cerebrium account, run the following:

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

This creates a folder with two files:

main.py - The entrypoint file where application code lives.
cerebrium.toml - A configuration file for build and environment settings.

The rest of this tutorial continues in this folder.

Setup Wandb

Weights & Biases (Wandb) tracks, visualizes, and manages machine learning experiments in real-time. It logs hyperparameters, metrics, and results for comparing models and optimizing performance.

pip install wandb
wandb login

A link prints in the terminal — click it and copy the API key back into the terminal. Add the W&B API key to Cerebrium secrets. In the Cerebrium Dashboard, navigate to the “secrets” tab in the left sidebar. Add the following:

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

Click the “Save All Changes” button.

Wandb authentication is now complete.

Training Script

To train with Llama 3.2, you’ll need:

Model access permission:
- Visit the Llama 3.2 model page on Hugging Face
- Accept all permissions
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”

The 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:
Takes a dictionary of parameters for flexible training configurations — the hyperparameter sweep.
Loads a customer support dataset from Hugging Face and formats it into chat template format
Implements QLoRA (Quantized Low-Rank Adaptation) for efficient fine-tuning.
Uses Weights & Biases (Wandb) for experiment tracking, logging results to the Wandb dashboard.
Saves the final model to a Cerebrium volume and returns a “success” message.

Deploy the training endpoint:

cerebrium deploy

This command:

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

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