Finetuning distribué de Llama-3.2-3B avec Unsloth sur plusieurs GPU
Affiner le Llama-3.2-3B LLM sur 8 GPU H100 à l'aide des bibliothèques Unsloth et serverless_gpu sur AI Runtime.
remarque
Cet exemple requiert l'environnement Databricks AI de version 6 ou une version ultérieure.
Connectez-vous au compute
- Sélectionnez 8xH100 comme accélérateur dans le panneau d’environnement.
- Choisissez AI v6 dans l'option Environnement de base
- Cliquez sur "Appliquer".
Notez que cela peut prendre jusqu'à 8 minutes.
L'environnement AI v6 n'inclut pas Unsloth. Installez Unsloth avant de continuer.
Python
%pip install unsloth==2026.9.4
Python
import os
os.environ["UNSLOTH_COMPILE_DISABLE"] = "1"
Python
dbutils.widgets.text("uc_catalog", "main")
dbutils.widgets.text("uc_schema", "default")
dbutils.widgets.text("uc_model_name", "llama-3_2-3b")
dbutils.widgets.text("uc_volume", "checkpoints")
UC_CATALOG = dbutils.widgets.get("uc_catalog")
UC_SCHEMA = dbutils.widgets.get("uc_schema")
UC_MODEL_NAME = dbutils.widgets.get("uc_model_name")
UC_VOLUME = dbutils.widgets.get("uc_volume")
print(f"UC_CATALOG: {UC_CATALOG}")
print(f"UC_SCHEMA: {UC_SCHEMA}")
print(f"UC_MODEL_NAME: {UC_MODEL_NAME}")
print(f"UC_VOLUME: {UC_VOLUME}")
# Model selection - Choose based on your compute constraints
MODEL_NAME = "unsloth/Llama-3.2-3B-Instruct" # or choose "unsloth/Llama-3.2-1B-Instruct"
OUTPUT_DIR = f"/Volumes/{UC_CATALOG}/{UC_SCHEMA}/{UC_VOLUME}/{UC_MODEL_NAME}" # Save checkpoint to UC Volume
DATASET_NAME = "mlabonne/FineTome-100k"
print(f"MODEL_NAME: {MODEL_NAME}")
print(f"OUTPUT_DIR: {OUTPUT_DIR}")
print(f"DATASET_NAME: {DATASET_NAME}")
Python
from serverless_gpu import distributed
from serverless_gpu import runtime as rt
@distributed(gpus=8, gpu_type='h100')
def run_train():
from datasets import load_dataset
import logging
import mlflow
import os
import torch
# Set up device for distributed training
local_rank = int(os.environ.get("LOCAL_RANK", 0))
torch.cuda.set_device(local_rank)
# IMPORTANT: import unsloth BEFORE trl
from unsloth import FastLanguageModel, is_bfloat16_supported
from unsloth.chat_templates import get_chat_template, standardize_sharegpt, train_on_responses_only
from trl import SFTTrainer, SFTConfig
from transformers import DataCollatorForSeq2Seq
from transformers.integrations import MLflowCallback
max_seq_length = 2048 # Choose any!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = False # Use 4bit quantization to reduce memory usage. Can be False.
# Load model and tokenizer
model, tokenizer = FastLanguageModel.from_pretrained(
model_name=MODEL_NAME,
max_seq_length=max_seq_length,
dtype=dtype,
load_in_4bit=load_in_4bit,
device_map={'': local_rank},
# token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)
model = FastLanguageModel.get_peft_model(
model,
r=16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
"gate_proj", "up_proj", "down_proj",],
lora_alpha=16,
lora_dropout=0, # Supports any, but = 0 is optimized
bias="none", # Supports any, but = "none" is optimized
use_gradient_checkpointing=True,
random_state=3407,
use_rslora=False, # We support rank stabilized LoRA
loftq_config=None, # And LoftQ
)
# DDP needs non-reentrant gradient checkpointing to avoid "mark a variable ready only once"
model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
# Process data
tokenizer = get_chat_template(
tokenizer,
chat_template="llama-3.1",
)
def formatting_prompts_func(examples):
convos = examples["conversations"]
texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False) for convo in convos]
return { "text" : texts, }
dataset = load_dataset(DATASET_NAME, split="train")
dataset = standardize_sharegpt(dataset)
dataset = dataset.map(formatting_prompts_func, batched=True,)
trainer = SFTTrainer(
model = model,
processing_class = tokenizer,
train_dataset = dataset,
data_collator = DataCollatorForSeq2Seq(tokenizer = tokenizer),
args = SFTConfig(
dataset_text_field = "text",
max_length = max_seq_length,
dataset_num_proc = 6,
packing = False, # Can make training 5x faster for short sequences.
per_device_train_batch_size = 2,
gradient_accumulation_steps = 4,
warmup_steps = 5,
# num_train_epochs = 1, # Set this for 1 full training run.
max_steps = 25,
learning_rate = 2e-4,
fp16 = not is_bfloat16_supported(),
bf16 = is_bfloat16_supported(),
logging_steps = 1,
optim = "adamw_8bit",
weight_decay = 0.01,
lr_scheduler_type = "linear",
seed = 3407,
output_dir = OUTPUT_DIR,
report_to = "mlflow", # Use MLflow to track model metrics,
run_name = f"{MODEL_NAME}-finetune-unsloth",
),
)
trainer = train_on_responses_only(
trainer,
instruction_part="<|start_header_id|>user<|end_header_id|>\n\n",
response_part="<|start_header_id|>assistant<|end_header_id|>\n\n",
num_proc=1
)
trainer.train()
# Save model
if rt.get_global_rank() == 0:
logging.info("\nSaving trained model...")
trainer.save_model(OUTPUT_DIR)
logging.info("✓ LoRA adapters saved - use with base model for inference")
tokenizer.save_pretrained(OUTPUT_DIR)
logging.info("✓ Tokenizer saved with model")
logging.info(f"\n🎉 All artifacts saved to: {OUTPUT_DIR}")
mlflow_run_id = None
if mlflow.last_active_run() is not None:
mlflow_run_id = mlflow.last_active_run().info.run_id
return mlflow_run_id
Python
run_id = run_train.distributed()[0]
Enregistrement de MLflow et Unity Catalog
Stratégie d’enregistrement du modèle
- MLflow Tracking : Enregistrer les artefacts et les métadonnées du modèle
- Unity Catalog : Enregistrer le modèle pour la gouvernance et le déploiement
- Gestion des versions de modèle : Versioning automatique pour la gestion du cycle de vie du modèle
- Métadonnées : Informations complètes sur le modèle pour la reproductibilité
Python
print("\nRegistering model with MLflow and Unity Catalog...")
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import mlflow
# Load the trained model for registration
print("Loading LoRA model for registration...")
# For LoRA models, we need both base model and adapter
base_model = AutoModelForCausalLM.from_pretrained(
MODEL_NAME,
trust_remote_code=True
)
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
adapter_dir = OUTPUT_DIR
peft_model = PeftModel.from_pretrained(base_model, adapter_dir)
# Merge LoRA into base and drop PEFT wrappers
merged_model = peft_model.merge_and_unload()
components = {
"model": merged_model,
"tokenizer": tokenizer,
}
# Create Unity Catalog model name
full_model_name = f"{UC_CATALOG}.{UC_SCHEMA}.{UC_MODEL_NAME}"
print(f"Registering model as: {full_model_name}")
# Start MLflow run and log model
task = "llm/v1/chat"
with mlflow.start_run(run_id=run_id):
model_info = mlflow.transformers.log_model(
transformers_model=components,
name="model",
task=task,
registered_model_name=full_model_name,
metadata={
"task": task,
"pretrained_model_name": MODEL_NAME,
"databricks_model_family": "Llama3.2",
},
)
print(f"✓ Model successfully registered in Unity Catalog: {full_model_name}")
print(f"✓ MLflow model URI: {model_info.model_uri}")
print(f"✓ Model version: {model_info.registered_model_version}")
# Print deployment information
print(f"\n📦 Model Registration Complete!")
print(f"Unity Catalog Path: {full_model_name}")
print(f"Optimization: Liger Kernels + LoRA")