scripts/train_n8n_sft.py

# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "transformers>=4.45.0",
#     "trl>=0.12.0",
#     "peft>=0.13.0",
#     "datasets>=3.0.0",
#     "accelerate>=1.0.0",
#     "bitsandbytes>=0.44.0",
#     "huggingface_hub>=0.26.0",
#     "torch>=2.4.0",
#     "einops>=0.8.0",
#     "sentencepiece>=0.2.0",
# ]
# [tool.uv]
# index-strategy = "unsafe-best-match"
# extra-index-url = ["https://download.pytorch.org/whl/cu124"]
# ///
"""
Script d'entraînement SFT pour le modèle n8n Expert.

Usage sur HuggingFace Jobs:
    hf jobs uv run \
        --script train_n8n_sft.py \
        --flavor h100x1 \
        --name n8n-expert-sft \
        --timeout 24h

Variables d'environnement requises:
    - HF_TOKEN: Token HuggingFace avec accès en écriture
    - WANDB_API_KEY: (optionnel) Pour le tracking W&B
"""

import os
import json
import torch
from datasets import Dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer, SFTConfig
from huggingface_hub import login, hf_hub_download

# ============================================================================
# CONFIGURATION
# ============================================================================

# Modele de base
MODEL_NAME = os.environ.get("BASE_MODEL", "Qwen/Qwen2.5-14B-Instruct")

# Dataset
DATASET_REPO = "stmasson/n8n-agentic-multitask"
TRAIN_FILE = "data/multitask_large/train.jsonl"
VAL_FILE = "data/multitask_large/val.jsonl"

# Output
OUTPUT_DIR = "./n8n-expert-sft"
HF_REPO = os.environ.get("HF_REPO", "stmasson/n8n-expert-14b-sft")

# Hyperparametres
NUM_EPOCHS = int(os.environ.get("NUM_EPOCHS", "3"))
BATCH_SIZE = int(os.environ.get("BATCH_SIZE", "2"))
GRAD_ACCUM = int(os.environ.get("GRAD_ACCUM", "8"))
LEARNING_RATE = float(os.environ.get("LEARNING_RATE", "2e-5"))
MAX_SEQ_LENGTH = int(os.environ.get("MAX_SEQ_LENGTH", "8192"))

# LoRA
LORA_R = int(os.environ.get("LORA_R", "64"))
LORA_ALPHA = int(os.environ.get("LORA_ALPHA", "128"))
LORA_DROPOUT = float(os.environ.get("LORA_DROPOUT", "0.05"))

# Quantization (pour economiser la VRAM) - 4-bit par defaut pour H100
USE_4BIT = os.environ.get("USE_4BIT", "true").lower() == "true"

# ============================================================================
# AUTHENTIFICATION
# ============================================================================

print("=" * 60)
print("ENTRAINEMENT SFT - N8N EXPERT")
print("=" * 60)

hf_token = os.environ.get("HF_TOKEN")
if hf_token:
    login(token=hf_token)
    print("Authentifie sur HuggingFace")
else:
    print("Warning: HF_TOKEN non defini, push desactive")

# Desactive wandb pour eviter les conflits de dependances
report_to = "none"
print("Tracking desactive (pas de wandb)")

# ============================================================================
# CHARGEMENT DU MODELE
# ============================================================================

print(f"\nChargement du modele: {MODEL_NAME}")

# Configuration quantization si necessaire
if USE_4BIT:
    print("Mode 4-bit active")
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=True,
    )
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        quantization_config=bnb_config,
        device_map="auto",
        trust_remote_code=True,
    )
    model = prepare_model_for_kbit_training(model)
else:
    print("Mode bfloat16")
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        torch_dtype=torch.bfloat16,
        attn_implementation="sdpa",  # SDPA is built into PyTorch, no extra install needed
        device_map="auto",
        trust_remote_code=True,
    )

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

print(f"Modele charge: {model.config.num_hidden_layers} layers, {model.config.hidden_size} hidden size")

# ============================================================================
# CONFIGURATION LORA
# ============================================================================

print(f"\nConfiguration LoRA: r={LORA_R}, alpha={LORA_ALPHA}")

lora_config = LoraConfig(
    r=LORA_R,
    lora_alpha=LORA_ALPHA,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
    lora_dropout=LORA_DROPOUT,
    bias="none",
    task_type="CAUSAL_LM"
)

# ============================================================================
# CHARGEMENT DU DATASET (FIX: chargement direct JSON pour eviter les conflits de schema)
# ============================================================================

print(f"\nChargement du dataset: {DATASET_REPO}")

def load_jsonl_dataset(repo_id: str, filename: str) -> Dataset:
    """
    Charge un dataset JSONL directement en ne gardant que la colonne 'messages'.
    Evite les problemes de schema avec les colonnes struct comme 'nodes_used'.
    """
    # Telecharger le fichier
    local_path = hf_hub_download(
        repo_id=repo_id,
        filename=filename,
        repo_type="dataset"
    )

    # Lire le JSONL et extraire uniquement 'messages'
    messages_list = []
    with open(local_path, 'r', encoding='utf-8') as f:
        for line in f:
            data = json.loads(line)
            messages_list.append({"messages": data["messages"]})

    # Creer le Dataset
    return Dataset.from_list(messages_list)

# Charger train et validation
train_dataset = load_jsonl_dataset(DATASET_REPO, TRAIN_FILE)
val_dataset = load_jsonl_dataset(DATASET_REPO, VAL_FILE)

print(f"Train: {len(train_dataset)} exemples")
print(f"Validation: {len(val_dataset)} exemples")

# Fonction de formatage
def format_example(example):
    """Formate les messages en texte pour l'entrainement"""
    messages = example["messages"]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=False
    )
    return {"text": text}

# Appliquer le formatage
print("Formatage des donnees...")
train_dataset = train_dataset.map(format_example, remove_columns=train_dataset.column_names)
val_dataset = val_dataset.map(format_example, remove_columns=val_dataset.column_names)

# Afficher un exemple
print("\nExemple de donnees formatees:")
print(train_dataset[0]["text"][:500] + "...")

# ============================================================================
# CONFIGURATION D'ENTRAINEMENT
# ============================================================================

print(f"\nConfiguration d'entrainement:")
print(f"  - Epochs: {NUM_EPOCHS}")
print(f"  - Batch size: {BATCH_SIZE}")
print(f"  - Gradient accumulation: {GRAD_ACCUM}")
print(f"  - Effective batch size: {BATCH_SIZE * GRAD_ACCUM}")
print(f"  - Learning rate: {LEARNING_RATE}")
print(f"  - Max sequence length: {MAX_SEQ_LENGTH}")

training_args = SFTConfig(
    output_dir=OUTPUT_DIR,
    num_train_epochs=NUM_EPOCHS,
    per_device_train_batch_size=BATCH_SIZE,
    per_device_eval_batch_size=BATCH_SIZE,
    gradient_accumulation_steps=GRAD_ACCUM,
    learning_rate=LEARNING_RATE,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    weight_decay=0.01,
    bf16=True,
    tf32=True,
    logging_steps=10,
    save_strategy="steps",
    save_steps=500,
    save_total_limit=3,
    eval_strategy="steps",
    eval_steps=500,
    max_length=MAX_SEQ_LENGTH,  # renamed from max_seq_length in TRL 0.12+
    packing=False,  # Disabled: packing requires flash attention for proper cross-attention masking
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    dataset_text_field="text",
    report_to=report_to,
    run_name="n8n-expert-sft",
    hub_model_id=HF_REPO if hf_token else None,
    push_to_hub=bool(hf_token),
    hub_strategy="checkpoint",
)

# ============================================================================
# ENTRAINEMENT
# ============================================================================

print("\nInitialisation du trainer...")

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=val_dataset,
    peft_config=lora_config,
    processing_class=tokenizer,  # renamed from tokenizer in TRL 0.12+
)

# Afficher les parametres entrainables
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
total_params = sum(p.numel() for p in model.parameters())
print(f"\nParametres entrainables: {trainable_params:,} / {total_params:,} ({100 * trainable_params / total_params:.2f}%)")

print("\n" + "=" * 60)
print("DEMARRAGE DE L'ENTRAINEMENT")
print("=" * 60)

trainer.train()

# ============================================================================
# SAUVEGARDE
# ============================================================================

print("\nSauvegarde du modele...")
trainer.save_model(f"{OUTPUT_DIR}/final")

if hf_token:
    print(f"Push vers {HF_REPO}...")
    trainer.push_to_hub()
    print(f"Modele disponible sur: https://huggingface.co/{HF_REPO}")

print("\n" + "=" * 60)
print("ENTRAINEMENT TERMINE")
print("=" * 60)