train_lora_mistral.py

import sys, os, zipfile, shutil, time, traceback, threading, uvicorn
from fastapi import FastAPI
from fastapi.responses import JSONResponse
from datetime import datetime
from datasets import load_dataset
from huggingface_hub import HfApi
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
from peft import get_peft_model, LoraConfig, TaskType
import torch

# === Sabitler ===
START_NUMBER = 0
END_NUMBER = 9
MODEL_NAME = "mistralai/Mistral-7B-Instruct-v0.2"
TOKENIZED_DATASET_ID = "UcsTurkey/turkish-general-culture-tokenized"
ZIP_UPLOAD_REPO = "UcsTurkey/trained-zips"
HF_TOKEN = os.environ.get("HF_TOKEN")
BATCH_SIZE = 1
EPOCHS = 2
MAX_LENGTH = 2048
OUTPUT_DIR = "/data/output"
ZIP_FOLDER = "/data/zip_temp"
zip_name = f"trained_model_{START_NUMBER:03d}_{END_NUMBER:03d}.zip"
ZIP_PATH = os.path.join(ZIP_FOLDER, zip_name)

# === Health check
app = FastAPI()

@app.get("/")
def health():
    return JSONResponse(content={"status": "ok"})

def run_health_server():
    uvicorn.run(app, host="0.0.0.0", port=7860)

threading.Thread(target=run_health_server, daemon=True).start()

# === Log
def log(message):
    timestamp = datetime.now().strftime("%H:%M:%S")
    print(f"[{timestamp}] {message}")
    sys.stdout.flush()

# === Eğitim Başlıyor
log("🛠️ Ortam hazırlanıyor...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

log("🧠 Model indiriliyor...")
base_model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16)
base_model.config.pad_token_id = tokenizer.pad_token_id

log("🎯 LoRA adapter uygulanıyor...")
peft_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=64, lora_alpha=16, lora_dropout=0.1,
    bias="none", fan_in_fan_out=False
)
model = get_peft_model(base_model, peft_config)
model.print_trainable_parameters()

log("📦 Parquet dosyaları listeleniyor...")
api = HfApi()
files = api.list_repo_files(repo_id=TOKENIZED_DATASET_ID, repo_type="dataset", token=HF_TOKEN)
selected_files = sorted([f for f in files if f.startswith("chunk_") and f.endswith(".parquet")])[START_NUMBER:END_NUMBER+1]
if not selected_files:
    log("⚠️ Parquet bulunamadı. Eğitim iptal.")
    exit(0)

training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,
    per_device_train_batch_size=BATCH_SIZE,
    num_train_epochs=EPOCHS,
    save_strategy="epoch",
    save_total_limit=2,
    learning_rate=2e-4,
    disable_tqdm=True,
    logging_strategy="steps",
    logging_steps=10,
    report_to=[],
    bf16=True,
    fp16=False
)

for file in selected_files:
    try:
        log(f"\n📄 Yükleniyor: {file}")
        dataset = load_dataset(
            path=TOKENIZED_DATASET_ID,
            data_files={"train": file},
            split="train",
            token=HF_TOKEN
        )
        log(f"🔍 {len(dataset)} örnek")
        if len(dataset) == 0:
            continue
        trainer = Trainer(model=model, args=training_args, train_dataset=dataset)
        log("🚀 Eğitim başlıyor...")
        trainer.train()
        log("✅ Eğitim tamam.")
    except Exception as e:
        log(f"❌ Hata: {file} → {e}")
        traceback.print_exc()

# === Zip
log("📦 Model zipleniyor...")
try:
    tmp_dir = os.path.join(ZIP_FOLDER, "temp_save")
    os.makedirs(tmp_dir, exist_ok=True)
    model.save_pretrained(tmp_dir)
    tokenizer.save_pretrained(tmp_dir)

    with zipfile.ZipFile(ZIP_PATH, "w", zipfile.ZIP_DEFLATED) as zipf:
        for root, _, files in os.walk(tmp_dir):
            for file in files:
                filepath = os.path.join(root, file)
                arcname = os.path.relpath(filepath, tmp_dir)
                zipf.write(filepath, arcname=os.path.join("output", arcname))
    log(f"✅ Zip oluşturuldu: {ZIP_PATH}")
except Exception as e:
    log(f"❌ Zipleme hatası: {e}")
    traceback.print_exc()

# === Upload
try:
    log("☁️ Hugging Face'e yükleniyor...")
    api.upload_file(
        path_or_fileobj=ZIP_PATH,
        path_in_repo=zip_name,
        repo_id=ZIP_UPLOAD_REPO,
        repo_type="model",
        token=HF_TOKEN
    )
    log("✅ Upload tamam.")
except Exception as e:
    log(f"❌ Upload hatası: {e}")
    traceback.print_exc()

log("⏸️ Eğitim tamamlandı. Servis bekleme modunda...")
while True:
    time.sleep(60)