import os
import threading
import uvicorn
from fastapi import FastAPI
from fastapi.responses import HTMLResponse, JSONResponse
from pydantic import BaseModel
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch
from huggingface_hub import hf_hub_download
import zipfile
from datetime import datetime
import random
# ✅ Sabitler
HF_TOKEN = os.environ.get("HF_TOKEN")
MODEL_BASE = "UcsTurkey/kanarya-750m-fixed"
FINE_TUNE_ZIP = "trained_model_000_000.zip"
FINE_TUNE_REPO = "UcsTurkey/trained-zips"
CONFIDENCE_THRESHOLD = -1.5
USE_SAMPLING = False # ✅ Sampling kapalı (test modu)
FALLBACK_ANSWERS = [
"Bu konuda maalesef bilgim yok.",
"Ne demek istediğinizi tam anlayamadım.",
"Bu soruya şu an yanıt veremiyorum."
]
def log(message):
timestamp = datetime.now().strftime("%H:%M:%S")
try:
print(f"[{timestamp}] {message}")
except UnicodeEncodeError:
safe_message = message.encode("utf-8", errors="replace").decode("utf-8", errors="ignore")
print(f"[{timestamp}] {safe_message}")
os.sys.stdout.flush()
app = FastAPI()
chat_history = []
model = None
tokenizer = None
class Message(BaseModel):
user_input: str
def detect_environment():
device = "cuda" if torch.cuda.is_available() else "cpu"
supports_bfloat16 = False
gpu_name = "Yok"
if device == "cuda":
props = torch.cuda.get_device_properties(0)
gpu_name = props.name
major, _ = torch.cuda.get_device_capability(0)
supports_bfloat16 = major >= 8
return {
"device": device,
"gpu_name": gpu_name,
"supports_bfloat16": supports_bfloat16,
"expected_config": {
"gpu": "Nvidia A100", "min_vram": "16GB", "cpu": "8 vCPU"
}
}
@app.get("/")
def health():
return {"status": "ok"}
@app.get("/status")
def status():
env = detect_environment()
return {
"device": env["device"],
"gpu": env["gpu_name"],
"supports_bfloat16": env["supports_bfloat16"],
"expected_config": env["expected_config"],
"note": "Sistem bu bilgilerle çalışıyor. bfloat16 desteklenmiyorsa performans sınırlı olabilir."
}
@app.get("/start", response_class=HTMLResponse)
def root():
return """
Fine-Tune Chat
Fine-tune Chat Test
"""
@app.post("/chat")
def chat(msg: Message):
try:
log(f"Kullanıcı mesajı alındı: {msg}")
global model, tokenizer
if model is None or tokenizer is None:
log("Hata: Model henüz yüklenmedi.")
return {"error": "Model yüklenmedi. Lütfen birkaç saniye sonra tekrar deneyin."}
user_input = msg.user_input.strip()
if not user_input:
return {"error": "Boş giriş"}
full_prompt = f"SORU: {user_input}\nCEVAP:"
log(f"Prompt: {full_prompt}")
inputs = tokenizer(full_prompt, return_tensors="pt")
inputs = {k: v.to(model.device) for k, v in inputs.items()}
log(f"Tokenizer input_ids: {inputs['input_ids']}")
log(f"input shape: {inputs['input_ids'].shape}")
with torch.no_grad():
if USE_SAMPLING:
output = model.generate(
**inputs,
max_new_tokens=100,
do_sample=True,
temperature=0.7,
top_k=50,
top_p=0.95,
return_dict_in_generate=True,
output_scores=True,
suppress_tokens=[tokenizer.pad_token_id] if tokenizer.pad_token_id else None
)
else:
output = model.generate(
**inputs,
max_new_tokens=100,
do_sample=False,
return_dict_in_generate=True,
output_scores=True,
suppress_tokens=[tokenizer.pad_token_id] if tokenizer.pad_token_id else None
)
generated_ids = output.sequences[0]
generated_text = tokenizer.decode(generated_ids, skip_special_tokens=True)
answer = generated_text[len(full_prompt):].strip()
if output.scores and len(output.scores) > 0:
first_token_logit = output.scores[0][0]
if torch.isnan(first_token_logit).any() or torch.isinf(first_token_logit).any():
log("Geçersiz logit (NaN/Inf) tespit edildi, fallback cevabı gönderiliyor.")
return {"answer": random.choice(FALLBACK_ANSWERS), "chat_history": chat_history}
top_logit_score = torch.max(first_token_logit).item()
log(f"İlk token logit skoru: {top_logit_score:.4f}")
if top_logit_score < CONFIDENCE_THRESHOLD:
fallback = random.choice(FALLBACK_ANSWERS)
log(f"Düşük güven: fallback cevabı gönderiliyor: {fallback}")
answer = fallback
chat_history.append({"user": user_input, "bot": answer})
log(f"Soru: {user_input} → Yanıt: {answer[:60]}...")
return {"answer": answer, "chat_history": chat_history}
except Exception as e:
log(f"/chat sırasında hata oluştu: {e}")
return {"error": str(e)}
def setup_model():
try:
global model, tokenizer
log("Fine-tune zip indiriliyor...")
zip_path = hf_hub_download(
repo_id=FINE_TUNE_REPO,
filename=FINE_TUNE_ZIP,
repo_type="model",
token=HF_TOKEN
)
extract_dir = "/app/extracted"
os.makedirs(extract_dir, exist_ok=True)
with zipfile.ZipFile(zip_path, "r") as zip_ref:
zip_ref.extractall(extract_dir)
log("Zip başarıyla açıldı.")
tokenizer = AutoTokenizer.from_pretrained(os.path.join(extract_dir, "output"))
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
env = detect_environment()
device = env["device"]
dtype = torch.bfloat16 if env["supports_bfloat16"] else (torch.float16 if device == "cuda" else torch.float32)
log(f"Ortam: GPU = {env['gpu_name']}, Device = {device}, bfloat16 destekleniyor mu: {env['supports_bfloat16']}")
log(f"Model {device.upper()} üzerinde {dtype} precision ile yüklenecek.")
log("Beklenen minimum sistem konfigürasyonu:")
log(f"- GPU: {env['expected_config']['gpu']}")
log(f"- GPU Bellek: {env['expected_config']['min_vram']}")
log(f"- CPU: {env['expected_config']['cpu']}")
base_model = AutoModelForCausalLM.from_pretrained(MODEL_BASE, torch_dtype=dtype).to(device)
peft_model = PeftModel.from_pretrained(base_model, os.path.join(extract_dir, "output"))
model = peft_model.model.to(device)
model.eval()
log(f"Model başarıyla yüklendi. dtype={next(model.parameters()).dtype}, device={next(model.parameters()).device}")
except Exception as e:
log(f"setup_model() sırasında hata oluştu: {e}")
def run_server():
log("Uvicorn sunucusu başlatılıyor...")
uvicorn.run(app, host="0.0.0.0", port=7860)
log("===== Application Startup =====")
threading.Thread(target=setup_model, daemon=True).start()
threading.Thread(target=run_server, daemon=True).start()
log("Model yükleniyor, istekler ve API sunucusu hazırlanıyor...")
while True:
try:
import time
time.sleep(60)
except Exception as e:
log(f"Ana bekleme döngüsünde hata: {e}")