File size: 6,355 Bytes
6734e84 d0e6741 2a16a45 6734e84 06fedb7 6734e84 0a5b12b d0e6741 06fedb7 d0e6741 6734e84 be00f5e 6734e84 0a5b12b 6734e84 9a3135b 06fedb7 6734e84 06fedb7 6734e84 a5cd092 b5c1d21 2a16a45 b5c1d21 cc493ed 2a16a45 cc493ed 2a16a45 cc493ed 6734e84 d0e6741 3977e82 06fedb7 d0e6741 1f2f347 430e8dc 06fedb7 1a7401c 06fedb7 430e8dc 06fedb7 430e8dc cee75e6 430e8dc 06fedb7 584aa58 06fedb7 430e8dc 6e08bcb 430e8dc 6e08bcb 430e8dc 6e08bcb 430e8dc 1a7401c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 |
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 datasets import load_dataset
from peft import PeftModel
import torch
from huggingface_hub import hf_hub_download
import zipfile
from datetime import datetime
import random
# ✅ Zamanlı log fonksiyonu (flush destekli)
def log(message):
timestamp = datetime.now().strftime("%H:%M:%S")
print(f"[{timestamp}] {message}")
os.sys.stdout.flush()
# ✅ Sabitler
HF_TOKEN = os.environ.get("HF_TOKEN")
MODEL_BASE = "UcsTurkey/kanarya-750m-fixed"
FINE_TUNE_ZIP = "trained_model_002_005.zip"
FINE_TUNE_REPO = "UcsTurkey/trained-zips"
RAG_DATA_FILE = "merged_dataset_000_100.parquet"
RAG_DATA_REPO = "UcsTurkey/turkish-general-culture-tokenized"
USE_RAG = False # ✅ RAG kullanımını opsiyonel hale getiren sabit
CONFIDENCE_THRESHOLD = -1.5 # ✅ Logit skorlarına göre eşik değeri
FALLBACK_ANSWERS = [
"Bu konuda maalesef bilgim yok.",
"Ne demek istediğinizi tam anlayamadım.",
"Bu soruya şu an yanıt veremiyorum."
]
app = FastAPI()
chat_history = []
model = None
tokenizer = None
class Message(BaseModel):
user_input: str
@app.get("/")
def health():
return {"status": "ok"}
@app.get("/start", response_class=HTMLResponse)
def root():
return """
<html>
<head><title>Fine-Tune Chat</title></head>
<body>
<h2>📘 Fine-tune Chat Test</h2>
<textarea id=\"input\" rows=\"4\" cols=\"60\" placeholder=\"Bir şeyler yaz...\"></textarea><br><br>
<button onclick=\"send()\">Gönder</button>
<pre id=\"output\"></pre>
<script>
async function send() {
const input = document.getElementById(\"input\").value;
const res = await fetch("/chat", {
method: "POST",
headers: { "Content-Type": "application/json" },
body: JSON.stringify({ user_input: input })
});
const data = await res.json();
document.getElementById("output").innerText = data.answer || data.error || "Hata oluştu.";
}
</script>
</body>
</html>
"""
@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()}
with torch.no_grad():
output = model.generate(
**inputs,
max_new_tokens=200,
do_sample=True,
temperature=0.7,
return_dict_in_generate=True,
output_scores=True
)
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.")
fallback = random.choice(FALLBACK_ANSWERS)
answer = fallback
return {"answer": answer, "chat_history": chat_history} # ilk tokenin logits
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.")
log("🔁 Tokenizer yükleniyor...")
tokenizer = AutoTokenizer.from_pretrained(os.path.join(extract_dir, "output"))
log("🧠 Base model indiriliyor...")
base_model = AutoModelForCausalLM.from_pretrained(
MODEL_BASE,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)
log("➕ LoRA adapter uygulanıyor...")
peft_model = PeftModel.from_pretrained(base_model, os.path.join(extract_dir, "output"))
model = peft_model.model
model.eval()
log("✅ Model başarıyla yüklendi.")
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)
# ✅ Uygulama başlangıcı
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}")
|