Update fine_tune_inference_test_mistral.py

fine_tune_inference_test_mistral.py

@@ -14,6 +14,7 @@ MODEL_BASE = "mistralai/Mistral-7B-Instruct-v0.2"
 FINE_TUNE_ZIP = "trained_model_000_009.zip"
 FINE_TUNE_REPO = "UcsTurkey/trained-zips"
 USE_SAMPLING = False
+USE_FINE_TUNE = False  # ✅ Ana model ile test için False yap
 CONFIDENCE_THRESHOLD = -1.5
 FALLBACK_ANSWERS = [
     "Bu konuda maalesef bilgim yok.",
@@ -116,27 +117,33 @@ def detect_env():
 def setup_model():
     global model, tokenizer
     try:
-        log("📦 Zip indiriliyor...")
-        zip_path = hf_hub_download(
-            repo_id=FINE_TUNE_REPO,
-            filename=FINE_TUNE_ZIP,
-            repo_type="model",
-            token=HF_TOKEN
-        )
         extract_path = "/app/extracted"
-        os.makedirs(extract_path, exist_ok=True)
-        with zipfile.ZipFile(zip_path, "r") as zip_ref:
-            zip_ref.extractall(extract_path)
-        tokenizer = AutoTokenizer.from_pretrained(os.path.join(extract_path, "output"))
-        if tokenizer.pad_token is None:
-            tokenizer.pad_token = tokenizer.eos_token
-
         device, supports_bf16 = detect_env()
         dtype = torch.bfloat16 if supports_bf16 else torch.float32
         log(f"🧠 Ortam: {device.upper()}, dtype: {dtype}")
-        base = AutoModelForCausalLM.from_pretrained(MODEL_BASE, torch_dtype=dtype).to(device)
-        peft = PeftModel.from_pretrained(base, os.path.join(extract_path, "output"))
-        model = peft.model.to(device)
+
+        if USE_FINE_TUNE:
+            log("📦 Zip indiriliyor...")
+            zip_path = hf_hub_download(
+                repo_id=FINE_TUNE_REPO,
+                filename=FINE_TUNE_ZIP,
+                repo_type="model",
+                token=HF_TOKEN
+            )
+            os.makedirs(extract_path, exist_ok=True)
+            with zipfile.ZipFile(zip_path, "r") as zip_ref:
+                zip_ref.extractall(extract_path)
+            tokenizer = AutoTokenizer.from_pretrained(os.path.join(extract_path, "output"))
+            if tokenizer.pad_token is None:
+                tokenizer.pad_token = tokenizer.eos_token
+            base = AutoModelForCausalLM.from_pretrained(MODEL_BASE, torch_dtype=dtype).to(device)
+            peft = PeftModel.from_pretrained(base, os.path.join(extract_path, "output"))
+            model = peft.model.to(device)
+        else:
+            log("🧪 Sadece ana model yüklenecek.")
+            tokenizer = AutoTokenizer.from_pretrained(MODEL_BASE)
+            model = AutoModelForCausalLM.from_pretrained(MODEL_BASE, torch_dtype=dtype).to(device)
+
         model.eval()
         log("✅ Model yüklendi.")
     except Exception as e: