Update llm_model.py

llm_model.py

@@ -3,27 +3,36 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForSequen
 from log import log
 from pydantic import BaseModel
 
-model = None
-tokenizer = None
-eos_token_id = None
+_model = None
+_tokenizer = None
+_eos_token_id = None
+
+def get_model():
+    return _model
+
+def get_tokenizer():
+    return _tokenizer
+
+def get_eos_token_id():
+    return _eos_token_id
 
 class Message(BaseModel):
     user_input: str
 
 def setup_model(s_config):
-    global model, tokenizer, eos_token_id
+    global _model, _tokenizer, _eos_token_id
     try:
         log("🧠 setup_model() başladı")
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         log(f"📡 Kullanılan cihaz: {device}")
-        tokenizer = AutoTokenizer.from_pretrained(s_config.MODEL_BASE, use_fast=False)
+        _tokenizer = AutoTokenizer.from_pretrained(s_config.MODEL_BASE, use_fast=False)
         log("📦 Tokenizer yüklendi. Ana model indiriliyor...")
-        model = AutoModelForCausalLM.from_pretrained(s_config.MODEL_BASE, torch_dtype=torch.float32).to(device)
+        _model = AutoModelForCausalLM.from_pretrained(s_config.MODEL_BASE, torch_dtype=torch.float32).to(device)
         log("📦 Ana model indirildi ve yüklendi. eval() çağırılıyor...")
-        tokenizer.pad_token = tokenizer.pad_token or tokenizer.eos_token
-        model.config.pad_token_id = tokenizer.pad_token_id
-        eos_token_id = tokenizer("<|im_end|>", add_special_tokens=False)["input_ids"][0]
-        model.eval()
+        _tokenizer.pad_token = _tokenizer.pad_token or _tokenizer.eos_token
+        _model.config.pad_token_id = _tokenizer.pad_token_id
+        _eos_token_id = _tokenizer("<|im_end|>", add_special_tokens=False)["input_ids"][0]
+        _model.eval()
         log("✅ Ana model eval() çağrıldı")
         log(f"📦 Intent modeli indiriliyor: {s_config.INTENT_MODEL_ID}")
         _ = AutoTokenizer.from_pretrained(s_config.INTENT_MODEL_ID)
@@ -35,9 +44,12 @@ def setup_model(s_config):
         traceback.print_exc()
 
 async def generate_response(text, app_config):
+    model = get_model()
+    tokenizer = get_tokenizer()
+    eos_token_id = get_eos_token_id()
+
     messages = [{"role": "user", "content": text}]
     encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
-    eos_token = tokenizer("<|im_end|>", add_special_tokens=False)["input_ids"][0]
     input_ids = encodeds.to(model.device)
     attention_mask = (input_ids != tokenizer.pad_token_id).long()
 
@@ -47,7 +59,7 @@ async def generate_response(text, app_config):
             attention_mask=attention_mask,
             max_new_tokens=128,
             do_sample=app_config.USE_SAMPLING,
-            eos_token_id=eos_token,
+            eos_token_id=eos_token_id,
             pad_token_id=tokenizer.pad_token_id,
             return_dict_in_generate=True,
             output_scores=True
@@ -67,9 +79,3 @@ async def generate_response(text, app_config):
             decoded = decoded[start + len(tag):].strip()
             break
     return decoded, top_conf
-
-def get_model():
-    return model
-
-def get_tokenizer():
-    return tokenizer