Resolve continous aborted GPU tasks

models.py

@@ -43,7 +43,6 @@ def is_model_loaded(model_name: str) -> bool:
                 config.global_medical_models[model_name] is not None and
                 _model_loading_states.get(model_name) == "loaded")
 
-@spaces.GPU(max_duration=120)
 def initialize_medical_model(model_name: str):
     """Initialize medical model (MedSwin) - download on demand"""
     if model_name not in config.global_medical_models or config.global_medical_models[model_name] is None:
@@ -73,7 +72,6 @@ def initialize_medical_model(model_name: str):
             set_model_loading_state(model_name, "loaded")
     return config.global_medical_models[model_name], config.global_medical_tokenizers[model_name]
 
-@spaces.GPU(max_duration=120)
 def initialize_tts_model():
     """Initialize TTS model for text-to-speech"""
     if not TTS_AVAILABLE:
@@ -90,7 +88,6 @@ def initialize_tts_model():
             config.global_tts_model = None
     return config.global_tts_model
 
-@spaces.GPU(max_duration=120)
 def get_or_create_embed_model():
     """Reuse embedding model to avoid reloading weights each request"""
     if config.global_embed_model is None:
@@ -98,7 +95,6 @@ def get_or_create_embed_model():
         config.global_embed_model = HuggingFaceEmbedding(model_name=config.EMBEDDING_MODEL, token=config.HF_TOKEN)
     return config.global_embed_model
 
-@spaces.GPU(max_duration=120)
 def get_llm_for_rag(temperature=0.7, max_new_tokens=256, top_p=0.95, top_k=50):
     """Get LLM for RAG indexing (uses medical model)"""
     medical_model_obj, medical_tokenizer = initialize_medical_model(config.DEFAULT_MEDICAL_MODEL)