import os import threading import torch from functools import lru_cache from transformers import pipeline from gliner import GLiNER from sentence_transformers import SentenceTransformer DEVICE = "cuda" if torch.cuda.is_available() else "cpu" DTYPE = ( torch.float16 if DEVICE == "cuda" else torch.float32 ) NER_MODEL = os.getenv( "NER_MODEL", "urchade/gliner_multi-v2.1" ) SENTIMENT_MODEL = os.getenv( "SENTIMENT_MODEL", "nahiar/sentiment-analysis-v2" ) SARCASM_MODEL = os.getenv( "SARCASM_MODEL", # "helinivan/multilingual-sarcasm-detector" # "jonathanjordan21/mdeberta-v3-base-indonesian-sarcasm" "jonathanjordan21/mdeberta-id-sarcasm-robust" ) EMBEDDING_MODEL = os.getenv( "EMBEDDING_MODEL", "sentence-transformers/paraphrase-multilingual-mpnet-base-v2" ) ABSA_MODEL = os.getenv( "ABSA_MODEL", "damasukma/indobert-absa" ) class ModelRegistry: """Lazy-loaded, process-local model registry.""" def __init__(self): self._models = {} self._lock = threading.RLock() def get(self, name, loader): if name not in self._models: with self._lock: if name not in self._models: self._models[name] = loader() return self._models[name] def ner(self): return self.get( "ner", lambda: GLiNER.from_pretrained(NER_MODEL) ) def sentiment(self): return self.get( "sentiment", lambda: pipeline( "text-classification", model=SENTIMENT_MODEL, device=0 if DEVICE == "cuda" else -1, torch_dtype=DTYPE, truncation=True ) ) def sarcasm(self): return self.get( "sarcasm", lambda: pipeline( "text-classification", model=SARCASM_MODEL, device=0 if DEVICE == "cuda" else -1, torch_dtype=DTYPE, truncation=True ) ) def embedding(self): return self.get( "embedding", lambda: SentenceTransformer( EMBEDDING_MODEL, device=DEVICE ) ) def absa(self): return self.get( "absa", lambda : pipeline( "text-classification", model=ABSA_MODEL, device=0 if DEVICE == "cuda" else -1, torch_dtype=DTYPE, top_k=None ) ) models = ModelRegistry()