models/model_loader.py

import torch
from transformers import AutoModel, AutoModelForSequenceClassification, AutoProcessor, AutoConfig
from functools import lru_cache

@lru_cache(maxsize=2)
def load_embed_model(model_path: str = "nvidia/llama-nemotron-embed-vl-1b-v2"):
    """Load embedding model (cached)."""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    
    print(f"🔄 Loading embedding model on {device}...")
    
    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
    config._attn_implementation = "sdpa"
    if hasattr(config, 'llm_config'):
        config.llm_config._attn_implementation = "sdpa"
    
    # ✅ FIX: Use manual device instead of device_map="auto"
    model = AutoModel.from_pretrained(
        model_path,
        config=config,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        trust_remote_code=True,
        low_cpu_mem_usage=True,  # ✅ CPU optimization
    ).to(device).eval()
    
    print(f"✅ Embedding model loaded on {device}")
    return model, device


@lru_cache(maxsize=2)
def load_rerank_model(model_path: str = "nvidia/llama-nemotron-rerank-vl-1b-v2"):
    """Load reranking model (cached)."""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    
    print(f"🔄 Loading reranking model on {device}...")
    
    # ✅ FIX: Use manual device instead of device_map="auto"
    model = AutoModelForSequenceClassification.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        trust_remote_code=False,
        attn_implementation="eager",
    ).to(device).eval()
    
    processor = AutoProcessor.from_pretrained(
        model_path,
        trust_remote_code=True,
        max_input_tiles=6,
        use_thumbnail=True,
        rerank_max_length=2048
    )
    
    print(f"✅ Reranking model loaded on {device}")
    return model, processor, device