eval_pvsg/model_loader.py

"""
Model loading utilities for Qwen-VL evaluation
"""
import sys
import os
import torch
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from pathlib import Path

# Add the Qwen2-VL-Finetune src to path for imports
sys.path.append("/weka/royg/vsg_train/Qwen2-VL-Finetune/src")

from config import MODELS


class QwenVLModelLoader:
    """Load and manage Qwen-VL models for evaluation"""
    
    def __init__(self, model_variant: str, device: str = "auto"):
        self.model_variant = model_variant
        self.device = device
        self.model_config = MODELS[model_variant]
        self.model = None
        self.processor = None
        
    def load_model(self):
        """Load the specified model variant"""
        model_path = self.model_config["path"]
        
        if not os.path.exists(model_path):
            raise FileNotFoundError(f"Model path not found: {model_path}")
            
        print(f"Loading {self.model_variant} model from {model_path}")
        
        # Load processor - use base model processor for consistency
        base_model_name = "Qwen/Qwen2.5-VL-3B-Instruct"
        self.processor = AutoProcessor.from_pretrained(base_model_name)
        
        # Load fine-tuned model
        self.model = Qwen2VLForConditionalGeneration.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            device_map="auto" if self.device == "auto" else None,
            trust_remote_code=True
        )
        
        if self.device != "auto":
            self.model = self.model.to(self.device)
            
        self.model.eval()
        print(f"Successfully loaded {self.model_variant} model")
        
    def generate_response(self, messages, max_new_tokens=1024, temperature=0.1):
        """Generate response from the model"""
        if self.model is None or self.processor is None:
            raise RuntimeError("Model not loaded. Call load_model() first.")
            
        # Prepare inputs
        text = self.processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )
        
        # Process inputs
        inputs = self.processor(
            text=[text],
            images=messages[0].get("content", [{}])[0].get("image") if "image" in str(messages) else None,
            videos=messages[0].get("content", [{}])[0].get("video") if "video" in str(messages) else None,
            padding=True,
            return_tensors="pt"
        )
        
        # Move to device
        if self.device != "auto":
            inputs = {k: v.to(self.device) if isinstance(v, torch.Tensor) else v 
                     for k, v in inputs.items()}
        
        # Generate
        with torch.no_grad():
            generated_ids = self.model.generate(
                **inputs,
                max_new_tokens=max_new_tokens,
                temperature=temperature,
                do_sample=temperature > 0,
                pad_token_id=self.processor.tokenizer.eos_token_id
            )
        
        # Decode response
        generated_ids_trimmed = [
            out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
        ]
        
        response = self.processor.batch_decode(
            generated_ids_trimmed, 
            skip_special_tokens=True, 
            clean_up_tokenization_spaces=False
        )[0]
        
        return response
        
    def unload_model(self):
        """Free model memory"""
        if self.model is not None:
            del self.model
            self.model = None
        if self.processor is not None:
            del self.processor  
            self.processor = None
        torch.cuda.empty_cache()
        

def load_all_models(device="auto"):
    """Load all three model variants"""
    models = {}
    for variant in MODELS.keys():
        loader = QwenVLModelLoader(variant, device)
        loader.load_model()
        models[variant] = loader
    return models


def get_model_prompt_template(model_variant: str):
    """Get the appropriate prompt template for model variant"""
    from config import VSG_PROMPT_TEMPLATE, VSG_PROMPT_NO_ID, VSG_PROMPT_SPECIAL_TOKENS
    
    if model_variant == "with_id":
        return VSG_PROMPT_TEMPLATE
    elif model_variant == "without_id":
        return VSG_PROMPT_NO_ID
    elif model_variant == "special_tokens":
        return VSG_PROMPT_SPECIAL_TOKENS
    else:
        raise ValueError(f"Unknown model variant: {model_variant}")