modules/resume_parser.py

import torch

class ResumeParser:
    def __init__(self, model, tokenizer):
        self.model = model
        self.tokenizer = tokenizer
    
    def format_data(self, txt):
            prompt = f"""<|im_start|>system
Extract and format resume: name | role | years | skills | core skills | experiences | education | certifications. 
Extract ONLY from input. NO inference or additions.
<|im_end|>
 
<|im_start|>user
{txt}
<|im_end|>
 
<|im_start|>assistant
"""
            return prompt


    def parse(self, resume):
        
        inputs = self.tokenizer(resume, return_tensors="pt", padding=False, truncation=True)
            
        with torch.no_grad():
                # Mixed precision for faster CPU inference
                with torch.cpu.amp.autocast():
                    output = self.model.generate(
                        **inputs,
                        max_new_tokens=300,
                        do_sample=False,
                        use_cache=True  # KV cache speeds up token generation
                    )
                    
        decoded = self.tokenizer.decode(output[0], skip_special_tokens=False)
        decoded = decoded.split("<|im_start|>assistant\n")[1].split("\n<|im_end|>")[0]
                
        return decoded