app.py

import os
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
import gradio as gr
from peft import PeftModel
import spaces

# Define the base model ID
base_model_id = "meta-llama/Llama-2-13b-hf"

# Ensure you have the Hugging Face token set as an environment variable
huggingface_token = os.getenv('HUGGINGFACE_TOKEN')
if not huggingface_token:
    raise Exception("Hugging Face token not found. Please set it as an environment variable 'HUGGINGFACE_TOKEN'.")

# Load the base model without quantization configuration
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,  
    trust_remote_code=True,
    use_auth_token=huggingface_token  # Use the correct parameter
).to("cuda")  # Move model to CUDA

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    base_model_id, 
    add_bos_token=True, 
    trust_remote_code=True, 
    use_auth_token=huggingface_token
)

# Load the fine-tuned model and move to CUDA
ft_model = PeftModel.from_pretrained(base_model, "checkpoint-2800").to("cuda")  # Move model to CUDA

def formatting_func(job_description):
    text = f"### The job description: {job_description}\n ### The skills: "
    return text

@spaces.GPU  # Decorate the function to ensure it uses GPU
def generate_skills(job_description):
    formatted_text = formatting_func(job_description)
    model_input = tokenizer(formatted_text, return_tensors="pt").to("cuda")  # Ensure input is on CUDA

    ft_model.eval()
    with torch.no_grad():
        output_tokens = ft_model.generate(**model_input, max_new_tokens=200)[0]
        generated_text = tokenizer.decode(output_tokens, skip_special_tokens=True)
    
    # Extract the text after "### The skills:" and before "### The qualifications:"
    skills_start_index = generated_text.find("### The skills:") + len("### The skills:")
    qualifications_start_index = generated_text.find("### The qualifications:")
    
    if qualifications_start_index != -1:
        skills_text = generated_text[skills_start_index:qualifications_start_index].strip()
    else:
        skills_text = generated_text[skills_start_index:].strip()

    # Clear CUDA memory
    torch.cuda.empty_cache()

    return skills_text

# Define the Gradio interface
inputs = gr.Textbox(lines=10, label="Job description:", placeholder="Enter or paste the job description here...")
outputs = gr.Textbox(label="Required skills:", placeholder="The required skills will be displayed here...")

gr.Interface(fn=generate_skills, inputs=inputs, outputs=outputs, title="Job Skills Analysis",
             description="Paste the job description in the text box below and the model will show the required skills for candidates.").launch(share=True)