app.py

import torch

import torch
torch.cuda.is_available = lambda: False  # Force torch to disable CUDA

from unsloth import FastLanguageModel


# Force CPU mode
device = "cpu"

# Load the base model in CPU mode
base_model_name = "unsloth/Llama-3.2-3B-Instruct"
base_model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=base_model_name,
    max_seq_length=2048,
    dtype="float32",  # Use float32 for CPU
    load_in_4bit=False  # Disable 4-bit quantization for CPU
)
base_model.to(device)

# Apply LoRA adapters in CPU mode
from peft import PeftModel

lora_model_name = "oskaralf/lora_model"  # Replace with your LoRA model path
model = PeftModel.from_pretrained(base_model, lora_model_name)
model.to(device)

# Prepare for inference in CPU mode
FastLanguageModel.for_inference(model)

# Gradio interface
import gradio as gr

def chatbot(input_text):
    inputs = tokenizer(input_text, return_tensors="pt").to(device)
    outputs = model.generate(input_ids=inputs["input_ids"], max_new_tokens=64)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

iface = gr.Interface(fn=chatbot, inputs="text", outputs="text", title="Chatbot")
iface.launch()