app.py

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

# Directory where your fine-tuned Phi-2 model and associated files are stored.
model_dir = "./phi2-qlora-finetuned"

# Directory to store offloaded model parts (for large models).
offload_dir = "./offload"

# Load the tokenizer.
tokenizer = AutoTokenizer.from_pretrained(model_dir)

# Load the base model with offloading support.
# base_model = AutoModelForCausalLM.from_pretrained(
#     model_dir,
#     device_map="auto",         # Automatically use available devices (GPU/CPU).
#     offload_folder=offload_dir # Directory to offload layers (for larger models).
# )

# CPU
base_model = AutoModelForCausalLM.from_pretrained(
    model_dir,
    device_map="cpu",  # Force CPU usage
    torch_dtype=torch.float32,  # Use float32 for CPU
    trust_remote_code=True,
    offload_folder=offload_dir # Directory to offload layers (for larger models).
)


# Load the adapter (PEFT) weights.
model = PeftModel.from_pretrained(base_model, model_dir)

def generate_response(prompt, max_new_tokens=200, temperature=0.7):
    """
    Generate a response from the fine-tuned Phi-2 model given a prompt.
    """
    # Tokenize the prompt and move tensors to the model's device.
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    # Generate output text using sampling.
    outputs = model.generate(
        **inputs,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        temperature=temperature
    )

    # Decode the generated tokens and return the response.
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

# Create a Gradio interface with example prompts.
demo = gr.Interface(
    fn=generate_response,
    inputs=[
        gr.Textbox(lines=4, label="Input Prompt"),
        gr.Slider(50, 500, value=200, label="Max New Tokens"),
        gr.Slider(0.0, 1.0, value=0.7, label="Temperature")
    ],
    outputs=gr.Textbox(label="Response"),
    title="Phi-2 Fine-tuned Chat",
    description="A Hugging Face Space app serving the fine-tuned Phi-2 model trained on OpenAssistant/oasst1 data.",
    examples=[
        ["Hello, how are you today?", 150, 0.7],
        ["Translate this sentence from English to French: I love programming.", 200, 0.8],
        ["Tell me a joke about artificial intelligence.", 180, 0.6],
        ["what is value of 2 + 2: ", 150, 0.9],
        ["Explain what about economics and how does it impact the individuals financial sector: ", 250, 0.7],
        ["Who is Randy orton?", 200, 0.8]
    ]
)

if __name__ == "__main__":
    demo.launch()