app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
from transformers import pipeline


title = "CodeParrot Generator"
description = "This is a subspace to make code generation with CodeParrot, it is used in a larger space for model comparison."
example = [
    ["def print_hello_world():", "Sample", 8, 42],
    ["\"""import GPT2 from transformers\""" ", "Sample", 16, 42],
    ["def get_file_size(filepath):", "Sample", 8, 42]]
tokenizer = AutoTokenizer.from_pretrained("lvwerra/codeparrot")
model = AutoModelForCausalLM.from_pretrained("lvwerra/codeparrot", low_cpu_mem_usage)
    

def code_generation(gen_prompt, strategy, max_tokens, seed=42):
    set_seed(seed)
    gen_kwargs = {}
    gen_kwargs["do_sample"] = strategy == "Sample"
    gen_kwargs["max_new_tokens"] = max_tokens
    if gen_kwargs["do_sample"]:
        gen_kwargs["temperature"] = 0.2
        gen_kwargs["top_k"] = 0
        gen_kwargs["top_p"] = 0.95
    pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
    generated_text = pipe(gen_prompt, **gen_kwargs)[0]['generated_text']
    return generated_text


iface = gr.Interface(
    fn=code_generation, 
    inputs=[
        gr.Textbox(lines=10, label="Input code"),
        gr.Dropdown(choices=["Greedy", "Sample"], value="Greedy"),
        gr.inputs.Slider(
            minimum=8,
            maximum=256,
            step=1,
            default=8,
            label="Number of tokens to generate",
        ),
        gr.inputs.Slider(
            minimum=0,
            maximum=1000,
            step=1,
            default=42,
            label="Random seed to use for the generation"
        )
    ],
    outputs=gr.Textbox(label="Predicted code", lines=10),
    examples=example,
    layout="horizontal",
    theme="peach",
    description=description,
    title=title
)
iface.launch()