app.py

import gradio as gr
from llama_cpp import Llama

model = Llama(model_path="CodeLlama_7B_nlp_pp_q8_0.gguf")

# These phrases/tokens indicate the start of a pass. For demonstration purposes, it's
# safe to assume that these should not be encountered in the output and represent a hallucination.
stop = ["@NODES", "@CODE", "@DECL"]

def generate(input_text):
    output = model(input_text, max_tokens=128, stop=stop, echo=True)
    return output['choices'][0]['text']

input_text = gr.inputs.Textbox(lines= 10, label="Enter your code to autocomplete")
output_text = gr.outputs.Textbox(label="Output code")

description = "Code generation for NLP++ with CodeLlama"

examples = [
    '# Find concept named parent under root and print "num" val for each child attribute\n',
    'L("iter") = getconcept(findroot(), L("parent_con"));\n',
    '# Match node _noun when preceded by _noun\n'
]

gr.Interface(fn=generate, inputs=input_text, outputs=output_text, title="CodeLlama for NLP++", description=description, examples=examples).launch()