app.py

import gradio as gr
import os
from text_generation import Client


FIM_PREFIX = "<fim_prefix>"
FIM_MIDDLE = "<fim_middle>"
FIM_SUFFIX = "<fim_suffix>"
FIM_INDICATOR = "<FILL_HERE>"

API_URL_BASE ="https://api-inference.huggingface.co/models/bigcode/starcoderbase"
HF_TOKEN = os.environ.get("HF_TOKEN", None)

theme = gr.themes.Monochrome(
    primary_hue="indigo",
    secondary_hue="blue",
    neutral_hue="slate",
    radius_size=gr.themes.sizes.radius_sm,
    font=[
        gr.themes.GoogleFont("Open Sans"),
        "ui-sans-serif",
        "system-ui",
        "sans-serif",
    ],
)
css = ".generating {visibility: hidden}"

monospace_css = """
#q-input textarea {
    font-family: monospace, 'Consolas', Courier, monospace;
}
"""


css +=  monospace_css + ".gradio-container {color: black}"

description = """
<div style="text-align: center;">
    <p>This is a demo to generate code with <a href="https://huggingface.co/bigcode/starcoder" style='color: #e6b800;'>StarCoder</a></p>
</div>
"""

examples = [
    "X_train, y_train, X_test, y_test = train_test_split(X, y, test_size=0.1)\n\n# Train a logistic regression model, predict the labels on the test set and compute the accuracy score",
    "// Returns every other value in the array as a new array.\nfunction everyOther(arr) {",
    "def alternating(list1, list2):\n   results = []\n   for i in range(min(len(list1), len(list2))):\n       results.append(list1[i])\n       results.append(list2[i])\n   if len(list1) > len(list2):\n       <FILL_HERE>\n   else:\n       results.extend(list2[i+1:])\n   return results",
]

client_base = Client(
    API_URL_BASE, headers={"Authorization": f"Bearer {HF_TOKEN}"},
)

def generate(
    prompt, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0, version="StarCoder",
):

    temperature = float(temperature)
    if temperature < 1e-2:
        temperature = 1e-2
    top_p = float(top_p)
    fim_mode = False

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        seed=42,
    )

    if FIM_INDICATOR in prompt:
        fim_mode = True
        try:
            prefix, suffix = prompt.split(FIM_INDICATOR)
        except:
            raise ValueError(f"Only one {FIM_INDICATOR} allowed in prompt!")
        prompt = f"{FIM_PREFIX}{prefix}{FIM_SUFFIX}{suffix}{FIM_MIDDLE}"

    if version == "StarCoder":
        stream = client.generate_stream(prompt, **generate_kwargs)
    else:
        stream = client_base.generate_stream(prompt, **generate_kwargs)

    if fim_mode:
        output = prefix
    else:
        output = prompt

    previous_token = ""
    for response in stream:
        if response.token.text == "<|endoftext|>":
            if fim_mode:
                output += suffix
            else:
                return output
        else:
            output += response.token.text
        previous_token = response.token.text
        yield output
    return output
def process_example(args):
    for x in generate(args):
        pass
    return x

with gr.Blocks(theme=theme, analytics_enabled=False, css=css) as demo:
    with gr.Column():
        gr.Markdown(description)
        with gr.Row():
            with gr.Column():
                instruction = gr.Textbox(
                    placeholder="Enter your code here",
                    label="Code",
                    elem_id="q-input",
                )
                submit = gr.Button("Generate", variant="primary")
                output = gr.Code(elem_id="q-output", lines=30)
        gr.Examples(
                examples=examples,
                inputs=[instruction],
                cache_examples=False,
                fn=process_example,
                outputs=[output],
            )

demo.queue(concurrency_count=16).launch(debug=True)