import json import os import shutil import requests import gradio as gr from huggingface_hub import Repository from text_generation import Client from share_btn import community_icon_html, loading_icon_html, share_js, share_btn_css HF_TOKEN = os.environ.get("HF_TOKEN", None) API_URL = os.environ.get("API_URL") FIM_PREFIX = "" FIM_MIDDLE = "" FIM_SUFFIX = "" FIM_INDICATOR = "" FORMATS = """## Model formats The model is pretrained on code and is formatted with special tokens in addition to the pure code data, such as prefixes specifying the source of the file or tokens separating code from a commit message. Use these templates to explore the model's capacities: ### 1. Prefixes 🏷️ For pure code files, use any combination of the following prefixes: ``` REPONAMEFILENAMESTARS\ncode<|endoftext|> ``` STARS can be one of: 0, 1-10, 10-100, 100-1000, 1000+ ### 2. Commits 💾 The commits data is formatted as follows: ``` codetextcode<|endoftext|> ``` ### 3. Jupyter Notebooks 📓 The model is trained on Jupyter notebooks as Python scripts and structured formats like: ``` textcodeoutput ``` ### 4. Issues 🐛 We also trained on GitHub issues using the following formatting: ``` text... ``` ### 5. Fill-in-the-middle 🧩 Fill in the middle requires rearranging the model inputs. The playground handles this for you - all you need is to specify where to fill: ``` code beforecode after ``` """ 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", ], ) client = Client( API_URL, headers={"Authorization": f"Bearer {HF_TOKEN}"}, ) def generate( prompt, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0 ): 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}" stream = client.generate_stream(prompt, **generate_kwargs) if fim_mode: output = prefix else: output = prompt previous_token = "" for response in stream: if fim_mode and response.token.text == "<|endoftext|>": output += (suffix + "\n" + response.token.text) else: output += response.token.text previous_token = response.token.text yield output return output examples = [ "def print_hello_world():", 'def fibonacci(n: int) -> int:\n """ Compute the n-th Fibonacci number. """', 'from typing import List, Tuple\n\ndef sum_and_product(numbers: List[int]) -> Tuple[int, int]:\n """ Return the sum and the product of the integers in the list as a tuple. Here is the answer of the exercise"""', "class ComplexNumbers:", ] def process_example(args): for x in generate(args): pass return x css = ".generating {visibility: hidden}" monospace_css = """ #q-input textarea { font-family: monospace, 'Consolas', Courier, monospace; } """ custom_output_css = """ #q-output textarea { min-height: 800px; } """ css += share_btn_css + monospace_css + custom_output_css + ".gradio-container {color: black}" description = """

💫 StarCoder - Playground

This is a demo to generate code with StarCoder, a 15B parameter model for code generation in 86 programming languages.

""" disclaimer = """⚠️ **Intended Use**: this app and its [supporting model](https://huggingface.co/bigcode) are provided for demonstration purposes; not to serve as replacement for human expertise. For more details on the model's limitations in terms of factuality and biases, see the [model card.](hf.co/bigcode)""" 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 prompt here", label="Prompt", elem_id="q-input", ) submit = gr.Button("Generate", variant="primary") output = gr.Code(elem_id="q-output", lines=30) with gr.Accordion("Advanced settings", open=False): with gr.Row(): column_1, column_2 = gr.Column(), gr.Column() with column_1: temperature = gr.Slider( label="Temperature", value=0.2, minimum=0.0, maximum=1.0, step=0.05, interactive=True, info="Higher values produce more diverse outputs", ) max_new_tokens = gr.Slider( label="Max new tokens", value=256, minimum=0, maximum=8192, step=64, interactive=True, info="The maximum numbers of new tokens", ) with column_2: top_p = gr.Slider( label="Top-p (nucleus sampling)", value=0.90, minimum=0.0, maximum=1, step=0.05, interactive=True, info="Higher values sample more low-probability tokens", ) repetition_penalty = gr.Slider( label="Repetition penalty", value=1.2, minimum=1.0, maximum=2.0, step=0.05, interactive=True, info="Penalize repeated tokens", ) gr.Markdown(disclaimer) with gr.Group(elem_id="share-btn-container"): community_icon = gr.HTML(community_icon_html, visible=True) loading_icon = gr.HTML(loading_icon_html, visible=True) share_button = gr.Button( "Share to community", elem_id="share-btn", visible=True ) gr.Examples( examples=examples, inputs=[instruction], cache_examples=False, fn=process_example, outputs=[output], ) gr.Markdown(FORMATS) submit.click( generate, inputs=[instruction, temperature, max_new_tokens, top_p, repetition_penalty], outputs=[output], ) share_button.click(None, [], [], _js=share_js) demo.queue(concurrency_count=16).launch(debug=True)