app.py

from openai import OpenAI
import gradio as gr
import os
import json
import functools
import random
import datetime
from transformers import AutoTokenizer

reflection_tokenizer = AutoTokenizer.from_pretrained("mattshumer/Reflection-Llama-3.1-70B")

api_key = os.environ.get('FEATHERLESS_API_KEY')
client = OpenAI(
    base_url="https://api.featherless.ai/v1",
    api_key=api_key
)

SYSTEM_PROMPT = """You are a helpful AI assistant. Your responses should be informative, polite, and tailored to the user's needs. 

Please think carefully."""

def respond(message, history, model):
    history_openai_format = [{"role": "system", "content": SYSTEM_PROMPT}]
    for human, assistant in history:
        history_openai_format.append({"role": "user", "content": human })
        history_openai_format.append({"role": "assistant", "content":assistant})
    history_openai_format.append({"role": "user", "content": message})


    if model == "mattshumer/Reflection-Llama-3.1-70B":
        # chat/completions not working for this model;
        # apply chat template locally
        response = client.completions.create(
            model=model,
            prompt=reflection_tokenizer.apply_chat_template(history_openai_format, tokenize=False),
            temperature=1.0,
            stream=True,
            max_tokens=2000,
            extra_headers={
                'HTTP-Referer': 'https://huggingface.co/spaces/featherless-ai/try-this-model',
                'X-Title': "HF's missing inference widget"
            }
        )

        # debugger_ran = False
        partial_message = ""
        for chunk in response:
            # if not debugger_ran:
            #     import code
            #     code.InteractiveConsole(locals=locals()).interact()
            #     debugger_ran = True
            if chunk.choices[0].text is not None:
                partial_message = partial_message + chunk.choices[0].text
                prefix_to_strip = "<|start_header_id|>assistant<|end_header_id|>\n\n"
                yield partial_message[len(prefix_to_strip):]
    else:
        response = client.chat.completions.create(
            model=model,
            messages= history_openai_format,
            temperature=1.0,
            stream=True,
            max_tokens=2000,
            extra_headers={
                'HTTP-Referer': 'https://huggingface.co/spaces/featherless-ai/try-this-model',
                'X-Title': "HF's missing inference widget"
            }
        )

        partial_message = ""
        for chunk in response:
            if chunk.choices[0].delta.content is not None:
                  partial_message = partial_message + chunk.choices[0].delta.content
                  yield partial_message

logo = open('./logo.svg').read()

with open('./model-cache.json', 'r') as f_model_cache:
    model_cache = json.load(f_model_cache)


model_class_filter = {
    "mistral-v02-7b-std-lc": True,
    "llama3-8b-8k": True,
    "llama2-solar-10b7-4k": True,
    "mistral-nemo-12b-lc": True,
    "llama2-13b-4k": True,
    "llama3-15b-8k": True,

    "qwen2-32b-lc":False,
    "llama3-70b-8k":False,
    "qwen2-72b-lc":False,
    "mixtral-8x22b-lc":False,
    "llama3-405b-lc":False,
}

def build_model_choices():
    all_choices = []
    for model_class in model_cache:
        if model_class not in model_class_filter:
            print(f"Warning: new model class {model_class}. Treating as blacklisted")
            continue

        if not model_class_filter[model_class]:
            continue
        all_choices += [ (f"{model_id} ({model_class})", model_id) for model_id in model_cache[model_class] ]

    # and add one more ...
    model_class = "llama3-70b-8k"
    model_id = "mattshumer/Reflection-Llama-3.1-70B"
    all_choices += [(f"{model_id} ({model_class})", model_id)]

    return all_choices

model_choices = build_model_choices()

def initial_model(referer=None):
    return "mattshumer/Reflection-Llama-3.1-70B"

    # if referer == 'http://127.0.0.1:7860/':
    #     return 'Sao10K/Venomia-1.1-m7'

    # if referer and referer.startswith("https://huggingface.co/"):
    #     possible_model = referer[23:]
    #     full_model_list = functools.reduce(lambda x,y: x+y, model_cache.values(), [])
    #     model_is_supported = possible_model in full_model_list
    #     if model_is_supported:
    #         return possible_model

    # # let's use a random but different model each day.
    # key=os.environ.get('RANDOM_SEED', 'kcOtfNHA+e')
    # o = random.Random(f"{key}-{datetime.date.today().strftime('%Y-%m-%d')}")
    # return o.choice(model_choices)[1]

title_text="HuggingFace's missing inference widget"
css = """
.logo-mark { fill: #ffe184; }

/* from https://github.com/gradio-app/gradio/issues/4001
 * necessary as putting ChatInterface in gr.Blocks changes behaviour
 */

.contain { display: flex; flex-direction: column; }
.gradio-container { height: 100vh !important; }
#component-0 { height: 100%; }
#chatbot { flex-grow: 1; overflow: auto;}
"""
with gr.Blocks(title_text, css=css) as demo:
    gr.HTML("""
        <h1 align="center">HuggingFace's missing inference widget</h1>
        <h2 align="center">
            Please select your model from the list 👇
        </h2>
    """)

    # hidden_state = gr.State(value=initial_model)
    with gr.Row():
        model_selector = gr.Dropdown(
            label="Select your Model",
            choices=build_model_choices(),
            value=initial_model,
            # value=hidden_state,
            scale=4
        )
        gr.Button(
            value="Visit Model Card ↗️",
            scale=1
        ).click(
            inputs=[model_selector],
            js="(model_selection) => { window.open(`https://huggingface.co/${model_selection}`, '_blank') }",
            fn=None,
        )

    gr.ChatInterface(
        respond,
        additional_inputs=[model_selector],
        head=""",
        <script>console.log("Hello from gradio!")</script>
        """,
    )
    gr.HTML(f"""
        <p align="center">
            Inference by <a href="https://featherless.ai">{logo}</a>
        </p>
    """)
    def update_initial_model_choice(request: gr.Request):
        return initial_model(request.headers.get('referer'))

    demo.load(update_initial_model_choice, outputs=model_selector)

demo.launch(concurrency_count=6)