app.py1

import gradio as gr
import subprocess
import os
from huggingface_hub import hf_hub_download

# MODEL REGISTRY
MODELS = {
    "Bonsai 1.7B (248MB)": {"repo": "prism-ml/Bonsai-1.7B-gguf", "file": "Bonsai-1.7B-v1.0.gguf"},
    "Bonsai 4B (572MB)": {"repo": "prism-ml/Bonsai-4B-gguf", "file": "Bonsai-4B-v1.0.gguf"},
    "Bonsai 8B (1.15GB)": {"repo": "prism-ml/Bonsai-8B-gguf", "file": "Bonsai-8B-v1.0.gguf"}
}

def chat(message, history, system_prompt, model_choice, temp):
    # 1. Download/Path Setup
    config = MODELS[model_choice]
    model_path = hf_hub_download(repo_id=config["repo"], filename=config["file"])
    
    # 2. Build the Prompt (Standard Format)
    prompt = f"System: {system_prompt}\n"
    for human, assistant in history:
        prompt += f"User: {human}\nAssistant: {assistant}\n"
    prompt += f"User: {message}\nAssistant:"

    # 3. Subprocess Call (The Old Way)
    # Using the binary we moved in the Dockerfile
    cmd = [
        "./llama-cli", "-m", model_path,
        "-p", prompt,
        "-n", "512",
        "--threads", "4",
        "--temp", str(temp),
        "--repeat_penalty", "1.1",
        "--no-display-prompt"
    ]

    try:
        process = subprocess.Popen(cmd, stdout=subprocess.PIPE, text=True, bufsize=1)
        response = ""
        for line in process.stdout:
            response += line
            yield response
    except Exception as e:
        yield f"Inference Error: {str(e)}"

# GRADIO UI
with gr.Blocks(theme=gr.themes.Default()) as demo:
    gr.Markdown("# 🌿 Bonsai 1-Bit AI Sandbox")
    
    with gr.Row():
        with gr.Column(scale=1):
            model_select = gr.Dropdown(list(MODELS.keys()), value="Bonsai 1.7B (248MB)", label="Model Selector")
            sys_input = gr.Textbox(
                value="You are a helpful AI assistant. Be concise and prioritize logic.",
                label="System Prompt", lines=4
            )
            temp_slider = gr.Slider(0.1, 1.0, value=0.7, label="Temperature")
            
            gr.Markdown("### Standard Benchmarks")
            btn_math = gr.Button("Logic: Math Problem")
            btn_code = gr.Button("Code: C Implementation")

        with gr.Column(scale=3):
            chatbot = gr.ChatInterface(
                fn=chat,
                additional_inputs=[sys_input, model_select, temp_slider]
            )

    # Simple Test Triggers
    btn_math.click(fn=lambda: "Explain why 1+1=2 logically.", outputs=None)
    btn_code.click(fn=lambda: "Write a C function to reverse a string in-place.", outputs=None)

demo.queue().launch(server_name="0.0.0.0", server_port=7860)