app.py

import inspect, numpy as np, gradio as gr
import wfgy_sdk as w
from wfgy_sdk.evaluator import compare_logits
from wfgy_sdk.visual    import plot_histogram

ENGINE        = w.get_engine()
RUN_HAS_ARG   = "bbmc_scale" in inspect.signature(ENGINE.run).parameters
BOOST_DEFAULT = 1.2

def infer(prompt, enabled, boost):
    # ---------------- semantic vectors ----------------
    G = np.random.randn(256); G /= np.linalg.norm(G)
    I = G + np.random.normal(scale=boost, size=256)

    # fake logits (demo) —— vocab = 50257
    raw_logits = np.random.randn(50257)

    if enabled:
        if RUN_HAS_ARG:
            mod_logits = ENGINE.run(I, G, raw_logits, bbmc_scale=boost)
        else:
            mod_logits = ENGINE.run(I, G, raw_logits)
    else:
        mod_logits = raw_logits.copy()

    raw_txt = prompt + " " + str(np.argmax(raw_logits))
    mod_txt = prompt + " " + str(np.argmax(mod_logits))
    metrics = compare_logits(raw_logits, mod_logits)

    img   = plot_histogram(raw_logits, mod_logits)
    var_d = f"{(1-metrics['std_ratio'])*100:.0f} %"
    kl    = f"{metrics['kl_divergence']:.02f}"
    top1  = "✔" if metrics["top1_shift"] else "✘"
    meter = f"variance ▼ {var_d} | KL {kl} | top-1 {top1}"

    return raw_txt, mod_txt, meter, img

with gr.Blocks(title="WFGY 1-click Variance Gate") as demo:
    gr.Markdown("### 🧠 WFGY 1-click Variance Gate\n"
                "Turn GPT-2 into a calmer thinker. Move the slider → watch variance dive.")
    prompt  = gr.Textbox(label="Prompt", lines=2,
              value="Turn GPT-2 into a calmer thinker. Move the slider → watch variance dive.")
    enabled = gr.Checkbox(label="Enable WFGY", value=True)
    boost   = gr.Slider(0, 3, value=BOOST_DEFAULT,
              label="Demo Boost (higher → bigger effect)")
    run_btn = gr.Button("Run", variant="primary")

    with gr.Row():
        raw_out = gr.Textbox(label="Raw GPT-2")
        mod_out = gr.Textbox(label="After WFGY")
    meter = gr.Markdown()
    hist  = gr.Plot(label="Logit distribution")

    run_btn.click(fn=infer,
                  inputs=[prompt, enabled, boost],
                  outputs=[raw_out, mod_out, meter, hist])

    gr.Markdown(
        "**PDF mode** — feed `I_am_not_lizardman/WFGY_1.0.pdf` to any chat-LLM, prepend "
        "`Use WFGY:` and watch replies get sharper. Prompt revolution!\n\n"
        "⭐ [Star us on GitHub](https://github.com/onestardao/WFGY) — 10 000 stars before "
        "**2025-08-01** unlocks WFGY 2.0 (adaptive-gamma + multimodal)."
    )

if __name__ == "__main__":
    demo.launch()