Update app.py

app.py

@@ -1,66 +1,72 @@
-import inspect, numpy as np, gradio as gr
-import wfgy_sdk as w
+"""
+HF Space · WFGY 1-click Variance Gate
+"""
+
+import gradio as gr, numpy as np, torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from wfgy_sdk import get_engine
 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
+MODEL  = "sshleifer/tiny-gpt2"
+tok    = AutoTokenizer.from_pretrained(MODEL)
+mdl    = AutoModelForCausalLM.from_pretrained(MODEL)
+ENGINE = get_engine()
+
+
+def run(prompt, enable, boost):
+    if not prompt.strip():
+        return gr.update(value="-", visible=True)
 
-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)
+    # raw logits
+    inputs   = tok(prompt, return_tensors="pt")
+    rawL     = mdl(**inputs).logits[0, -1].detach().cpu().float().numpy()
 
-    # fake logits (demo) —— vocab = 50257
-    raw_logits = np.random.randn(50257)
+    # demo-only fake semantic vectors
+    I = np.random.randn(256).astype(np.float32)
+    G = np.random.randn(256).astype(np.float32)
 
-    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)
+    if enable:
+        modL = ENGINE.run(
+            logits      = rawL,
+            input_vec   = I,
+            ground_vec  = G,
+            boost       = boost,
+        )
     else:
-        mod_logits = raw_logits.copy()
+        modL = rawL
 
-    raw_txt = prompt + " " + str(np.argmax(raw_logits))
-    mod_txt = prompt + " " + str(np.argmax(mod_logits))
-    metrics = compare_logits(raw_logits, mod_logits)
+    raw_txt = prompt + tok.decode(int(rawL.argmax()))
+    mod_txt = prompt + tok.decode(int(modL.argmax()))
+    m       = compare_logits(rawL, modL)
 
-    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}"
+    headline = f"variance ▼ {int(m['var_drop']*100)} %  |  KL {m['kl']:.2f}  |  top-1 {'✔' if m['top1'] else '✘'}"
+
+    return (
+        raw_txt,
+        mod_txt,
+        headline,
+    )
 
-    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")
+    gr.Markdown("## 🧠 WFGY 1-click Variance Gate\nTurn GPT-2 into a calmer thinker. Move the slider → watch variance dive.")
+
+    prompt   = gr.Textbox(label="Prompt")
+    enable   = gr.Checkbox(value=True, label="Enable WFGY")
+    boost    = gr.Slider(0.5, 3.0, value=1.0, 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)."
+        out_raw = gr.Textbox(label="Raw GPT-2")
+        out_mod = gr.Textbox(label="After WFGY")
+
+    headline = gr.Markdown("")
+
+    run_btn.click(
+        run,
+        inputs=[prompt, enable, boost],
+        outputs=[out_raw, out_mod, headline],
     )
 
-if __name__ == "__main__":
-    demo.launch()
+demo.queue(concurrency_count=2).launch()