feat: ZeroGPU demo (spaces.GPU + CUDA torch)

app.py

@@ -86,9 +86,19 @@ def _forward(premise, hypothesis, keys):
 
 
 def grade(premise, hypothesis, selection):
+    # Generator: emit an immediate status line so the click has visible feedback while the (potentially
+    # slow) cold start runs — first request downloads the weights to CPU AND cold-starts a shared ZeroGPU
+    # allocation, which together can take ~30-60s. Warm requests only wait on the GPU handoff.
     if not premise.strip() or not hypothesis.strip():
-        return "Enter a document premise and a claim to check."
+        yield "Enter a document premise and a claim to check."
+        return
     keys = SELECTION.get(selection, ["en", "multi"])
+    cold = [s for s in MODELS if s["key"] in keys and s["key"] not in _loaded]
+    if cold:
+        yield ("⏳ **Waking up…** the first check downloads the model and cold-starts a shared GPU — "
+               "this can take **~30–60s**. Later checks are near-instant.")
+    else:
+        yield "⏳ Scoring on the GPU…"
     for spec in MODELS:            # download/load on CPU first, so the GPU window is inference-only
         if spec["key"] in keys:
             _ensure(spec)
@@ -105,7 +115,7 @@ def grade(premise, hypothesis, selection):
         verdict = "✅ grounded" if p >= 0.5 else "❌ not grounded"
         bar = "█" * round(p * 10) + "░" * (10 - round(p * 10))
         lines.append(f"| **{spec['name']}** | {spec['tag']} | `{bar}` **{p:.3f}** | {verdict} |")
-    return "\n".join(lines)
+    yield "\n".join(lines)
 
 
 from examples import EXAMPLES  # noqa: E402  (real, labeled rows; single source of truth, verified)
@@ -129,8 +139,9 @@ with gr.Blocks(title="Grounding demo", theme=gr.themes.Soft()) as demo:
     expected = gr.Textbox(label="Expected answer (ground truth for the selected example)",
                           interactive=False)
     btn = gr.Button("Check grounding", variant="primary")
+    gr.Markdown("<sub>⏱️ The first check cold-starts a shared GPU (~30–60s); later checks are fast.</sub>")
     out = gr.Markdown()
-    btn.click(grade, [premise, hypothesis, model_sel], out)
+    btn.click(grade, [premise, hypothesis, model_sel], out, show_progress="full")
     gr.Examples(EXAMPLES, [premise, hypothesis, expected])  # fills the boxes incl. expected answer
     gr.Markdown(
         '## About the author\n'