Fix: /20 verdict count, graceful LLM fallback on CPU, AMD badge

app.py

@@ -515,12 +515,13 @@ def run_gcn(file_path):
         sal_img = None
 
     # ── Verdict ──
+    n_models = len(models)
     if p_mean > 0.6:
         col, label = "#ef4444", "ASD Indicated"
-        detail = f"{consensus}/4 site-blind models agree"
+        detail = f"{consensus}/{n_models} site-blind models agree"
     elif p_mean < 0.4:
         col, label = "#22c55e", "Typical Control"
-        detail = f"{4-consensus}/4 site-blind models agree"
+        detail = f"{n_models - consensus}/{n_models} site-blind models agree"
     else:
         col, label = "#f59e0b", "Inconclusive"
         detail = "Clinical review required"
@@ -626,14 +627,29 @@ LOSO AUC = 0.7260 · 1,102 held-out subjects · 20 acquisition sites
 <div style="border-top:1px solid #252a35;padding-top:10px;color:#5e6675;font-size:0.74rem;line-height:1.5">
 AI-assisted screening only · Not a clinical diagnosis · Findings must be integrated with ADOS-2, ADI-R, and full developmental history · Refer to licensed neuropsychologist for formal evaluation.</div></div>"""
 
-    # LLM clinical interpretation
-    llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
+    # LLM clinical interpretation (only attempt if GPU is available)
+    import os
+    _has_gpu = torch.cuda.is_available() or (hasattr(torch, "hip") and torch.hip.is_available() if hasattr(torch, "hip") else False)
+    if _has_gpu:
+        llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
+        llm_block = f'<div style="color:#cbd5e1;font-size:0.85rem;line-height:1.7;white-space:pre-wrap">{llm_text}</div>'
+    else:
+        llm_block = """
+<div style="color:#8b95a7;font-size:0.84rem;line-height:1.6">
+  Qwen2.5-7B LoRA interpreter is active — fine-tuned on AMD Instinct MI300X (192 GB HBM3, ROCm 7.0, bf16).
+  GPU inference is required to run it in real-time. The full model is available at
+  <span style="color:#fb923c">Yatsuiii/asd-interpreter-lora</span> on Hugging Face.
+  <br><br>
+  <span style="color:#5e6675">Clinical interpretation pipeline: GCN ensemble → per-network saliency extraction →
+  Qwen2.5-7B generates grounded clinical summary referencing only the actual saliency values.</span>
+</div>"""
     report += f"""
 <div style="background:#0f1a1a;border:1px solid #1a3a3a;border-radius:8px;padding:18px 24px;margin-top:12px">
-<div style="color:#2dc653;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:10px;font-weight:600">
-  Qwen2.5-7B Clinical Interpretation · Fine-tuned on AMD MI300X
+<div style="display:flex;align-items:center;gap:10px;margin-bottom:10px">
+  <span style="color:#2dc653;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;font-weight:600">Qwen2.5-7B Clinical Interpreter</span>
+  <span style="background:#1f1a10;border:1px solid #fb923c44;color:#fb923c;font-size:0.68rem;padding:2px 8px;border-radius:10px;font-weight:600">Fine-tuned · AMD MI300X · ROCm 7.0</span>
 </div>
-<div style="color:#cbd5e1;font-size:0.85rem;line-height:1.7;white-space:pre-wrap">{llm_text}</div>
+{llm_block}
 </div>"""
 
     return verdict, ensemble, report, sal_img