Keep local model benchmarks CPU only

app.py

@@ -29,10 +29,8 @@ if hf_token:
 def _gpu_startup_touch():
     return "ok"
 
-try:
-    _gpu_startup_touch()
-except Exception as e:
-    print(f"Warning: ZeroGPU touch failed: {e}", flush=True)
+# Keep the decorated symbol so ZeroGPU Spaces recognize the app, but do not call it.
+# The Respite API and all CPU benchmarks must not consume GPU time.
 
 
 def _log(msg):
@@ -252,8 +250,10 @@ def _run_bench(model_filter=None):
             mp = hf_hub_download(repo_id=mi["repo"], filename=mi["file"], cache_dir=_BENCH_CACHE)
             env = os.environ.copy()
             env["LD_LIBRARY_PATH"] = os.path.dirname(cli)
+            env["CUDA_VISIBLE_DEVICES"] = ""
+            env["GGML_CUDA"] = "0"
             t0 = time.time()
-            to = 300 if mi["size_gb"] <= 6 else 600
+            to = 90 if mi["size_gb"] <= 6 else 180
             command = [bench, "-m", mp, "-t", str(nt), "-ngl", "0", "-p", "32", "-n", "64", "-o", "json"]
             p = subprocess.run(command, capture_output=True, text=True, timeout=to, env=env)
             wt = time.time()-t0
@@ -265,7 +265,9 @@ def _run_bench(model_filter=None):
             row = parsed[0] if isinstance(parsed, list) and parsed else parsed
             r.update({"status":"success", "total_time_sec":round(wt,2), "prompt_tokens":32, "completion_tokens":64, "prompt_tokens_per_sec":row.get("avg_ts", 0), "generation_tokens_per_sec":row.get("avg_ts", 0), "raw":row})
         except subprocess.TimeoutExpired:
-            raise RuntimeError(f"llama-bench timed out after {to}s")
+            r["error"] = f"llama-bench timed out after {to}s on CPU"
+            results.append(r)
+            continue
         except Exception as e:
             r["error"] = str(e)
         results.append(r)