Serialize CPU benchmark requests

app.py

@@ -201,6 +201,7 @@ _BENCH_MODELS = [
 ]
 _BENCH_CACHE = os.path.join(tempfile.gettempdir(), "llm_bench")
 _BENCH_RESULTS = []
+_BENCH_LOCK = threading.Lock()
 
 
 def _ensure_llama():
@@ -232,6 +233,11 @@ def _ensure_llama():
 
 
 def _run_bench(model_filter=None):
+    with _BENCH_LOCK:
+        return _run_bench_locked(model_filter)
+
+
+def _run_bench_locked(model_filter=None):
     global _BENCH_RESULTS
     if _BENCH_RESULTS and not model_filter: return _BENCH_RESULTS
     import subprocess, re
@@ -254,7 +260,8 @@ def _run_bench(model_filter=None):
             env["GGML_CUDA"] = "0"
             t0 = time.time()
             to = 90 if mi["size_gb"] <= 6 else 180
-            command = [bench, "-m", mp, "-t", str(nt), "-ngl", "0", "-p", "32", "-n", "64", "-o", "json"]
+            command = [bench, "-m", mp, "-t", str(nt), "-ngl", "0", "-p", "32", "-n", "64", "-r", "1", "-o", "json"]
+            _log(f"Running llama-bench for {mi['name']} on CPU with {nt} threads")
             p = subprocess.run(command, capture_output=True, text=True, timeout=to, env=env)
             wt = time.time()-t0
             diagnostics = (p.stderr or "").strip()