Use RAM loaded CPU model benchmark

app.py

@@ -260,7 +260,7 @@ def _run_bench_locked(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", "-r", "1", "-o", "json"]
+            command = [bench, "-m", mp, "-t", str(nt), "-ngl", "0", "-c", "256", "-p", "32", "-n", "64", "-r", "1", "--no-mmap", "-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
@@ -327,7 +327,7 @@ def _run_cpu_smoke(model_filter):
         result["phases"]["download"] = {"ok": False, "elapsed_sec": round(time.time() - t0, 2), "error": str(exc)}
         return result
 
-    command = [cli, "-m", path, "-t", str(result["threads"]), "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt", "--no-warmup"]
+    command = [cli, "-m", path, "-t", str(result["threads"]), "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt", "--no-warmup", "--no-mmap", "-fa", "0"]
     t0 = time.time()
     try:
         run = subprocess.run(command, capture_output=True, text=True, timeout=180, env=env)