Fix CPU benchmark diagnostics

app.py

@@ -218,7 +218,7 @@ def _ensure_llama():
     bd = os.path.join(_BENCH_CACHE, "build")
     os.makedirs(bd, exist_ok=True)
     nt = _parse_cpu_count() or 8
-    subprocess.check_call(["cmake","-S",src,"-B",bd,"-DLLAMA_NATIVE=ON","-DBUILD_SHARED_LIBS=OFF"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    subprocess.check_call(["cmake","-S",src,"-B",bd,"-DLLAMA_NATIVE=ON","-DLLAMA_CUDA=OFF","-DGGML_CUDA=OFF","-DBUILD_SHARED_LIBS=OFF"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
     subprocess.check_call(["cmake","--build",bd,"-j",str(nt),"--target","llama-cli"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
     import glob as _g
     cands = _g.glob(os.path.join(bd,"**","llama-cli"), recursive=True)
@@ -244,21 +244,39 @@ 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)
-            np = 128 if mi["size_gb"]<=6 else 256
-            to = 300 if mi["size_gb"]<=6 else 600
+            np = 48 if mi["size_gb"] <= 6 else 24
+            to = 240 if mi["size_gb"] <= 6 else 600
             t0 = time.time()
-            p = subprocess.run([cli,"-m",mp,"-t",str(nt),"-c","2048","-n",str(np),"--temp","0.7","--top-p","0.9","-p",prompt,"--no-display-prompt","--log-disable"], capture_output=True, text=True, timeout=to, env=env)
+            command = [
+                cli, "-m", mp, "-t", str(nt), "-ngl", "0",
+                "-c", "1024", "-n", str(np), "-b", "128",
+                "--temp", "0.7", "--top-p", "0.9", "-p", prompt,
+                "--no-display-prompt",
+            ]
+            try:
+                p = subprocess.run(command, capture_output=True, text=True, timeout=to, env=env)
+            except subprocess.TimeoutExpired as e:
+                partial = e.stdout or e.stderr or ""
+                if isinstance(partial, bytes):
+                    partial = partial.decode("utf-8", "replace")
+                raise RuntimeError(f"llama.cpp timed out after {to}s; partial output: {partial[-1000:]}")
             wt = time.time()-t0
-            pt=gt=0.0; ptk=gtk=0
-            for line in p.stderr.splitlines():
+            if p.returncode != 0:
+                diagnostics = (p.stderr or p.stdout or "").strip()
+                raise RuntimeError(f"llama.cpp exited with code {p.returncode}: {diagnostics[-1500:]}")
+            combined_output = (p.stderr or "") + "\n" + (p.stdout or "")
+            pt = gt = 0.0
+            ptk = gtk = 0
+            for line in combined_output.splitlines():
                 tm = re.search(r"(\d+\.?\d*) tokens per second", line)
                 km = re.search(r"/ (\d+) tokens", line)
-                if "prompt eval" in line.lower():
-                    if tm: pt=float(tm.group(1))
-                    if km: ptk=int(km.group(1))
-                elif "eval time" in line.lower() and "prompt" not in line.lower():
-                    if tm: gt=float(tm.group(1))
-                    if km: gtk=int(km.group(1))
+                lower = line.lower()
+                if "prompt eval" in lower:
+                    if tm: pt = float(tm.group(1))
+                    if km: ptk = int(km.group(1))
+                elif "eval time" in lower and "prompt" not in lower:
+                    if tm: gt = float(tm.group(1))
+                    if km: gtk = int(km.group(1))
             r.update({"status":"success","total_time_sec":round(wt,2),"prompt_tokens":ptk,"completion_tokens":gtk,"prompt_tokens_per_sec":round(pt,1),"generation_tokens_per_sec":round(gt,1),"output":p.stdout.strip()[:500]})
         except Exception as e:
             r["error"] = str(e)