Rebuild llama.cpp with stable CPU backend

app.py

@@ -200,7 +200,7 @@ _BENCH_MODELS = [
     {"name": "Qwen3 8B", "repo": "Qwen/Qwen3-8B-GGUF", "file": "Qwen3-8B-Q4_K_M.gguf", "size_gb": 5.0},
     {"name": "Qwen3.8 27B", "repo": "unsloth/Qwen3.8-27B-GGUF", "file": "Qwen3.8-27B-Q4_0.gguf", "size_gb": 16.0},
 ]
-_BENCH_CACHE = os.path.join(tempfile.gettempdir(), "llm_bench")
+_BENCH_CACHE = os.path.join(tempfile.gettempdir(), "llm_bench_cpu_v2")
 _BENCH_RESULTS = []
 _BENCH_LOCK = threading.Lock()
 _BENCH_JOBS = {}
@@ -221,7 +221,7 @@ 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","-DLLAMA_CUDA=OFF","-DGGML_CUDA=OFF","-DBUILD_SHARED_LIBS=OFF"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    subprocess.check_call(["cmake","-S",src,"-B",bd,"-DGGML_NATIVE=OFF","-DGGML_OPENMP=OFF","-DGGML_CUDA=OFF","-DGGML_AVX=ON","-DGGML_AVX2=ON","-DGGML_FMA=ON","-DGGML_AVX512=OFF","-DBUILD_SHARED_LIBS=OFF"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
     subprocess.check_call(["cmake","--build",bd,"-j",str(nt),"--target","llama-cli","llama-bench"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
     import glob as _g
     cands = [p for p in _g.glob(os.path.join(bd,"**","llama-cli"), recursive=True) if os.path.isfile(p)]