Fix: build llama.cpp with static linking, find binary after build

app.py

@@ -424,8 +424,7 @@ def _ensure_llama_cpp():
     import subprocess
 
     os.makedirs(_BENCHMARK_CACHE, exist_ok=True)
-    build_dir = os.path.join(_BENCHMARK_CACHE, "llama_build")
-    bin_dir = os.path.join(build_dir, "build", "bin")
+    bin_dir = os.path.join(_BENCHMARK_CACHE, "llama_bin")
     llama_cli = os.path.join(bin_dir, "llama-cli")
 
     if os.path.isfile(llama_cli):
@@ -445,18 +444,28 @@ def _ensure_llama_cpp():
     )
     _log("Benchmark: building llama.cpp (this takes 2-5 minutes)...")
     n_threads = _parse_cpu_count() or 8
+    build_dir = os.path.join(_BENCHMARK_CACHE, "llama_build")
     os.makedirs(build_dir, exist_ok=True)
+    os.makedirs(bin_dir, exist_ok=True)
     subprocess.check_call(
         ["cmake", "-S", src_dir, "-B", build_dir,
-         "-DLLAMA_NATIVE=ON", "-DBUILD_SHARED_LIBS=ON"],
+         "-DLLAMA_NATIVE=ON", "-DBUILD_SHARED_LIBS=OFF"],
         stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
     )
     subprocess.check_call(
         ["cmake", "--build", build_dir, "-j", str(n_threads), "--target", "llama-cli"],
         stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
     )
-    _log(f"Benchmark: llama.cpp built at {llama_cli}")
-    _log(f"Benchmark: files: {os.listdir(bin_dir)}")
+    # Find the built binary
+    import glob as _glob
+    candidates = _glob.glob(os.path.join(build_dir, "**", "llama-cli"), recursive=True)
+    if candidates:
+        import shutil as _shutil2
+        _shutil2.copy2(candidates[0], llama_cli)
+        os.chmod(llama_cli, 0o755)
+        _log(f"Benchmark: llama.cpp built at {llama_cli}")
+    else:
+        raise RuntimeError(f"llama-cli not found after build in {build_dir}")
 
     _LLAMA_CPP_BIN = llama_cli
     return llama_cli