Fix: extract all llama.cpp libs, set LD_LIBRARY_PATH

app.py

@@ -445,23 +445,41 @@ def _ensure_llama_cpp():
 
     _log("Benchmark: extracting llama.cpp...")
     with zipfile.ZipFile(zip_path) as zf:
-        _log(f"Benchmark: zip contents: {zf.namelist()}")
-        for name in zf.namelist():
-            if name.endswith("llama-cli") or name.endswith("llama-cli.exe"):
-                data = zf.read(name)
-                with open(llama_cli, "wb") as f:
-                    f.write(data)
-                _log(f"Benchmark: extracted {name} ({len(data)} bytes)")
-                break
-    os.chmod(llama_cli, 0o755)
+        _log(f"Benchmark: zip has {len(zf.namelist())} files")
+        # Extract ALL files (libs + binaries) to bin_dir
+        zf.extractall(bin_dir)
     os.unlink(zip_path)
 
+    # The files are under build/bin/ - move them up and fix permissions
+    build_bin = os.path.join(bin_dir, "build", "bin")
+    if os.path.isdir(build_bin):
+        for fname in os.listdir(build_bin):
+            src = os.path.join(build_bin, fname)
+            dst = os.path.join(bin_dir, fname)
+            os.rename(src, dst)
+        os.rmdir(os.path.join(bin_dir, "build", "bin"))
+        os.rmdir(os.path.join(bin_dir, "build"))
+
+    # Make all files executable
+    for fname in os.listdir(bin_dir):
+        fpath = os.path.join(bin_dir, fname)
+        if os.path.isfile(fpath):
+            os.chmod(fpath, 0o755)
+
+    _log(f"Benchmark: extracted to {bin_dir}: {os.listdir(bin_dir)}")
+
     # Verify binary works
     _log("Benchmark: verifying llama-cli binary...")
     try:
-        check = subprocess.run([llama_cli, "--help"], capture_output=True, text=True, timeout=10)
-        _log(f"Benchmark: llama-cli --help exit code: {check.returncode}")
-        _log(f"Benchmark: llama-cli --help output: {check.stdout[:200]}")
+        # Set LD_LIBRARY_PATH so the binary can find shared libs
+        env = os.environ.copy()
+        env["LD_LIBRARY_PATH"] = bin_dir + ":" + env.get("LD_LIBRARY_PATH", "")
+        check = subprocess.run(
+            [llama_cli, "--version"], capture_output=True, text=True, timeout=10, env=env,
+        )
+        _log(f"Benchmark: llama-cli --version exit code: {check.returncode}")
+        _log(f"Benchmark: llama-cli --version stdout: {check.stdout[:200]}")
+        _log(f"Benchmark: llama-cli --version stderr: {check.stderr[:200]}")
     except Exception as e:
         _log(f"Benchmark: llama-cli verification failed: {e}")
 
@@ -495,6 +513,9 @@ def _run_benchmark():
             )
 
             _log(f"Benchmark: running {model_info['name']} with {n_threads} threads...")
+            env = os.environ.copy()
+            bin_dir = os.path.dirname(llama_cli)
+            env["LD_LIBRARY_PATH"] = bin_dir + ":" + env.get("LD_LIBRARY_PATH", "")
             cmd = [
                 llama_cli,
                 "-m", model_path,
@@ -509,10 +530,11 @@ def _run_benchmark():
             ]
             _log(f"Benchmark: cmd: {' '.join(cmd[:6])}...")
             _log(f"Benchmark: model exists: {os.path.isfile(model_path)}, size: {os.path.getsize(model_path) if os.path.isfile(model_path) else 0}")
+            _log(f"Benchmark: LD_LIBRARY_PATH={bin_dir}")
             t0 = time.time()
             try:
                 proc = subprocess.run(
-                    cmd, capture_output=True, text=True, timeout=300,
+                    cmd, capture_output=True, text=True, timeout=300, env=env,
                 )
             except subprocess.TimeoutExpired:
                 result["error"] = "Timeout after 300s"