Build llama.cpp from source instead of pre-built binaries

app.py

@@ -416,82 +416,47 @@ _LLAMA_CPP_BIN = None  # path to downloaded llama-cli binary
 
 
 def _ensure_llama_cpp():
-    """Download pre-built llama.cpp binary if not cached."""
+    """Build or download llama.cpp binary if not cached."""
     global _LLAMA_CPP_BIN
     if _LLAMA_CPP_BIN and os.path.isfile(_LLAMA_CPP_BIN):
         return _LLAMA_CPP_BIN
 
     import subprocess
-    import zipfile
-    import urllib.request
 
     os.makedirs(_BENCHMARK_CACHE, exist_ok=True)
-    bin_dir = os.path.join(_BENCHMARK_CACHE, "llama_cpp")
-    os.makedirs(bin_dir, exist_ok=True)
-
+    build_dir = os.path.join(_BENCHMARK_CACHE, "llama_build")
+    bin_dir = os.path.join(build_dir, "build", "bin")
     llama_cli = os.path.join(bin_dir, "llama-cli")
-    lib_llama = os.path.join(bin_dir, "libllama.so")
-    if os.path.isfile(llama_cli) and os.path.isfile(lib_llama):
+
+    if os.path.isfile(llama_cli):
         _LLAMA_CPP_BIN = llama_cli
         return llama_cli
-    # Stale cache: remove old files and re-extract
-    if os.path.isfile(llama_cli):
-        _log("Benchmark: stale llama.cpp cache (missing libs), re-downloading...")
+
+    # Clone and build llama.cpp from source
+    _log("Benchmark: cloning llama.cpp...")
+    src_dir = os.path.join(_BENCHMARK_CACHE, "llama_src")
+    if os.path.isdir(src_dir):
         import shutil as _shutil
-        _shutil.rmtree(bin_dir, ignore_errors=True)
-        os.makedirs(bin_dir, exist_ok=True)
-
-    _log("Benchmark: downloading llama.cpp binary...")
-    # Use a known stable release
-    release = "b5440"
-    url = f"https://github.com/ggml-org/llama.cpp/releases/download/{release}/llama-{release}-bin-ubuntu-x64.zip"
-    _log(f"Benchmark: URL: {url}")
-    zip_path = os.path.join(bin_dir, "llama.zip")
-    urllib.request.urlretrieve(url, zip_path)
-    _log(f"Benchmark: downloaded {os.path.getsize(zip_path)} bytes")
-
-    _log("Benchmark: extracting llama.cpp...")
-    with zipfile.ZipFile(zip_path) as zf:
-        _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}")
-    _log(f"Benchmark: files: {os.listdir(bin_dir)}")
-    _log(f"Benchmark: libllama.so exists: {os.path.isfile(os.path.join(bin_dir, 'libllama.so'))}")
-    _log(f"Benchmark: libllama.so size: {os.path.getsize(os.path.join(bin_dir, 'libllama.so')) if os.path.isfile(os.path.join(bin_dir, 'libllama.so')) else 'N/A'}")
+        _shutil.rmtree(src_dir, ignore_errors=True)
 
-    # Verify binary works
-    _log("Benchmark: verifying llama-cli binary...")
-    try:
-        env = os.environ.copy()
-        env["LD_LIBRARY_PATH"] = bin_dir + ":" + env.get("LD_LIBRARY_PATH", "")
-        _log(f"Benchmark: LD_LIBRARY_PATH={env['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}")
+    subprocess.check_call(
+        ["git", "clone", "--depth=1", "https://github.com/ggml-org/llama.cpp.git", src_dir],
+        stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+    )
+    _log("Benchmark: building llama.cpp (this takes 2-5 minutes)...")
+    n_threads = _parse_cpu_count() or 8
+    os.makedirs(build_dir, exist_ok=True)
+    subprocess.check_call(
+        ["cmake", "-S", src_dir, "-B", build_dir,
+         "-DLLAMA_NATIVE=ON", "-DBUILD_SHARED_LIBS=ON"],
+        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)}")
 
     _LLAMA_CPP_BIN = llama_cli
     return llama_cli