fix: pre-compile vllm-metal paged_ops extension in tarball build

vllm-metal v0.2.0 JIT-compiles a paged_ops C++ extension using clang++
at runtime. This fails inside the macOS sandbox which blocks compiler
invocations. Instead, compile the extension during the tarball build
(where Xcode CLT is available) and ship the .so in a prebuilt/ dir.

At install time, model-runner copies the pre-built .so into the user's
~/.cache/vllm-metal/ cache directory. vllm-metal's build.py sees the
cached .so is newer than the sources and skips JIT compilation.

This also reverts the include/ directory preservation since the Python
headers are only needed for compilation, which now happens at build time.

Author: ilopezluna

pkg/inference/backends/vllm/vllm_metal.go

@@ -176,6 +176,29 @@ func (v *vllmMetal) downloadAndExtract(ctx context.Context, _ *http.Client) erro
 		return fmt.Errorf("failed to make python3 executable: %w", err)
 	}
 
+	// Copy pre-built Metal kernel extension to the user's cache directory
+	// so vllm-metal skips JIT compilation at runtime (the macOS sandbox
+	// blocks clang++ invocations needed by the JIT compiler).
+	homeDir, err := os.UserHomeDir()
+	if err == nil {
+		cacheDir := filepath.Join(homeDir, ".cache", "vllm-metal")
+		prebuiltDir := filepath.Join(v.installDir, "prebuilt")
+		if entries, readErr := os.ReadDir(prebuiltDir); readErr == nil {
+			if mkErr := os.MkdirAll(cacheDir, 0755); mkErr == nil {
+				for _, entry := range entries {
+					src := filepath.Join(prebuiltDir, entry.Name())
+					dst := filepath.Join(cacheDir, entry.Name())
+					if data, cpErr := os.ReadFile(src); cpErr == nil {
+						if wErr := os.WriteFile(dst, data, 0755); wErr != nil {
+							v.log.Warn("failed to copy prebuilt extension", "file", entry.Name(), "error", wErr)
+						}
+					}
+				}
+				v.log.Info("Copied pre-built Metal kernel extension to cache", "cacheDir", cacheDir)
+			}
+		}
+	}
+
 	v.log.Info("vllm-metal installed successfully", "version", vllmMetalVersion)
 	return nil
 }

scripts/build-vllm-metal-tarball.sh

@@ -70,10 +70,18 @@ curl -fsSL -O "$VLLM_METAL_WHEEL_URL"
 uv pip install --python "$PYTHON_DIR/bin/python3" --system vllm_metal-*.whl
 rm -f vllm_metal-*.whl
 
+# Pre-compile the paged_ops Metal kernel extension so users don't need Xcode CLT
+# at runtime (the macOS sandbox blocks clang++ invocations).  build.py caches the
+# compiled .so under ~/.cache/vllm-metal/; we redirect $HOME so the artefact
+# lands in a known temp location we can bundle into the tarball.
+echo "Pre-compiling vllm-metal paged_ops extension..."
+HOME="$WORK_DIR" "$PYTHON_DIR/bin/python3" -c "from vllm_metal.metal.build import build; build()"
+mkdir -p "$PYTHON_DIR/prebuilt"
+cp "$WORK_DIR/.cache/vllm-metal/"*_paged_ops* "$PYTHON_DIR/prebuilt/"
+
 # Strip files not needed at runtime to reduce tarball size
 echo "Stripping unnecessary files..."
-# Keep include/python3.12 (needed by vllm-metal to compile Metal kernels at runtime)
-find "$PYTHON_DIR/include" -mindepth 1 -maxdepth 1 ! -name 'python3.12' -exec rm -rf {} + 2>/dev/null || true
+rm -rf "$PYTHON_DIR/include"
 rm -rf "$PYTHON_DIR/share"
 PYLIB="$PYTHON_DIR/lib/python3.12"
 rm -rf "$PYLIB/test" "$PYLIB/tests"