Add model package support

docs/source/reference/cli.rst

@@ -170,6 +170,17 @@ Run benchmarking using llm-eval.
     :prog: olive
     :path: benchmark
 
+Generate Model Package
+======================
+
+Merge multiple model outputs into a model package with manifest and per-component metadata.
+
+.. argparse::
+    :module: olive.cli.launcher
+    :func: get_cli_parser
+    :prog: olive
+    :path: generate-model-package
+
 Providing Input Models
 ======================
 

olive/cache.py

@@ -384,10 +384,10 @@ def save_model(
     ):
         """Save a model from the cache to a given path."""
         output_dir = Path(output_dir) if output_dir else Path.cwd()
-
-        # If output_dir has a suffix (like .onnx), it's a file path
-        # Use parent directory for saving files
-        actual_output_dir = output_dir.parent if output_dir.suffix else output_dir
+        if output_dir.suffix and not output_dir.is_dir():
+            actual_output_dir = output_dir.parent
+        else:
+            actual_output_dir = output_dir
         actual_output_dir.mkdir(parents=True, exist_ok=True)
 
         model_json = self.load_model(model_id)

olive/cli/launcher.py

@@ -17,6 +17,7 @@
 from olive.cli.generate_adapter import GenerateAdapterCommand
 from olive.cli.generate_cost_model import GenerateCostModelCommand
 from olive.cli.init import InitCommand
+from olive.cli.model_package import ModelPackageCommand
 from olive.cli.optimize import OptimizeCommand
 from olive.cli.quantize import QuantizeCommand
 from olive.cli.run import WorkflowRunCommand
@@ -54,6 +55,7 @@ def get_cli_parser(called_as_console_script: bool = True) -> ArgumentParser:
     ConfigureQualcommSDKCommand.register_subcommand(commands_parser)
     SharedCacheCommand.register_subcommand(commands_parser)
     ExtractAdaptersCommand.register_subcommand(commands_parser)
+    ModelPackageCommand.register_subcommand(commands_parser)
     BenchmarkCommand.register_subcommand(commands_parser)
 
     return parser