Fix

Author: nargokul

sagemaker-serve/src/sagemaker/serve/model_builder.py

@@ -514,12 +514,13 @@ def _resolve_model_artifact_uri(self) -> Optional[str]:
         fine-tuned model.
 
         Returns:
-            Optional[str]: S3 URI to model artifacts, or None for LORA adapters
+            Optional[str]: S3 URI to model artifacts, or None when not needed
 
         Logic:
             - For LORA adapters: Returns None (adapter weights are separate)
+            - For fine-tuned models: Returns None (model data is handled by the recipe/container)
             - For base models: Uses HostingArtifactUri from JumpStart hub metadata
-            - For full fine-tuned models: Uses ModelPackage artifact location
+            - For non-model-customization: Returns None
 
         Raises:
             ValueError: If model package or hub metadata is unavailable when needed
@@ -534,22 +535,23 @@ def _resolve_model_artifact_uri(self) -> Optional[str]:
         if self._is_model_customization():
             model_package = self._fetch_model_package()
             if model_package:
-                # Check if this is a full fine-tuned model (has its own artifacts)
                 if (hasattr(model_package, 'inference_specification') and
                     model_package.inference_specification and
                     hasattr(model_package.inference_specification, 'containers') and
                     model_package.inference_specification.containers):
 
                     container = model_package.inference_specification.containers[0]
 
-                    # If container has model_data_source, use it (full fine-tuned model)
+                    # For fine-tuned models (have model_data_source), return None.
+                    # The model data is handled by the recipe/container configuration,
+                    # not via artifact_url in CreateInferenceComponent.
                     if (hasattr(container, 'model_data_source') and
                         container.model_data_source and
                         hasattr(container.model_data_source, 's3_data_source') and
                         container.model_data_source.s3_data_source):
-                        return container.model_data_source.s3_data_source.s3_uri
+                        return None
 
-                    # Otherwise, this is a base model - get HostingArtifactUri from JumpStart
+                    # For base models, get HostingArtifactUri from JumpStart
                     if hasattr(container, 'base_model') and container.base_model:
                         try:
                             hub_document = self._fetch_hub_document_for_custom_model()
@@ -570,7 +572,6 @@ def _resolve_model_artifact_uri(self) -> Optional[str]:
                             return None
 
         # For non-model-customization deployments, return None
-        # (artifact handling is done differently for those cases)
         return None
 
     def _infer_instance_type_from_jumpstart(self) -> str:

sagemaker-serve/tests/unit/test_artifact_path_propagation.py

@@ -150,9 +150,8 @@ def test_fine_tuned_model_artifact_uri_propagated_to_inference_component(
         mock_is_model_customization.return_value = True
         mock_fetch_peft.return_value = "FULL"
 
-        # Setup: Artifact URI resolution returns model package S3 URI
-        expected_artifact_uri = 's3://my-bucket/fine-tuned-model/model.tar.gz'
-        mock_resolve_artifact.return_value = expected_artifact_uri
+        # Setup: Artifact URI resolution returns None for fine-tuned models
+        mock_resolve_artifact.return_value = None
 
         # Setup: Model package
         mock_package = Mock()
@@ -209,8 +208,8 @@ def test_fine_tuned_model_artifact_uri_propagated_to_inference_component(
         # Extract the specification
         ic_spec = call_kwargs['specification']
 
-        # Verify artifact_url matches the resolved URI
-        assert ic_spec.container.artifact_url == expected_artifact_uri
+        # Verify artifact_url is None for fine-tuned models (model data handled by recipe)
+        assert ic_spec.container.artifact_url is None
 
     @patch('sagemaker.core.resources.InferenceComponent.create')
     @patch('sagemaker.core.resources.InferenceComponent.get_all')

sagemaker-serve/tests/unit/test_artifact_path_resolution.py

@@ -95,7 +95,7 @@ def test_fine_tuned_adapter_artifact_location(
         mock_fetch_peft,
         mock_fetch_package
     ):
-        """Test fine-tuned adapter artifact location from ModelPackage."""
+        """Test fine-tuned model returns None (model data handled by recipe/container)."""
         # Setup: Full fine-tuned model (not LORA)
         mock_is_model_customization.return_value = True
         mock_fetch_peft.return_value = "FULL"
@@ -127,8 +127,8 @@ def test_fine_tuned_adapter_artifact_location(
         # Execute
         artifact_uri = builder._resolve_model_artifact_uri()
 
-        # Verify: Should return model_data_source S3 URI
-        assert artifact_uri == 's3://my-bucket/fine-tuned-model/model.tar.gz'
+        # Verify: Fine-tuned models return None - model data is handled by recipe/container
+        assert artifact_uri is None
 
     @patch('sagemaker.serve.model_builder.ModelBuilder._fetch_peft')
     @patch('sagemaker.serve.model_builder.ModelBuilder._is_model_customization')
@@ -434,7 +434,7 @@ def test_fine_tuned_model_with_nested_s3_data_source(
         mock_fetch_peft,
         mock_fetch_package
     ):
-        """Test fine-tuned model with properly nested s3_data_source structure."""
+        """Test fine-tuned model with nested s3_data_source returns None."""
         # Setup: Full fine-tuned model with nested structure
         mock_is_model_customization.return_value = True
         mock_fetch_peft.return_value = "FULL"
@@ -443,7 +443,6 @@ def test_fine_tuned_model_with_nested_s3_data_source(
         mock_package = Mock()
         mock_container = Mock()
 
-        # Create nested structure: container -> model_data_source -> s3_data_source -> s3_uri
         mock_s3_data_source = Mock()
         mock_s3_data_source.s3_uri = 's3://custom-bucket/my-fine-tuned-model/artifacts.tar.gz'
 
@@ -470,8 +469,8 @@ def test_fine_tuned_model_with_nested_s3_data_source(
         # Execute
         artifact_uri = builder._resolve_model_artifact_uri()
 
-        # Verify: Should return the nested S3 URI
-        assert artifact_uri == 's3://custom-bucket/my-fine-tuned-model/artifacts.tar.gz'
+        # Verify: Fine-tuned models return None - model data handled by recipe/container
+        assert artifact_uri is None
 
     @patch('sagemaker.serve.model_builder.ModelBuilder._fetch_hub_document_for_custom_model')
     @patch('sagemaker.serve.model_builder.ModelBuilder._fetch_model_package')