patch to fix missing fine_tune_weights in MULTI_MODEL_CONFIG

Author: elizjo

ads/aqua/modeldeployment/model_group_config.py

@@ -44,8 +44,8 @@ class BaseModelSpec(BaseModel):
         Additional vLLM launch parameters for this model (e.g. parallelism, max context).
     model_task : str, optional
         Model task type (e.g., text-generation, image-to-text).
-    fine_tune_weights : List[FineTunedModelSpec], optional
-        List of associated fine-tuned models.
+    fine_tune_weights : List[List[LoraModuleSpec]], optional
+        List of associated LoRA modules for fine-tuned models.
     """
 
     model_path: str = Field(..., description="Path to the base model.")
@@ -71,17 +71,13 @@ def clean_model_path(cls, artifact_path_prefix: str) -> str:
             "The base model path is not available in the model artifact."
         )
 
-    @field_validator("fine_tune_weights")
     @classmethod
-    def set_fine_tuned_weights(cls, fine_tuned_weights: List[LoraModuleSpec]):
-        """Removes duplicate LoRA Modules (duplicate model_names in fine_tuned_weights)"""
+    def dedup_lora_modules(cls, fine_tune_weights: List[LoraModuleSpec]):
+        """Removes duplicate LoRA Modules (duplicate model_names in fine_tune_weights)"""
         seen_modules = set()
         unique_modules: List[LoraModuleSpec] = []
 
-        if not fine_tuned_weights:
-            return None
-
-        for lora_module in fine_tuned_weights:
+        for lora_module in fine_tune_weights or []:
             if lora_module.model_name not in seen_modules:
                 seen_modules.add(lora_module.model_name)
                 unique_modules.append(lora_module)
@@ -101,7 +97,7 @@ def from_aqua_multi_model_ref(
             model_path=model.artifact_location,
             params=model_params,
             model_task=model.model_task,
-            fine_tuned_weights=model.fine_tune_weights,
+            fine_tune_weights=cls.dedup_lora_modules(model.fine_tune_weights),
         )
 
 
@@ -112,7 +108,7 @@ class ModelGroupConfig(Serializable):
     Attributes
     ----------
     models : List[BaseModelConfig]
-        List of base models (with optional fine-tuned weights) to be served.
+        List of base models (with optional fine-tune weights) to be served.
     """
 
     models: List[BaseModelSpec] = Field(
@@ -228,5 +224,4 @@ def from_create_model_deployment_details(
                     "Each base model must have a unique `model_name`. "
                     "Please remove or rename the duplicate model and register the model group again."
                 )
-
         return cls(models=models)

tests/unitary/with_extras/aqua/test_deployment.py

@@ -2306,60 +2306,64 @@ def test_validate_multimodel_deployment_feasibility_positive_single(
             "test_data/deployment/aqua_summary_multi_model_single.json",
         )
 
-
 class TestBaseModelSpec:
     VALID_WEIGHT = LoraModuleSpec(
         model_name="ft_model",
         model_path="oci://test_bucket@test_namespace/",
     )
 
     @pytest.mark.parametrize(
-        "model_path, ft_weights, expect_warning",
+        "model_path, ft_weights, expect_warning, expect_error",
         [
-            ("oci://test_location_3", [VALID_WEIGHT, VALID_WEIGHT], True),
-            ("oci://test_location_3", [], False),
-            ("not-a-valid-uri", [VALID_WEIGHT], False),
+            ("oci://test_location_3", [VALID_WEIGHT, VALID_WEIGHT], True, False),
+            ("oci://test_location_3", [], False, False),
+            ("not-a-valid-uri", [VALID_WEIGHT], False, True),
         ],
     )
-    def test_invalid_base_model_spec(
+    def test_invalid_from_aqua_multi_model_ref(
         self,
         model_path,
         ft_weights,
         expect_warning,
+        expect_error,
         caplog,
     ):
         logger = logging.getLogger("ads.aqua.modeldeployment.model_group_config")
         logger.propagate = True
 
         caplog.set_level(logging.WARNING, logger=logger.name)
 
-        with pytest.raises(ValidationError) as excinfo:
-            BaseModelSpec(
-                model_id="test_model_id_3",
-                model_name="test_model_3",
-                model_task="code_synthesis",
-                model_path=model_path,
-                fine_tune_weights=ft_weights,
-            )
+        model_ref = AquaMultiModelRef(
+            artifact_location=model_path,
+            model_task="code_synthesis",
+            model_name="test_model_3",
+            model_id="test_model_id_3",
+            fine_tune_weights=ft_weights,
+            env_var={},
+            gpu_count=1,
+        )
+
+        model_params = "--dummy-param"
+
+        if expect_error:
+            with pytest.raises(ValidationError) as excinfo:
+                BaseModelSpec.from_aqua_multi_model_ref(model_ref, model_params)
+            errs = excinfo.value.errors()
+            if not model_path.startswith("oci://"):
+                model_path_errors = [e for e in errs if e["loc"] == ("model_path",)]
+                assert model_path_errors, f"expected a model_path error, got: {errs!r}"
+                assert (
+                    "the base model path is not available in the model artifact."
+                    in model_path_errors[0]["msg"].lower()
+                )
+        else:
+            BaseModelSpec.from_aqua_multi_model_ref(model_ref, model_params)
 
         messages = [rec.getMessage().lower() for rec in caplog.records]
-
         if expect_warning:
             assert any(
                 "duplicate lora modules detected" in m for m in messages
             ), f"Expected warning, got none. Captured messages: {messages}"
         else:
             assert not messages, f"Did not expect any warnings, but got: {messages}"
 
-        # inspecting if errors are thrown
-        errs = excinfo.value.errors()
-        if not model_path.startswith("oci://"):
-            model_path_errors = [e for e in errs if e["loc"] == ("model_path",)]
-            assert model_path_errors, f"expected a model_path error, got: {errs!r}"
-            assert (
-                "the base model path is not available in the model artifact."
-                in model_path_errors[0]["msg"].lower()
-            )
-        else:
-            # e.g. for the duplicate‐weights case you might check for a different loc/msg
-            pass