fix: Resolve JSONDecodeError in LLM fine-tuning tune() API

sdk/python/v1beta1/kubeflow/katib/api/katib_client.py

@@ -619,6 +619,14 @@ class name in this argument.
                     "Trainer parameters must be an instance of HuggingFaceTrainerParams."
                 )
 
+            if not trainer_parameters.training_parameters:
+                raise ValueError(
+                    "trainer_parameters.training_parameters must be provided. "
+                    "Please use transformers.TrainingArguments(...) to configure "
+                    "training. See: https://www.kubeflow.org/docs/components/katib/"
+                    "user-guides/hp-tuning/configure-experiment/#tune-api"
+                )
+
             # Iterate over input parameters and do substitutions.
             experiment_parameters = []
             trial_parameters = []
@@ -665,9 +673,17 @@ class name in this argument.
                     "--dataset_dir",
                     VOLUME_PATH_DATASET,
                     "--lora_config",
-                    f"'{lora_config}'",
+                    (
+                        json.dumps(lora_config)
+                        if isinstance(lora_config, dict)
+                        else lora_config
+                    ),
                     "--training_parameters",
-                    f"'{training_args}'",
+                    (
+                        json.dumps(training_args)
+                        if isinstance(training_args, dict)
+                        else training_args
+                    ),
                 ],
                 volume_mounts=[STORAGE_INITIALIZER_VOLUME_MOUNT],
                 resources=(

sdk/python/v1beta1/kubeflow/katib/api/katib_client_test.py

@@ -458,6 +458,30 @@ def create_experiment(
         },
         ValueError,
     ),
+    (
+        "missing training_parameters in HuggingFaceTrainerParams - issue #2587",
+        {
+            "name": "tune_test",
+            "model_provider_parameters": HuggingFaceModelParams(
+                model_uri="hf://google-bert/bert-base-cased",
+                transformer_type=transformers.AutoModelForSequenceClassification,
+                num_labels=5,
+            ),
+            "dataset_provider_parameters": HuggingFaceDatasetParams(
+                repo_id="yelp_review_full",
+                split="train[:3000]",
+            ),
+            "trainer_parameters": HuggingFaceTrainerParams(
+                training_parameters=None,
+            ),
+            "resources_per_trial": types.TrainerResources(
+                num_workers=2,
+                num_procs_per_worker=2,
+                resources_per_worker={"gpu": "2"},
+            ),
+        },
+        ValueError,
+    ),
     (
         "pvc creation failed",
         {
@@ -881,6 +905,29 @@ def test_tune(katib_client, test_name, kwargs, expected_output):
                         additional_metric_names=[],
                     )
 
+                    # Verify JSON serialization fix
+                    # Container args should NOT have extra shell quotes like '{"key": "value"}'
+                    container_args = (
+                        experiment.spec.trial_template.trial_spec.spec.pytorch_replica_specs[
+                            "Master"
+                        ]
+                        .template.spec.containers[0]
+                        .args
+                    )
+                    for i, arg in enumerate(container_args):
+                        if arg in ("--training_parameters", "--lora_config"):
+                            assert i + 1 < len(
+                                container_args
+                            ), f"Missing value for {arg} in container args"
+                            next_arg = container_args[i + 1]
+                            # Should NOT start/end with extra single quotes
+                            assert not (
+                                next_arg.startswith("'") and next_arg.endswith("'")
+                            ), (
+                                f"{arg} value should not be wrapped with extra quotes. "
+                                f"Got: {next_arg[:50]}..."
+                            )
+
                 elif test_name == "valid flow with pip_index_urls":
                     # Verify pip install command in container args.
                     args_content = "".join(