From dbda3de2c5dc75a26b89f078e06ee6fb70f9be54 Mon Sep 17 00:00:00 2001
From: reginaldalfret <reginaldalfret@gmail.com>
Date: Fri, 9 Oct 2026 14:58:34 +0530
Subject: [PATCH] fix(train): omit input_data_config when no input channels are
 provided

When a ModelTrainer has no input channels specified, _create_training_job_args
serialized an empty list ('input_data_config: []'). Downstream CreatePipeline
rejects an empty InputDataConfig array with ValidationException because the
botocore service model enforces min=1 on InputDataConfig when present.

Omit input_data_config from training_request when final_input_data_config is
empty, allowing the SageMaker API and CreatePipeline to succeed as in v2.

Fixes #6156

Signed-off-by: reginaldalfret <reginaldalfret@gmail.com>
---
 .../src/sagemaker/train/model_trainer.py      |  6 ++-
 .../tests/unit/train/test_model_trainer.py    | 40 +++++++++++++++++++
 2 files changed, 44 insertions(+), 2 deletions(-)

diff --git a/sagemaker-train/src/sagemaker/train/model_trainer.py b/sagemaker-train/src/sagemaker/train/model_trainer.py
index 8816ea68f9..53b70ec515 100644
--- a/sagemaker-train/src/sagemaker/train/model_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/model_trainer.py
@@ -851,7 +851,6 @@ def _create_training_job_args(
             "training_job_name": current_training_job_name,
             "algorithm_specification": algorithm_specification,
             "hyper_parameters": string_hyper_parameters,
-            "input_data_config": final_input_data_config,
             "resource_config": resource_config,
             "vpc_config": vpc_config,
             "role_arn": self.role,
@@ -876,6 +875,9 @@ def _create_training_job_args(
             "session_chaining_config": self._session_chaining_config,
         }
 
+        if final_input_data_config:
+            training_request["input_data_config"] = final_input_data_config
+
         if self.model_package_config:
             training_request["model_package_config"] = self.model_package_config
 
@@ -995,7 +997,7 @@ def train(
                     "algorithm_specification"
                 ].container_entrypoint,
                 container_arguments=training_request["algorithm_specification"].container_arguments,
-                input_data_config=training_request["input_data_config"],
+                input_data_config=training_request.get("input_data_config"),
                 hyper_parameters=training_request["hyper_parameters"],
                 environment=training_request["environment"],
             )
diff --git a/sagemaker-train/tests/unit/train/test_model_trainer.py b/sagemaker-train/tests/unit/train/test_model_trainer.py
index f990b7ab88..2a3a4a965b 100644
--- a/sagemaker-train/tests/unit/train/test_model_trainer.py
+++ b/sagemaker-train/tests/unit/train/test_model_trainer.py
@@ -663,6 +663,46 @@ def _echo(path, bucket, key_prefix, extra_args=None):
     assert f"/{DEFAULT_BASE_NAME}/input" not in uris["sm_drivers"]
 
 
+def test_create_training_job_args_omits_empty_input_data_config():
+    """ModelTrainer with no input channels omits input_data_config to avoid ValidationException (issue #6156)."""
+    trainer = ModelTrainer(
+        training_image=DEFAULT_IMAGE,
+        role=DEFAULT_ROLE,
+        compute=DEFAULT_COMPUTE_CONFIG,
+        stopping_condition=DEFAULT_STOPPING_CONDITION,
+        output_data_config=DEFAULT_OUTPUT_DATA_CONFIG,
+    )
+    args = trainer._create_training_job_args()
+    assert "input_data_config" not in args
+
+
+def test_create_training_job_args_includes_provided_input_data_config():
+    """ModelTrainer with input channels includes input_data_config in training args."""
+    channel = Channel(
+        channel_name="train",
+        data_source=DataSource(
+            s3_data_source=S3DataSource(
+                s3_data_type="S3Prefix",
+                s3_uri=f"s3://{DEFAULT_BUCKET}/data",
+                s3_data_distribution_type="FullyReplicated",
+            )
+        ),
+    )
+    trainer = ModelTrainer(
+        training_image=DEFAULT_IMAGE,
+        role=DEFAULT_ROLE,
+        compute=DEFAULT_COMPUTE_CONFIG,
+        stopping_condition=DEFAULT_STOPPING_CONDITION,
+        output_data_config=DEFAULT_OUTPUT_DATA_CONFIG,
+        input_data_config=[channel],
+    )
+    args = trainer._create_training_job_args()
+    assert "input_data_config" in args
+    assert len(args["input_data_config"]) == 1
+    assert args["input_data_config"][0].channel_name == "train"
+
+
+
 HETEROGENEOUS_INSTANCE_GROUPS = [
     InstanceGroup(
         instance_type="ml.t3.large", instance_count=1, instance_group_name="head-instance-group"