fix: update job key used for show_metrics/stream_logs in MTRL

sagemaker-serve/src/sagemaker/serve/bedrock_model_builder.py

@@ -364,7 +364,7 @@ def deploy(
                         self._get_bedrock_client(), model_arn
                     )
                     if existing_deployment:
-                        logger.warning(
+                        logger.info(
                             "Reusing existing custom model %s and deployment %s "
                             "(matched model-source tag). No new resources were created. "
                             "Pass reuse_resources=False to force new resources.",
@@ -375,7 +375,7 @@ def deploy(
                             "modelArn": model_arn,
                             "customModelDeploymentArn": existing_deployment,
                         }
-                    logger.warning(
+                    logger.info(
                         "Reusing existing custom model %s (matched model-source tag); "
                         "creating a new deployment on it. Pass reuse_resources=False to "
                         "force a new model.",

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

@@ -4061,7 +4061,7 @@ def build(
                 reusable_endpoint = self._find_reusable_endpoint()
                 if reusable_endpoint:
                     self._reused_endpoint_name = reusable_endpoint
-                logger.warning(
+                logger.info(
                     "Reusing existing Model %r (matched model-source tag). "
                     "No new Model will be created. Pass reuse_resources=False "
                     "to force a new Model.",
@@ -5528,7 +5528,7 @@ def deploy(
                         endpoint_name,
                         reusable_endpoint,
                     )
-                logger.warning(
+                logger.info(
                     "Reusing existing endpoint %r (matched model-source tag and "
                     "deployment configuration). No new resources were created. "
                     "Pass reuse_resources=False to force a new endpoint.",

sagemaker-train/src/sagemaker/train/base_trainer.py

@@ -399,30 +399,40 @@ def show_metrics(
             ValueError: If no training job has been run yet, no logs/metrics
                 are found, or MLflow is not configured for OSS models.
         """
-        # Validate that we have a training job to get metrics from
-        if not hasattr(self, '_latest_training_job') or self._latest_training_job is None:
-            raise ValueError(
-                "No training job found. Call .train() first, then call .show_metrics() "
-                "to view training metrics."
+        # Resolve the job reference. Prefer _latest_training_job (CreateTrainingJob),
+        # fall back to _latest_job (generic CreateJob API used by MTRL).
+        resolved_job = getattr(self, '_latest_training_job', None)
+        if resolved_job is None:
+            latest_job = getattr(self, '_latest_job', None)
+            if latest_job is None:
+                raise ValueError(
+                    "No training job found. Call .train() first, then call .show_metrics() "
+                    "to view training metrics. If training has already completed, set the "
+                    "job name directly via trainer._latest_training_job = '<job-name>' or "
+                    "trainer._latest_job = '<job-name>'."
+                )
+            resolved_job = (
+                latest_job.job_name if hasattr(latest_job, 'job_name') else str(latest_job)
             )
 
         # Route based on model type
         model_name = getattr(self, '_model_name', None)
         is_nova = _is_nova_model(model_name) if model_name else False
 
         if is_nova:
-            return self._show_metrics_cloudwatch(metrics, starting_step, ending_step, start_time, end_time)
+            return self._show_metrics_cloudwatch(resolved_job, metrics, starting_step, ending_step, start_time, end_time)
         else:
-            return self._show_metrics_mlflow(metrics, starting_step, ending_step)
+            return self._show_metrics_mlflow(resolved_job, metrics, starting_step, ending_step)
 
     def _show_metrics_mlflow(
         self,
+        resolved_job,
         metrics: Optional[List[str]] = None,
         starting_step: Optional[int] = None,
         ending_step: Optional[int] = None,
     ) -> None:
         """Pull and plot training metrics from MLflow for non-Nova models."""
-        training_job = self._latest_training_job
+        training_job = resolved_job
 
         # Resolve the TrainingJob object if it's a string
         if isinstance(training_job, str):
@@ -456,6 +466,7 @@ def _show_metrics_mlflow(
 
     def _show_metrics_cloudwatch(
         self,
+        resolved_job,
         metrics: Optional[List[str]] = None,
         starting_step: Optional[int] = None,
         ending_step: Optional[int] = None,
@@ -464,7 +475,7 @@ def _show_metrics_cloudwatch(
     ) -> Any:
         """Parse and plot training metrics from CloudWatch logs (Nova models)."""
 
-        training_job = self._latest_training_job
+        training_job = resolved_job
         if hasattr(training_job, 'training_job_name'):
             job_id = training_job.training_job_name
         elif isinstance(training_job, str):
@@ -631,10 +642,21 @@ def stream_logs(self, poll: int = 5, start_time: Optional[Any] = None) -> None:
         Raises:
             ValueError: If no training job has been run yet.
         """
-        if not hasattr(self, '_latest_training_job') or self._latest_training_job is None:
-            raise ValueError(
-                "No training job found. Call .train(wait=False) first, "
-                "then call .stream_logs() to stream logs in real-time."
+        # Resolve the job reference. Prefer _latest_training_job (CreateTrainingJob),
+        # fall back to _latest_job (generic CreateJob API used by MTRL).
+        resolved_job = getattr(self, '_latest_training_job', None)
+        if resolved_job is None:
+            latest_job = getattr(self, '_latest_job', None)
+            if latest_job is None:
+                raise ValueError(
+                    "No training job found. Call .train(wait=False) first, "
+                    "then call .stream_logs() to stream logs in real-time. "
+                    "If training has already completed, set the job name directly via "
+                    "trainer._latest_training_job = '<job-name>' or "
+                    "trainer._latest_job = '<job-name>'."
+                )
+            resolved_job = (
+                latest_job.job_name if hasattr(latest_job, 'job_name') else str(latest_job)
             )
 
         # Resolve start_time for SMHP jobs
@@ -645,7 +667,7 @@ def stream_logs(self, poll: int = 5, start_time: Optional[Any] = None) -> None:
             else:
                 start_time_ms = int(start_time)
 
-        training_job = self._latest_training_job
+        training_job = resolved_job
         compute = getattr(self, 'compute', None)
 
         if isinstance(compute, HyperPodCompute):

sagemaker-train/src/sagemaker/train/common_utils/cloudwatch_metrics.py

@@ -35,11 +35,13 @@
         "SFT": {"training_loss": TRAINING_LOSS_REGEX, "lr": LEARNING_RATE_REGEX},
         "CPT": {"training_loss": TRAINING_LOSS_REGEX, "lr": LEARNING_RATE_REGEX},
         "RLVR": {"reward_score": SMTJ_RLVR_REWARD_SCORE_REGEX},
+        "MTRL": {"reward_score": SMTJ_RLVR_REWARD_SCORE_REGEX},
     },
     "smhp": {
         "SFT": {"training_loss": TRAINING_LOSS_REGEX, "lr": LEARNING_RATE_REGEX},
         "CPT": {"training_loss": TRAINING_LOSS_REGEX, "lr": LEARNING_RATE_REGEX},
         "RLVR": {"reward_score": SMHP_RLVR_REWARD_SCORE_REGEX},
+        "MTRL": {"reward_score": SMHP_RLVR_REWARD_SCORE_REGEX},
     },
 }
 

sagemaker-train/src/sagemaker/train/common_utils/data_utils.py

@@ -248,7 +248,7 @@ def is_multimodal_data(dataset: Union[str, "DataSet"]) -> bool:
         True if multimodal fields detected, False otherwise
     """
 
-    logger.info(f"Auto-detecting whether dataset is multimodal: {dataset}")
+    logger.debug(f"Auto-detecting whether dataset is multimodal: {dataset}")
 
     if isinstance(dataset, DataSet):
         data_s3_path = dataset.source

sagemaker-train/src/sagemaker/train/cpt_trainer.py

@@ -40,7 +40,6 @@
 from sagemaker.core.telemetry.constants import Feature
 
 logger = logging.getLogger(__name__)
-logger.setLevel(logging.INFO)
 
 
 class CPTTrainer(BaseTrainer):

sagemaker-train/src/sagemaker/train/dpo_trainer.py

@@ -31,7 +31,6 @@
 from sagemaker.train.constants import get_sagemaker_hub_name
 
 logger = logging.getLogger(__name__)
-logger.setLevel(logging.INFO)
 
 
 class DPOTrainer(BaseTrainer):

sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py

@@ -173,6 +173,8 @@ class MultiTurnRLTrainer(BaseTrainer):
             and 'job_name_prefix'. If not specified, no notifications are sent.
     """
 
+    _customization_technique = "MTRL"
+
     def __init__(
         self,
         model: Union[str, ModelPackage],

sagemaker-train/src/sagemaker/train/sft_trainer.py

@@ -41,7 +41,6 @@
 from sagemaker.core.training.constants import TrainingPlatform
 
 logger = logging.getLogger(__name__)
-logger.setLevel(logging.INFO)
 
 
 class SFTTrainer(BaseTrainer):