From 9dee3958ac6e562bea2256a01eb4fa569853dbd0 Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Mon, 13 Jul 2026 15:53:56 -0700
Subject: [PATCH 1/8] fix: filter full recipe template from serverless train( )

---
 .../src/sagemaker/train/base_trainer.py       | 90 +++++++++++++++++--
 .../src/sagemaker/train/dpo_trainer.py        |  9 +-
 .../sagemaker/train/multi_turn_rl_trainer.py  |  8 +-
 .../src/sagemaker/train/rlvr_trainer.py       | 11 ++-
 .../src/sagemaker/train/sft_trainer.py        |  9 +-
 5 files changed, 113 insertions(+), 14 deletions(-)

diff --git a/sagemaker-train/src/sagemaker/train/base_trainer.py b/sagemaker-train/src/sagemaker/train/base_trainer.py
index a453720f53..5edbc671bb 100644
--- a/sagemaker-train/src/sagemaker/train/base_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/base_trainer.py
@@ -236,19 +236,79 @@ def _patch_resolved_recipe(self, resolved: Dict[str, Any]) -> None:
             if dotpath:
                 _set_nested_value(resolved, dotpath, value)
 
-    def _apply_recipe_to_hyperparameters(self, final_hyperparameters: Dict[str, Any]) -> Dict[str, Any]:
+    def _get_user_provided_recipe_keys(self) -> set:
+        """Return the set of leaf keys the user explicitly provided.
+
+        Collects keys from every source that represents an explicit user choice:
+
+        * direct hyperparameter assignments (``trainer.hyperparameters.x = val``),
+          tracked in ``hyperparameters._user_set``;
+        * the programmatic ``overrides`` dict passed at construction;
+        * the user recipe YAML file passed at construction.
+
+        These are the keys that should be forwarded to a serverless training job,
+        alongside the Hub override spec. Full recipe-template internal keys that
+        the user never touched are intentionally excluded.
+
+        Returns:
+            Set of leaf key names the user explicitly provided. Empty when the
+            user provided nothing beyond Hub defaults.
+        """
+        keys: set = set()
+
+        # Direct hyperparameter assignments (always members of the Hub spec).
+        user_set = getattr(getattr(self, 'hyperparameters', None), '_user_set', None)
+        if isinstance(user_set, set):
+            keys.update(user_set)
+
+        # Programmatic overrides dict (may contain non-spec recipe keys).
+        overrides = getattr(self, '_overrides', None)
+        if isinstance(overrides, dict) and overrides:
+            keys.update(flatten_resolved_recipe(overrides).keys())
+
+        # User recipe YAML file (may contain non-spec recipe keys).
+        recipe_path = getattr(self, '_recipe_path', None)
+        if recipe_path:
+            try:
+                from sagemaker.train.recipe_resolver import _load_user_recipe
+
+                user_recipe = _load_user_recipe(recipe_path)
+                keys.update(flatten_resolved_recipe(user_recipe).keys())
+            except Exception as e:  # pragma: no cover - best-effort key discovery
+                logger.debug("Could not load user recipe to collect keys: %s", e)
+
+        return keys
+
+
+    def _apply_recipe_to_hyperparameters(
+        self,
+        final_hyperparameters: Dict[str, Any],
+        only_user_overrides: bool = False,
+    ) -> Dict[str, Any]:
         """Apply resolved recipe values to final_hyperparameters dict.
 
         If recipe/overrides were provided, or if the user set hyperparameters
         directly via ``.hyperparameters.*``, merges resolved recipe values into
-        the hyperparameters dict. All leaf values from the resolved recipe are
-        applied — including keys not in the Hub spec subset — enabling
-        power users to override any parameter in the full recipe.
+        the hyperparameters dict. By default all leaf values from the resolved
+        recipe are applied — including keys not in the Hub spec subset. This is
+        used by the serverful/HyperPod paths, which render the values into a
+        recipe YAML file (no key-count limit).
+
         Values are converted to strings (matching the SageMaker API
         expectation for hyperparameter values).
 
         Args:
             final_hyperparameters: The hyperparameters dict from to_dict().
+            only_user_overrides: When True, only apply resolved values for keys the
+                user explicitly provided — direct hyperparameter assignments, the
+                ``overrides`` dict, or a recipe file — instead of flattening the
+                entire resolved recipe. This is required for serverless SageMaker
+                training jobs, whose ``CreateTrainingJob`` ``HyperParameters`` map
+                is limited to 100 members by the CreateTrainingJob: the resolved full recipe
+                template can contain several hundred internal leaf keys the user
+                never touched. User-provided keys are not restricted to the Hub
+                spec — power users may override any recipe key, and validity is
+                enforced by ``get_resolved_recipe()``.
 
         Returns:
             The updated hyperparameters dict with recipe values applied.
@@ -259,12 +319,30 @@ def _apply_recipe_to_hyperparameters(self, final_hyperparameters: Dict[str, Any]
         try:
             resolved = self.get_resolved_recipe()
         except NoRecipeError:
+            
             return final_hyperparameters
 
         flat = flatten_resolved_recipe(resolved)
+
+        allowed_keys = None
+        if only_user_overrides:
+            try:
+                allowed_keys = self._get_user_provided_recipe_keys()
+            except Exception as e:
+                logger.warning(
+                    "Failed to determine user-provided recipe keys (%s); "
+                    "falling back to submitting the full resolved recipe.",
+                    e,
+                )
+                allowed_keys = None
+
         for k, v in flat.items():
-            if v is not None:
-                final_hyperparameters[k] = str(v) if not isinstance(v, str) else v
+            if v is None:
+                continue
+            if allowed_keys is not None and k not in allowed_keys:
+                logger.info("Skipping recipe key '%s' (not user-provided).", k)
+                continue
+            final_hyperparameters[k] = str(v) if not isinstance(v, str) else v
 
         return final_hyperparameters
 
diff --git a/sagemaker-train/src/sagemaker/train/dpo_trainer.py b/sagemaker-train/src/sagemaker/train/dpo_trainer.py
index a83d1ed2fe..e56adde46e 100644
--- a/sagemaker-train/src/sagemaker/train/dpo_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/dpo_trainer.py
@@ -308,8 +308,13 @@ def train(self,
 
         final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Serverless CreateTrainingJob limits HyperParameters to 100 members, so
+        # only forward the values the user explicitly overrode rather than the
+        # entire resolved recipe template.
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(
+            final_hyperparameters, only_user_overrides=True
+        )
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
             effective_training_dataset = training_dataset or self.training_dataset
diff --git a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
index daf3c8f5d9..35e137ee50 100644
--- a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
@@ -289,8 +289,12 @@ def train(
 
         self._final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
-        self._final_hyperparameters = self._apply_recipe_to_hyperparameters(self._final_hyperparameters)
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Only forward the values the user explicitly overrode rather than the
+        # entire resolved recipe template.
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(
+            final_hyperparameters, only_user_overrides=True
+        )
 
         _validate_hyperparameter_values(self._final_hyperparameters)
 
diff --git a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
index f20b54e7b1..279cd20857 100644
--- a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
@@ -471,8 +471,13 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None,
 
         final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Serverless CreateTrainingJob limits HyperParameters to 100 members, so
+        # only forward the values the user explicitly overrode rather than the
+        # entire resolved recipe template.
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(
+            final_hyperparameters, only_user_overrides=False
+        )
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
             effective_training_dataset = training_dataset or self.training_dataset
@@ -518,6 +523,8 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None,
         if self.stopping_condition is not None:
             create_args["stopping_condition"] = self.stopping_condition
 
+        logger.info("Final hyperparameters before submitting training job: %s", final_hyperparameters)
+
         try:
             training_job = TrainingJob.create(**create_args)
         except Exception as e:
diff --git a/sagemaker-train/src/sagemaker/train/sft_trainer.py b/sagemaker-train/src/sagemaker/train/sft_trainer.py
index 3f14c348f0..7892fb8b28 100644
--- a/sagemaker-train/src/sagemaker/train/sft_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/sft_trainer.py
@@ -361,8 +361,13 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None, validati
 
         final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Serverless CreateTrainingJob limits HyperParameters to 100 members, so
+        # only forward the values the user explicitly overrode rather than the
+        # entire resolved recipe template.
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(
+            final_hyperparameters, only_user_overrides=True
+        )
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
             effective_training_dataset = training_dataset or self.training_dataset

From 038a14d4c76b5efefdfbb8c47f4967cc618cdd4e Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Mon, 13 Jul 2026 15:58:15 -0700
Subject: [PATCH 2/8] fix: update hyperparam ref in mtrl trainer

---
 sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
index 35e137ee50..0c2a44e647 100644
--- a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
@@ -292,8 +292,8 @@ def train(
         # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
         # Only forward the values the user explicitly overrode rather than the
         # entire resolved recipe template.
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            final_hyperparameters, only_user_overrides=True
+        self._final_hyperparameters = self._apply_recipe_to_hyperparameters(
+            self._final_hyperparameters, only_user_overrides=True
         )
 
         _validate_hyperparameter_values(self._final_hyperparameters)

From c51d4e68ce6e4bd934cfb7eae076c33682a0be98 Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Mon, 13 Jul 2026 16:04:24 -0700
Subject: [PATCH 3/8] fix: enable only_user_overrides for RLVRTrainer
 serverless path

---
 sagemaker-train/src/sagemaker/train/rlvr_trainer.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
index 279cd20857..da151d7165 100644
--- a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
@@ -476,7 +476,7 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None,
         # only forward the values the user explicitly overrode rather than the
         # entire resolved recipe template.
         final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            final_hyperparameters, only_user_overrides=False
+            final_hyperparameters, only_user_overrides=True
         )
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:

From 15060015c9275a5cfa596d6f8312676f200ea3c2 Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Tue, 14 Jul 2026 12:11:48 -0700
Subject: [PATCH 4/8] chore: removing extra logging

---
 sagemaker-train/src/sagemaker/train/base_trainer.py | 1 -
 sagemaker-train/src/sagemaker/train/rlvr_trainer.py | 2 --
 2 files changed, 3 deletions(-)

diff --git a/sagemaker-train/src/sagemaker/train/base_trainer.py b/sagemaker-train/src/sagemaker/train/base_trainer.py
index 5edbc671bb..966bb960f4 100644
--- a/sagemaker-train/src/sagemaker/train/base_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/base_trainer.py
@@ -340,7 +340,6 @@ def _apply_recipe_to_hyperparameters(
             if v is None:
                 continue
             if allowed_keys is not None and k not in allowed_keys:
-                logger.info("Skipping recipe key '%s' (not user-provided).", k)
                 continue
             final_hyperparameters[k] = str(v) if not isinstance(v, str) else v
 
diff --git a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
index da151d7165..8ad88a6c00 100644
--- a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
@@ -523,8 +523,6 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None,
         if self.stopping_condition is not None:
             create_args["stopping_condition"] = self.stopping_condition
 
-        logger.info("Final hyperparameters before submitting training job: %s", final_hyperparameters)
-
         try:
             training_job = TrainingJob.create(**create_args)
         except Exception as e:

From 50b62405e412600c4246501b07fd91aa6c9ed763 Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Tue, 14 Jul 2026 12:39:24 -0700
Subject: [PATCH 5/8] testing: added new integ test for nemotron

---
 .../train/test_rlvr_trainer_integration.py    | 73 +++++++++++++++++++
 1 file changed, 73 insertions(+)

diff --git a/sagemaker-train/tests/integ/train/test_rlvr_trainer_integration.py b/sagemaker-train/tests/integ/train/test_rlvr_trainer_integration.py
index d3d244d2b9..420f72ec91 100644
--- a/sagemaker-train/tests/integ/train/test_rlvr_trainer_integration.py
+++ b/sagemaker-train/tests/integ/train/test_rlvr_trainer_integration.py
@@ -15,11 +15,14 @@
 
 import time
 import random
+import tempfile
 import pytest
 import boto3
+import yaml
 import logging
 
 from sagemaker.core.helper.session_helper import Session
+from sagemaker.core.resources import ModelPackageGroup
 from sagemaker.train.rlvr_trainer import RLVRTrainer
 from sagemaker.train.common import TrainingType
 from sagemaker.ai_registry.evaluator import Evaluator
@@ -161,6 +164,11 @@ def test_rlvr_trainer_nova_workflow(sagemaker_session_us_east_1):
     """Test RLVR training workflow with Nova model."""
     # sagemaker_session_us_east_1 fixture is defined in conftest.py (us-east-1 region)
 
+    overrides={
+        "training_config": {
+            "lambda_concurrency_limit": 64
+        }
+    }
     unique_id = f"{int(time.time())}-{random.randint(1000, 9999)}"
     rlvr_trainer = RLVRTrainer(
         model="nova-textgeneration-lite-v2",
@@ -176,9 +184,12 @@ def test_rlvr_trainer_nova_workflow(sagemaker_session_us_east_1):
         accept_eula=True,
         sagemaker_session=sagemaker_session_us_east_1,
         base_job_name=f"rlvr-nova-integ-{unique_id}",
+        overrides=overrides,
     )
     # Workaround: Hub spec has save_steps default=null but recipe template requires an integer
     rlvr_trainer.hyperparameters.save_steps = 10
+    rlvr_trainer.hyperparameters.global_batch_size = 32
+
     training_job = rlvr_trainer.train(wait=False)
     logger.info(f"Training job submitted: {training_job.training_job_arn}")
     
@@ -279,4 +290,66 @@ def test_rlvr_trainer_with_evaluator_object(sagemaker_session, evaluator):
     assert training_job.training_job_status == "Completed"
     assert hasattr(training_job, 'output_model_package_arn')
     assert training_job.output_model_package_arn is not None
+
+
+@pytest.mark.gpu_intensive
+def test_rlvr_trainer_nemotron_with_kl_and_recipe(sagemaker_session):
+    """Test RLVR training with Nemotron model, KL regularization, recipe overrides, and hyperparameter tuning."""
+    unique_id = f"{int(time.time())}-{random.randint(1000, 9999)}"
+
+    recipe_content = {
+        "training_config": {
+            "data": {
+                "max_prompt_length": 3070,
+            },
+        }
+    }
+    with tempfile.NamedTemporaryFile(mode="w", suffix=".yaml", delete=False) as f:
+        yaml.dump(recipe_content, f)
+        recipe_path = f.name
+
+    rlvr_trainer = RLVRTrainer(
+        model="huggingface-reasoning-nvidia-nemotron-3-nano-30b-a3b-bf16",
+        model_package_group="sdk-test-finetuned-models",
+        training_dataset="s3://mc-flows-sdk-testing/input_data/rlvr-rlaif-test-data/train_285.jsonl",
+        s3_output_path="s3://mc-flows-sdk-testing/output/",
+        sagemaker_session=sagemaker_session,
+        custom_reward_function="arn:aws:sagemaker:us-west-2:729646638167:hub-content/sdktest/JsonDoc/rlvr-test-rf/0.0.1",
+        accept_eula=True,
+        base_job_name=f"rlvr-nemotron-kl-integ-{unique_id}",
+        overrides={
+            "training_config": {
+                "learning_rate": 2e-5,
+                "max_epochs": 10,
+                "train_val_split_ratio": 0.8,
+                "temperature": 1.1,
+            }
+        },
+        recipe=recipe_path,
+    )
+
+    rlvr_trainer.hyperparameters.use_kl_loss = True
+    rlvr_trainer.hyperparameters.kl_loss_coef = 0.05
+    rlvr_trainer.hyperparameters.max_epochs = 1
+    rlvr_trainer.hyperparameters.clip_ratio = 0.2
+
+    training_job = rlvr_trainer.train(wait=False)
+    logger.info(f"Training job submitted: {training_job.training_job_arn}")
+
+    max_wait_time = 7200  # 2 hour timeout for larger model
+    poll_interval = 30
+    start_time = time.time()
+
+    while time.time() - start_time < max_wait_time:
+        training_job.refresh()
+        status = training_job.training_job_status
+
+        if status in ["Completed", "Failed", "Stopped"]:
+            break
+
+        time.sleep(poll_interval)
+
+    assert training_job.training_job_status == "Completed"
+    assert hasattr(training_job, 'output_model_package_arn')
+    assert training_job.output_model_package_arn is not None
     

From 3dd951570cbea09950d47cd61c67be8c98721c82 Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Tue, 14 Jul 2026 16:00:25 -0700
Subject: [PATCH 6/8] fix: use self.compute instead of user override flag for
 apply_recipe_hyperparams fix

---
 .../src/sagemaker/train/base_trainer.py       |  28 +--
 .../src/sagemaker/train/dpo_trainer.py        |   8 +-
 .../sagemaker/train/multi_turn_rl_trainer.py  |   4 +-
 .../src/sagemaker/train/rlvr_trainer.py       |   7 +-
 .../src/sagemaker/train/sft_trainer.py        |   7 +-
 .../train/test_recipe_override_integration.py | 229 ++++++++++++++++++
 6 files changed, 244 insertions(+), 39 deletions(-)

diff --git a/sagemaker-train/src/sagemaker/train/base_trainer.py b/sagemaker-train/src/sagemaker/train/base_trainer.py
index 966bb960f4..da94cc0982 100644
--- a/sagemaker-train/src/sagemaker/train/base_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/base_trainer.py
@@ -283,32 +283,23 @@ def _get_user_provided_recipe_keys(self) -> set:
     def _apply_recipe_to_hyperparameters(
         self,
         final_hyperparameters: Dict[str, Any],
-        only_user_overrides: bool = False,
     ) -> Dict[str, Any]:
         """Apply resolved recipe values to final_hyperparameters dict.
 
         If recipe/overrides were provided, or if the user set hyperparameters
         directly via ``.hyperparameters.*``, merges resolved recipe values into
-        the hyperparameters dict. By default all leaf values from the resolved
-        recipe are applied — including keys not in the Hub spec subset. This is
-        used by the serverful/HyperPod paths, which render the values into a
-        recipe YAML file (no key-count limit).
-
+        the hyperparameters dict. All leaf values from the resolved recipe are
+        applied — including keys not in the Hub spec subset — enabling
+        power users to override any parameter in the full recipe.
         Values are converted to strings (matching the SageMaker API
         expectation for hyperparameter values).
 
+        For serverless training (``self.compute`` is None), only user-provided
+        keys (from .hyperparameters.*, recipe or overrides dict) are included because CreateTrainingJob limits HyperParameters to
+        100 members and the full resolved recipe can exceed that.
+
         Args:
             final_hyperparameters: The hyperparameters dict from to_dict().
-            only_user_overrides: When True, only apply resolved values for keys the
-                user explicitly provided — direct hyperparameter assignments, the
-                ``overrides`` dict, or a recipe file — instead of flattening the
-                entire resolved recipe. This is required for serverless SageMaker
-                training jobs, whose ``CreateTrainingJob`` ``HyperParameters`` map
-                is limited to 100 members by the CreateTrainingJob: the resolved full recipe
-                template can contain several hundred internal leaf keys the user
-                never touched. User-provided keys are not restricted to the Hub
-                spec — power users may override any recipe key, and validity is
-                enforced by ``get_resolved_recipe()``.
 
         Returns:
             The updated hyperparameters dict with recipe values applied.
@@ -319,13 +310,14 @@ def _apply_recipe_to_hyperparameters(
         try:
             resolved = self.get_resolved_recipe()
         except NoRecipeError:
-            
+
             return final_hyperparameters
 
         flat = flatten_resolved_recipe(resolved)
 
+        # Serverless (compute is None) → only user-provided keys + defaults;
         allowed_keys = None
-        if only_user_overrides:
+        if getattr(self, 'compute', None) is None:
             try:
                 allowed_keys = self._get_user_provided_recipe_keys()
             except Exception as e:
diff --git a/sagemaker-train/src/sagemaker/train/dpo_trainer.py b/sagemaker-train/src/sagemaker/train/dpo_trainer.py
index e56adde46e..1883d6f6dc 100644
--- a/sagemaker-train/src/sagemaker/train/dpo_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/dpo_trainer.py
@@ -309,12 +309,8 @@ def train(self,
         final_hyperparameters = self.hyperparameters.to_dict()
 
         # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
-        # Serverless CreateTrainingJob limits HyperParameters to 100 members, so
-        # only forward the values the user explicitly overrode rather than the
-        # entire resolved recipe template.
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            final_hyperparameters, only_user_overrides=True
-        )
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
             effective_training_dataset = training_dataset or self.training_dataset
diff --git a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
index 0c2a44e647..84b11e5860 100644
--- a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
@@ -290,10 +290,8 @@ def train(
         self._final_hyperparameters = self.hyperparameters.to_dict()
 
         # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
-        # Only forward the values the user explicitly overrode rather than the
-        # entire resolved recipe template.
         self._final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            self._final_hyperparameters, only_user_overrides=True
+            self._final_hyperparameters
         )
 
         _validate_hyperparameter_values(self._final_hyperparameters)
diff --git a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
index 8ad88a6c00..7c0dc1535c 100644
--- a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
@@ -472,12 +472,7 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None,
         final_hyperparameters = self.hyperparameters.to_dict()
 
         # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
-        # Serverless CreateTrainingJob limits HyperParameters to 100 members, so
-        # only forward the values the user explicitly overrode rather than the
-        # entire resolved recipe template.
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            final_hyperparameters, only_user_overrides=True
-        )
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
             effective_training_dataset = training_dataset or self.training_dataset
diff --git a/sagemaker-train/src/sagemaker/train/sft_trainer.py b/sagemaker-train/src/sagemaker/train/sft_trainer.py
index 7892fb8b28..ac0f59e4a0 100644
--- a/sagemaker-train/src/sagemaker/train/sft_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/sft_trainer.py
@@ -362,12 +362,7 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None, validati
         final_hyperparameters = self.hyperparameters.to_dict()
 
         # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
-        # Serverless CreateTrainingJob limits HyperParameters to 100 members, so
-        # only forward the values the user explicitly overrode rather than the
-        # entire resolved recipe template.
-        final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            final_hyperparameters, only_user_overrides=True
-        )
+        final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
             effective_training_dataset = training_dataset or self.training_dataset
diff --git a/sagemaker-train/tests/integ/train/test_recipe_override_integration.py b/sagemaker-train/tests/integ/train/test_recipe_override_integration.py
index ec1de908b1..b1867436f1 100644
--- a/sagemaker-train/tests/integ/train/test_recipe_override_integration.py
+++ b/sagemaker-train/tests/integ/train/test_recipe_override_integration.py
@@ -13,6 +13,7 @@
 """Integration tests for recipe override feature (get_resolved_recipe)."""
 from __future__ import absolute_import
 
+import logging
 import os
 import tempfile
 import time
@@ -20,8 +21,12 @@
 import pytest
 import yaml
 
+logger = logging.getLogger(__name__)
+
 from sagemaker.train.sft_trainer import SFTTrainer
+from sagemaker.train.rlvr_trainer import RLVRTrainer
 from sagemaker.train.common import TrainingType
+from sagemaker.train.recipe_resolver import flatten_resolved_recipe
 
 
 # Ensure bundled service model is available for botocore
@@ -1002,3 +1007,227 @@ def test_model_trainer_get_resolved_recipe_is_idempotent(self):
 
         finally:
             os.unlink(recipe_path)
+
+
+class TestRLVRServerlessOnlyUserOverrideKeys:
+    """Integration tests for serverless path filtering (compute=None excludes non-overridden keys)."""
+
+    def test_rlvr_serverless_only_user_override_keys_applied(self, sagemaker_session):
+        """Test that serverless path (compute=None) excludes non-overridden recipe keys.
+
+        When compute is None, CreateTrainingJob limits HyperParameters to 100 members.
+        A full recipe template can have several hundred internal leaf keys. This test
+        verifies that after applying overrides, only the user-provided keys from
+        overrides/recipe/direct assignment appear — non-overridden recipe template
+        keys must NOT leak into the result.
+        """
+        recipe_content = {
+            "training_config": {
+                "data": {
+                    "max_prompt_length": 3070,
+                },
+            }
+        }
+        with tempfile.NamedTemporaryFile(mode="w", suffix=".yaml", delete=False) as f:
+            yaml.dump(recipe_content, f)
+            recipe_path = f.name
+
+        try:
+            rlvr_trainer = RLVRTrainer(
+                model="huggingface-reasoning-nvidia-nemotron-3-nano-30b-a3b-bf16",
+                model_package_group="sdk-test-finetuned-models",
+                training_dataset="s3://mc-flows-sdk-testing/input_data/rlvr-rlaif-test-data/train_285.jsonl",
+                s3_output_path="s3://mc-flows-sdk-testing/output/",
+                sagemaker_session=sagemaker_session,
+                custom_reward_function="arn:aws:sagemaker:us-west-2:729646638167:hub-content/sdktest/JsonDoc/rlvr-test-rf/0.0.1",
+                accept_eula=True,
+                base_job_name="rlvr-override-keys-integ",
+                overrides={
+                    "training_config": {
+                        "learning_rate": 2e-5,
+                        "max_epochs": 10,
+                        "train_val_split_ratio": 0.8,
+                        "temperature": 1.1,
+                    }
+                },
+                recipe=recipe_path,
+            )
+
+            rlvr_trainer.hyperparameters.use_kl_loss = True
+            rlvr_trainer.hyperparameters.kl_loss_coef = 0.05
+
+            # Get baseline hyperparameters (the Hub spec defaults)
+            baseline_hp = rlvr_trainer.hyperparameters.to_dict()
+
+            # Serverless path (compute=None): only user-provided keys applied
+            assert rlvr_trainer.compute is None
+            result_hp = rlvr_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
+
+            # Simulate serverful path (compute set): full recipe applied
+            from sagemaker.core.training.configs import TrainingJobCompute
+            rlvr_trainer.compute = TrainingJobCompute(instance_type="ml.p5.48xlarge", instance_count=1)
+            full_hp = rlvr_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
+            rlvr_trainer.compute = None  # reset
+
+            logger.info(f"Baseline HP keys: {len(baseline_hp)}")
+            logger.info(f"User-override-only HP keys: {len(result_hp)}")
+            logger.info(f"Full recipe HP keys: {len(full_hp)}")
+
+            # Keys the user explicitly provided
+            expected_override_keys = {"learning_rate", "max_epochs", "train_val_split_ratio", "temperature"}
+            expected_recipe_keys = {"max_prompt_length"}
+            expected_direct_hp_keys = {"use_kl_loss", "kl_loss_coef"}
+            all_user_keys = expected_override_keys | expected_recipe_keys | expected_direct_hp_keys
+
+            # All user-provided keys must be present in the user-override result
+            for key in all_user_keys:
+                assert key in result_hp, (
+                    f"User-provided key '{key}' missing from serverless (compute=None) result"
+                )
+
+            # Dynamically compute recipe keys NOT in the override spec by fetching
+            # the full resolved recipe and subtracting the spec keys + user-provided keys.
+            # Only consider keys with non-None values (None keys are skipped by _apply_recipe).
+            resolved_recipe = rlvr_trainer.get_resolved_recipe()
+            flat_recipe = flatten_resolved_recipe(resolved_recipe)
+            all_recipe_keys = {k for k, v in flat_recipe.items() if v is not None}
+            override_spec_keys = set(rlvr_trainer.hyperparameters._specs.keys())
+            recipe_internal_keys_not_in_spec = all_recipe_keys - override_spec_keys - all_user_keys
+
+            logger.info(f"All recipe keys from Hub ({len(all_recipe_keys)}): {sorted(all_recipe_keys)}")
+            logger.info(f"Override spec keys ({len(override_spec_keys)}): {sorted(override_spec_keys)}")
+            logger.info(
+                f"Recipe internal keys NOT in spec ({len(recipe_internal_keys_not_in_spec)}): "
+                f"{sorted(recipe_internal_keys_not_in_spec)}"
+            )
+
+            assert len(recipe_internal_keys_not_in_spec) > 0, (
+                "Expected recipe template to have keys beyond the override spec, but found none."
+            )
+
+            # These internal recipe keys must NOT appear in the serverless result
+            leaked_keys = recipe_internal_keys_not_in_spec & set(result_hp.keys())
+            assert not leaked_keys, (
+                f"Non-overridable recipe template keys leaked into the serverless hyperparameters: "
+                f"{sorted(leaked_keys)}. These keys are not in the override spec and should be "
+                f"excluded when compute=None (serverless)."
+            )
+
+            # The full recipe (compute set) MUST contain these internal keys
+            missing_from_full = recipe_internal_keys_not_in_spec - set(full_hp.keys())
+            assert not missing_from_full, (
+                f"Full recipe (compute set) is missing expected internal keys: "
+                f"{sorted(missing_from_full)}. The full recipe path should include all template keys."
+            )
+
+            # The user-override result should not exceed the 100-key CreateTrainingJob limit
+            assert len(result_hp) <= 100, (
+                f"Serverless hyperparameters have {len(result_hp)} keys, exceeding the "
+                f"CreateTrainingJob limit of 100. Non-overridden recipe keys are leaking through."
+            )
+
+        finally:
+            os.unlink(recipe_path)
+
+    def test_sft_nova_serverless_only_user_override_keys_applied(self, sagemaker_session_us_east_1):
+        """Test that Nova SFT serverless path (compute=None) excludes non-overridden recipe keys.
+
+        Same principle as the RLVR Nemotron test: when compute is None,
+        internal recipe template keys that the user never touched must not appear
+        in the hyperparameters sent to CreateTrainingJob.
+        """
+        sft_trainer = SFTTrainer(
+            model="nova-textgeneration-lite-v2",
+            training_type=TrainingType.LORA,
+            model_package_group="sdk-test-finetuned-models",
+            training_dataset="s3://sagemaker-us-east-1-784379639078/input_data/sft-nova/sft_200_samples.jsonl",
+            s3_output_path="s3://sagemaker-us-east-1-784379639078/output/",
+            sagemaker_session=sagemaker_session_us_east_1,
+            accept_eula=True,
+            base_job_name="sft-nova-override-keys-integ",
+            overrides={
+                "training_config": {
+                    "learning_rate": 3e-5,
+                    "max_steps": 50,
+                }
+            },
+        )
+
+        sft_trainer.hyperparameters.warmup_steps = 5
+        sft_trainer.hyperparameters.weight_decay = 0.01
+
+        baseline_hp = sft_trainer.hyperparameters.to_dict()
+
+        # Serverless path (compute=None): only user-provided keys applied
+        assert sft_trainer.compute is None
+        result_hp = sft_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
+
+        # Simulate serverful path (compute set): full recipe applied
+        from sagemaker.core.training.configs import TrainingJobCompute
+        sft_trainer.compute = TrainingJobCompute(instance_type="ml.p5.48xlarge", instance_count=1)
+        full_hp = sft_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
+        sft_trainer.compute = None  # reset
+
+        logger.info(f"Nova SFT — Baseline HP keys: {len(baseline_hp)}")
+        logger.info(f"Nova SFT — User-override-only HP keys: {len(result_hp)}")
+        logger.info(f"Nova SFT — Full recipe HP keys: {len(full_hp)}")
+
+        # Keys the user explicitly provided
+        expected_override_keys = {"learning_rate", "max_steps"}
+        expected_direct_hp_keys = {"warmup_steps", "weight_decay"}
+        all_user_keys = expected_override_keys | expected_direct_hp_keys
+
+        for key in all_user_keys:
+            assert key in result_hp, (
+                f"User-provided key '{key}' missing from serverless (compute=None) result"
+            )
+
+        # Full recipe should have more keys than the user-override-only result
+        full_only_keys = set(full_hp.keys()) - set(result_hp.keys())
+        assert len(full_only_keys) > 0, (
+            "Full recipe should contain additional keys beyond the user-override-only result."
+        )
+        logger.info(
+            f"Nova SFT — Keys excluded from serverless path: "
+            f"{len(full_only_keys)} keys — {sorted(list(full_only_keys))}"
+        )
+
+        # Dynamically compute recipe keys NOT in the override spec by fetching
+        # the full resolved recipe and subtracting the spec keys + user-provided keys.
+        # Only consider keys with non-None values (None keys are skipped by _apply_recipe).
+        resolved_recipe = sft_trainer.get_resolved_recipe()
+        flat_recipe = flatten_resolved_recipe(resolved_recipe)
+        all_recipe_keys = {k for k, v in flat_recipe.items() if v is not None}
+        override_spec_keys = set(sft_trainer.hyperparameters._specs.keys())
+        recipe_internal_keys_not_in_spec = all_recipe_keys - override_spec_keys - all_user_keys
+
+        logger.info(f"Nova SFT — All recipe keys from Hub ({len(all_recipe_keys)}): {sorted(all_recipe_keys)}")
+        logger.info(f"Nova SFT — Override spec keys ({len(override_spec_keys)}): {sorted(override_spec_keys)}")
+        logger.info(
+            f"Nova SFT — Recipe internal keys NOT in spec ({len(recipe_internal_keys_not_in_spec)}): "
+            f"{sorted(recipe_internal_keys_not_in_spec)}"
+        )
+
+        assert len(recipe_internal_keys_not_in_spec) > 0, (
+            "Expected Nova recipe template to have keys beyond the override spec, but found none."
+        )
+
+        # These internal recipe keys must NOT appear in the serverless result
+        leaked_keys = recipe_internal_keys_not_in_spec & set(result_hp.keys())
+        assert not leaked_keys, (
+            f"Non-overridable Nova recipe template keys leaked into serverless hyperparameters: "
+            f"{sorted(leaked_keys)}. These keys are not in the override spec and should be "
+            f"excluded when compute=None (serverless)."
+        )
+
+        # The full recipe (compute set) MUST contain these internal keys
+        missing_from_full = recipe_internal_keys_not_in_spec - set(full_hp.keys())
+        assert not missing_from_full, (
+            f"Full recipe (compute set) is missing expected internal keys: "
+            f"{sorted(missing_from_full)}. The full recipe path should include all template keys."
+        )
+
+        assert len(result_hp) <= 100, (
+            f"Serverless hyperparameters have {len(result_hp)} keys, exceeding the "
+            f"CreateTrainingJob limit of 100."
+        )

From 1030ea2c771ca24b803ae099e68e422f455d25bf Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Tue, 14 Jul 2026 16:03:36 -0700
Subject: [PATCH 7/8] chore: undo minor indent changes in trainers

---
 sagemaker-train/src/sagemaker/train/dpo_trainer.py          | 3 +--
 .../src/sagemaker/train/multi_turn_rl_trainer.py            | 6 ++----
 sagemaker-train/src/sagemaker/train/rlvr_trainer.py         | 2 +-
 sagemaker-train/src/sagemaker/train/sft_trainer.py          | 2 +-
 4 files changed, 5 insertions(+), 8 deletions(-)

diff --git a/sagemaker-train/src/sagemaker/train/dpo_trainer.py b/sagemaker-train/src/sagemaker/train/dpo_trainer.py
index 1883d6f6dc..a83d1ed2fe 100644
--- a/sagemaker-train/src/sagemaker/train/dpo_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/dpo_trainer.py
@@ -308,8 +308,7 @@ def train(self,
 
         final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
         final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
diff --git a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
index 84b11e5860..daf3c8f5d9 100644
--- a/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/multi_turn_rl_trainer.py
@@ -289,10 +289,8 @@ def train(
 
         self._final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
-        self._final_hyperparameters = self._apply_recipe_to_hyperparameters(
-            self._final_hyperparameters
-        )
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
+        self._final_hyperparameters = self._apply_recipe_to_hyperparameters(self._final_hyperparameters)
 
         _validate_hyperparameter_values(self._final_hyperparameters)
 
diff --git a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
index 7c0dc1535c..f20b54e7b1 100644
--- a/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/rlvr_trainer.py
@@ -471,7 +471,7 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None,
 
         final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
         final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:
diff --git a/sagemaker-train/src/sagemaker/train/sft_trainer.py b/sagemaker-train/src/sagemaker/train/sft_trainer.py
index ac0f59e4a0..3f14c348f0 100644
--- a/sagemaker-train/src/sagemaker/train/sft_trainer.py
+++ b/sagemaker-train/src/sagemaker/train/sft_trainer.py
@@ -361,7 +361,7 @@ def train(self, training_dataset: Optional[Union[str, DataSet]] = None, validati
 
         final_hyperparameters = self.hyperparameters.to_dict()
 
-        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults).
+        # Apply recipe/overrides if provided (overrides > recipe > Hub defaults)
         final_hyperparameters = self._apply_recipe_to_hyperparameters(final_hyperparameters)
         # Resolve is_multimodal: auto-detect from training dataset if not explicitly set
         if self.is_multimodal is None:

From 1e883b037d6062230be2c27b7ac5b612a0196aba Mon Sep 17 00:00:00 2001
From: Syed Jafri <syedjfr@amazon.com>
Date: Tue, 14 Jul 2026 16:21:10 -0700
Subject: [PATCH 8/8] code cleanup: recipe override tests

---
 .../tests/integ/train/test_recipe_override_integration.py      | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/sagemaker-train/tests/integ/train/test_recipe_override_integration.py b/sagemaker-train/tests/integ/train/test_recipe_override_integration.py
index b1867436f1..aeb7f43f0d 100644
--- a/sagemaker-train/tests/integ/train/test_recipe_override_integration.py
+++ b/sagemaker-train/tests/integ/train/test_recipe_override_integration.py
@@ -27,6 +27,7 @@
 from sagemaker.train.rlvr_trainer import RLVRTrainer
 from sagemaker.train.common import TrainingType
 from sagemaker.train.recipe_resolver import flatten_resolved_recipe
+from sagemaker.core.training.configs import TrainingJobCompute
 
 
 # Ensure bundled service model is available for botocore
@@ -1064,7 +1065,6 @@ def test_rlvr_serverless_only_user_override_keys_applied(self, sagemaker_session
             result_hp = rlvr_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
 
             # Simulate serverful path (compute set): full recipe applied
-            from sagemaker.core.training.configs import TrainingJobCompute
             rlvr_trainer.compute = TrainingJobCompute(instance_type="ml.p5.48xlarge", instance_count=1)
             full_hp = rlvr_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
             rlvr_trainer.compute = None  # reset
@@ -1163,7 +1163,6 @@ def test_sft_nova_serverless_only_user_override_keys_applied(self, sagemaker_ses
         result_hp = sft_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
 
         # Simulate serverful path (compute set): full recipe applied
-        from sagemaker.core.training.configs import TrainingJobCompute
         sft_trainer.compute = TrainingJobCompute(instance_type="ml.p5.48xlarge", instance_count=1)
         full_hp = sft_trainer._apply_recipe_to_hyperparameters(baseline_hp.copy())
         sft_trainer.compute = None  # reset