Doc update and added SFT integ test (#6018)

* Doc update and added SFT integ test

- Moved Nova-only CPT notebook from open-weight section to Nova section.
- Renamed "Customizing Models" → "Model Customization (Foundation Models)".
- Consolidated deployment docs under "Deploy Models for Inference" (including Bedrock).
- Added SFT SMTJ serverful integ test

* Update integ test

* Update instance type and model

Author: zhaoqizqwang

docs/inference/deploy_finetuned.rst

@@ -0,0 +1,10 @@
+Deploying Fine-Tuned Models
+============================
+
+Deploy models that have been fine-tuned through the :doc:`Model Customization <../model_customization/index>` workflow to either a SageMaker real-time endpoint or Amazon Bedrock.
+
+.. toctree::
+   :maxdepth: 1
+
+   Deploy to SageMaker Endpoint <../model_customization/deploy_sagemaker_endpoint>
+   Deploy to Amazon Bedrock <../../v3-examples/model-customization-examples/bedrock-modelbuilder-deployment>

docs/inference/index.rst

@@ -508,6 +508,12 @@ V3 Package Structure
      - Pipelines, processing, model registry, monitoring, Clarify
 
 
+.. toctree::
+   :maxdepth: 2
+
+   deploy_finetuned
+
+
 Inference Examples
 -----------------
 

docs/model_customization/deployment.rst

@@ -1,5 +1,5 @@
-Customizing Models
-===================
+Model Customization (Foundation Models)
+=========================================
 
 .. raw:: html
 
@@ -23,4 +23,3 @@ Key Benefits
    open_weight_model_customization
    nova
    evaluation
-   deployment

docs/model_customization/index.rst

@@ -24,8 +24,9 @@ Trainer Classes
 **SFTTrainer (Supervised Fine-Tuning)**
   Traditional fine-tuning with labeled datasets for task-specific adaptation
 
-**CPTTrainer (Continued Pre-Training)**
-  Continue pre-training on a raw corpus to extend model knowledge in a specific domain
+**CPTTrainer (Continued Pre-Training)** *(Nova models only)*
+  Continue pre-training on a raw corpus to extend model knowledge in a specific domain.
+  See the :doc:`Nova section <nova>` for CPT examples.
 
 **DPOTrainer (Direct Preference Optimization)**
   Fine-tune models using human preference data without reinforcement learning complexity
@@ -62,6 +63,5 @@ Key Features
    DPOTrainer Finetuning <../../v3-examples/model-customization-examples/dpo_trainer_example_notebook_v3_prod>
    RLAIF Finetuning <../../v3-examples/model-customization-examples/rlaif_finetuning_example_notebook_v3_prod>
    RLVR Finetuning <../../v3-examples/model-customization-examples/rlvr_finetuning_example_notebook_v3_prod>
-   CPT Training on HyperPod <../../v3-examples/model-customization-examples/cpt_data_mixing_hyperpod>
    Fine-Tuning with Serverful Training Jobs <finetuning_serverful>
    Fine-Tuning with HyperPod <finetuning_hyperpod>

docs/model_customization/model_customization.rst

@@ -13,4 +13,5 @@ Adapt Amazon Nova foundation models to your specific use cases through fine-tuni
    :caption: Nova Customization Guides
 
    nova_data_mixing
+      CPT Training on HyperPod <../../v3-examples/model-customization-examples/cpt_data_mixing_hyperpod>
    ../../v3-examples/model-customization-examples/sm-studio-nova-training-job-sample-notebook

docs/model_customization/nova.rst

@@ -0,0 +1,164 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+"""Integration test for SFT fine-tuning with serverful Training Jobs (SMTJ).
+
+Based on: v3-examples/model-customization-examples/sft_finetuning_serverful_smtj.ipynb
+
+Prerequisites:
+    - Active AWS credentials in us-east-1 with permissions to launch SageMaker
+      Training Jobs and access the default SageMaker bucket.
+
+The test is self-sufficient: it uses the account's default SageMaker bucket
+(``sagemaker-{region}-{account_id}``, auto-created by ``Session.default_bucket()``).
+The sample training data in ``tests/data/train/sft_smtj_sample_data.jsonl`` is
+uploaded under the ``sft-smtj-integ/`` prefix if not already present.
+
+Run with:
+    export AWS_DEFAULT_REGION=us-east-1
+    pytest tests/integ/train/test_sft_trainer_serverful_smtj.py -v -s
+"""
+from __future__ import absolute_import
+
+import logging
+import os
+import time
+import random
+
+import pytest
+
+from sagemaker.train import SFTTrainer
+from sagemaker.train.common import TrainingType
+from sagemaker.core.training.configs import TrainingJobCompute
+
+logger = logging.getLogger(__name__)
+
+# Test configuration
+DATA_PREFIX = "sft-smtj-integ"
+DATA_S3_KEY = f"{DATA_PREFIX}/sft_smtj_sample_data.jsonl"
+
+# Local sample data file
+DATA_DIR = os.path.join(os.path.dirname(__file__), "..", "..", "data", "train")
+LOCAL_TRAINING_DATA = os.path.join(DATA_DIR, "sft_smtj_sample_data.jsonl")
+
+
+def _ensure_training_data(s3_client, bucket_name: str) -> str:
+    """Upload local training data to S3 if not already present.
+
+    Returns the S3 URI for the training data file.
+    """
+    s3_uri = f"s3://{bucket_name}/{DATA_S3_KEY}"
+
+    # Check if already exists
+    try:
+        s3_client.head_object(Bucket=bucket_name, Key=DATA_S3_KEY)
+        logger.info(f"Training data already exists at {s3_uri}")
+        return s3_uri
+    except s3_client.exceptions.ClientError:
+        pass
+
+    logger.info(f"Uploading training data from {LOCAL_TRAINING_DATA} to {s3_uri}")
+    s3_client.upload_file(LOCAL_TRAINING_DATA, bucket_name, DATA_S3_KEY)
+    logger.info("Training data uploaded successfully.")
+    return s3_uri
+
+
+@pytest.fixture(scope="module")
+def training_resources(sagemaker_session_us_east_1):
+    """Ensure training data exists in the account's default SageMaker bucket.
+
+    Uses ``Session.default_bucket()`` (creates ``sagemaker-{region}-{account_id}``
+    if needed), then uploads the local sample data file under the
+    ``sft-smtj-integ/`` prefix if not already present.
+
+    Returns a dict with:
+        - training_dataset: S3 URI to the training data file
+        - s3_output_path: S3 URI for training output
+        - bucket_name: the default bucket name
+    """
+    bucket_name = sagemaker_session_us_east_1.default_bucket()
+    s3_client = sagemaker_session_us_east_1.boto_session.client("s3")
+    training_dataset = _ensure_training_data(s3_client, bucket_name)
+
+    return {
+        "training_dataset": training_dataset,
+        "s3_output_path": f"s3://{bucket_name}/{DATA_PREFIX}/output/",
+        "bucket_name": bucket_name,
+    }
+
+
+@pytest.mark.gpu_intensive
+@pytest.mark.us_east_1
+def test_sft_trainer_serverful_smtj(sagemaker_session_us_east_1, training_resources):
+    """Test SFT fine-tuning on serverful compute with recipe overrides.
+
+    Exercises the full notebook flow:
+    1. Create SFTTrainer with TrainingJobCompute (dedicated instances)
+    2. Apply recipe overrides and verify via get_resolved_recipe()
+    3. Set hyperparameters and submit training job
+    4. Poll until completion
+    """
+    unique_id = f"{int(time.time())}-{random.randint(1000, 9999)}"
+
+    sft_trainer = SFTTrainer(
+        model="amazon.nova-micro-v1",
+        training_type=TrainingType.LORA,
+        training_dataset=training_resources["training_dataset"],
+        s3_output_path=training_resources["s3_output_path"],
+        compute=TrainingJobCompute(
+            instance_type="ml.g6.48xlarge",
+            instance_count=1,
+        ),
+        sagemaker_session=sagemaker_session_us_east_1,
+        overrides={"training_config": {"max_epochs": 1}},
+        base_job_name=f"sft-smtj-integ-{unique_id}",
+    )
+
+    # Verify recipe overrides are resolved correctly
+    resolved = sft_trainer.get_resolved_recipe()
+    training_config = resolved["training_config"]
+    logger.info(f"Resolved training_config: {training_config}")
+
+    # Nova Micro uses trainer.max_epochs for step control
+    assert training_config["trainer"]["max_epochs"] == 1, (
+        f"Expected max_epochs=1, got: {training_config.get('trainer')}"
+    )
+
+    # Submit (non-blocking)
+    training_job = sft_trainer.train(wait=False)
+    assert training_job is not None
+    assert training_job.training_job_name is not None
+    logger.info(f"Training job submitted: {training_job.training_job_name}")
+
+    # Poll for completion
+    max_wait_time = 10800  # 3 hour timeout (Nova training can be slow)
+    poll_interval = 60
+    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
+
+        logger.info(
+            f"Job {training_job.training_job_name} status: {status} "
+            f"({int(time.time() - start_time)}s elapsed)"
+        )
+        time.sleep(poll_interval)
+
+    assert training_job.training_job_status == "Completed", (
+        f"Training job {training_job.training_job_name} ended with status: "
+        f"{training_job.training_job_status}"
+    )
+    logger.info(f"Training job completed successfully: {training_job.training_job_name}")