datamixing recipe path fix

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

@@ -480,16 +480,18 @@ def build_hyperpod_datamix_recipe_from_context(
     4. Inject customer_data_percent and nova_data_percentages into data_mixing.sources
     5. Validate non-zero categories exist in template's nova_data section
     6. Write final YAML to HyperPod CLI recipes directory
-    7. Return (relative_recipe_path, image_uri)
+    7. Return (recipe_path, image_uri)
 
     Args:
         context: The HyperPodTemplateContext from resolve_hyperpod_datamix_context.
         validated_config: A DataMixingConfig validated via validate_data_mixing_categories.
 
     Returns:
-        Tuple of (relative_recipe_path, image_uri). relative_recipe_path is relative
-        to the HyperPod CLI recipes_collection/recipes directory with .yaml extension
-        removed. image_uri is the container image URI from the context (or None).
+        Tuple of (recipe_path, image_uri). recipe_path is the absolute filesystem
+        path (including the .yaml extension) of the generated recipe written under
+        the HyperPod CLI recipes_collection/recipes directory; it is consumed by the
+        recipe resolver as a user recipe file. image_uri is the container image URI
+        from the context (or None).
 
     Raises:
         RuntimeError: If hyperpod_cli is not installed.
@@ -652,18 +654,11 @@ def _apply_overrides(recipe: dict, overrides: dict) -> dict:
     with open(recipe_path, "w") as f:
         f.write(recipe_output)
 
-    relative_path = (
-        recipe_path.split(HYPERPOD_RECIPE_PATH, 1)[1]
-        .lstrip("/").lstrip("\\")
-        .removesuffix(".yaml")
-    )
-
     logger.info(
-        "Generated HyperPod datamix recipe at '%s' (relative: '%s') from context '%s'.",
+        "Generated HyperPod datamix recipe at '%s' from context '%s'.",
         recipe_path,
-        relative_path,
         context.recipe_name,
     )
 
-    return relative_path, context.image_uri
+    return recipe_path, context.image_uri
 

sagemaker-train/tests/integ/train/test_sft_data_mixing_hyperpod.py

@@ -36,6 +36,7 @@
 import pytest
 from sagemaker.train.sft_trainer import SFTTrainer
 from sagemaker.train.common import TrainingType
+from sagemaker.train.base_trainer import BaseTrainer
 from sagemaker.train.data_mixing_config import DataMixingConfig
 from sagemaker.core.training.configs import HyperPodCompute
 
@@ -49,8 +50,8 @@
 REGION = "us-east-1"
 DATA_PREFIX = "test-sft-data-mixing-hyperpod-integ"
 NUM_TRAINING_SAMPLES = 300
-HYPERPOD_CLUSTER_NAME = "riv-rig"
-HYPERPOD_INSTANCE_TYPE = "ml.p5.48xlarge"
+HYPERPOD_CLUSTER_NAME = "pysdk-hp-integ-tests"
+HYPERPOD_INSTANCE_TYPE = "ml.g6.12xlarge"
 
 
 def _generate_training_data() -> str:
@@ -125,10 +126,10 @@ def training_resources(sagemaker_session_us_east_1):
 @pytest.mark.gpu_intensive
 @pytest.mark.us_east_1
 def test_sft_trainer_nova_micro_data_mixing_hyperpod(sagemaker_session_us_east_1, training_resources):
-    """Test SFT trainer with Nova Lite 2 model and data mixing on HyperPod.
+    """Test SFT trainer with Nova Micro model and data mixing on HyperPod.
 
     This end-to-end test submits a real HyperPod training job with DataMixingConfig
-    for Nova Lite 2. The SDK resolves the datamix recipe from SageMaker Hub, validates
+    for Nova Micro. The SDK resolves the datamix recipe from SageMaker Hub, validates
     categories, and includes the serialized config in the HyperPod override parameters.
     """
     unique_id = f"{int(time.time())}-{random.randint(1000, 9999)}"
@@ -184,4 +185,33 @@ def test_sft_trainer_nova_micro_data_mixing_hyperpod(sagemaker_session_us_east_1
     assert get_job_result.returncode == 0, (
         f"hyperpod get-job failed for '{job_name}': {get_job_result.stderr}"
     )
-    logger.info(f"Verified job '{job_name}' exists on the cluster.")
+    logger.info(f"Verified job '{job_name}' exists on the cluster using hp-cli.")
+
+    # Poll for job completion by checking for the manifest in S3.
+    # The manifest is written under {output_s3_path}/{job_name}/manifest.json
+    # once training finishes, so its presence confirms end-to-end completion.
+
+    output_s3_path = training_resources["s3_output_path"]
+    max_wait_time = 21600  # 6 hour timeout (HyperPod jobs can take longer)
+    poll_interval = 60  # Check every 60 seconds
+    start_time = time.time()
+    checkpoint_path = None
+
+    while time.time() - start_time < max_wait_time:
+        checkpoint_path = BaseTrainer._resolve_checkpoint_from_manifest(
+            job_name=job_name,
+            output_s3_path=output_s3_path,
+            sagemaker_session=sagemaker_session_us_east_1,
+        )
+        if checkpoint_path:
+            logger.info(f"Checkpoint resolved: {checkpoint_path}")
+            break
+
+        elapsed = int(time.time() - start_time)
+        logger.info(f"Waiting for manifest... ({elapsed}s elapsed)")
+        time.sleep(poll_interval)
+
+    assert checkpoint_path is not None, (
+        f"Job {job_name} did not produce a manifest within {max_wait_time}s"
+    )
+    logger.info(f"Training complete. Checkpoint: {checkpoint_path}")

sagemaker-train/tests/unit/train/common_utils/test_data_mixing_utils.py

@@ -13,6 +13,8 @@
 """Unit tests for data mixing utility functions."""
 from __future__ import absolute_import
 
+import os
+
 import pytest
 
 from sagemaker.train.data_mixing_config import DataMixingConfig
@@ -823,8 +825,12 @@ def capture_write(path, mode="r", **kwargs):
         assert parsed["data_mixing"]["sources"]["nova_data"]["en-entertainment"] == 20
         assert parsed["data_mixing"]["sources"]["nova_data"]["en-scientific"] == 10
 
-    def test_return_value_is_relative_path_and_image_uri(self):
-        """Return value should be (relative_path_without_extension, image_uri)."""
+    def test_return_value_is_absolute_recipe_path_and_image_uri(self):
+        """Return value should be (absolute_recipe_path_with_extension, image_uri).
+
+        The path must be a loadable filesystem path (absolute, .yaml extension
+        intact) because the recipe resolver opens it as a user recipe file.
+        """
         from unittest.mock import patch, MagicMock, mock_open
 
         from sagemaker.train.common_utils.data_mixing_utils import (
@@ -843,12 +849,14 @@ def test_return_value_is_relative_path_and_image_uri(self):
                 with patch("builtins.open", m_open):
                     result = build_hyperpod_datamix_recipe_from_context(context, config)
 
-        relative_path, image_uri = result
+        recipe_path, image_uri = result
 
-        # relative_path should not end with .yaml
-        assert not relative_path.endswith(".yaml")
-        # relative_path should contain fine-tuning/nova
-        assert "fine-tuning/nova" in relative_path
+        # recipe_path must be an absolute filesystem path
+        assert os.path.isabs(recipe_path)
+        # recipe_path must retain the .yaml extension so it can be opened
+        assert recipe_path.endswith(".yaml")
+        # recipe_path should live under the HyperPod CLI recipes fine-tuning/nova dir
+        assert "fine-tuning/nova" in recipe_path
         # image_uri should match the context
         assert image_uri == self.IMAGE_URI