test: disable HF Xet/hf_transfer in serve integ tests to fix endpoint deploy (#5992)

* test: disable HF Xet/hf_transfer in serve integ tests to fix endpoint deploy

The schema_builder and serve_tei SAGEMAKER_ENDPOINT tests began failing
around 2026-06-29 because HuggingFace Hub migrated model repos to the Xet
storage backend. The HF DLC container downloads the model at startup with
hf_transfer enabled, which now hits the Xet CDN (us.aws.cdn.hf.co) and gets
403 Forbidden. The container never passes the ping health check, the endpoint
enters Failed, and the test hits its 15-minute deploy timeout.

Inject HF_HUB_ENABLE_HF_TRANSFER=0 and HF_HUB_DISABLE_XET=1 into the container
env so the model is pulled over plain HTTPS, restoring the pre-Xet download
path. Evidence: endpoint container logs show hf_transfer 403 against the Xet
bridge URL; the same tests passed in build #2712 (2026-06-16) before the
rollout.

* test: Mock JumpStartModel in optimize/compile JS builder test

Author: lucasjia-aws

tests/integ/sagemaker/serve/test_schema_builder.py

@@ -92,6 +92,13 @@ def test_model_builder_happy_path_with_task_provided_local_schema_mode(
         model_metadata={"HF_TASK": task_provided},
         instance_type=instance_type_provided,
         sagemaker_session=sagemaker_session,
+        env_vars={
+            # Force plain HTTPS model download in the container. HF Hub repos are now
+            # backed by Xet storage, and the DLC's hf_transfer path returns 403 against
+            # the Xet CDN, so the container fails the ping health check on startup.
+            "HF_HUB_ENABLE_HF_TRANSFER": "0",
+            "HF_HUB_DISABLE_XET": "1",
+        },
     )
 
     model = model_builder.build(sagemaker_session=sagemaker_session)
@@ -169,6 +176,13 @@ def test_model_builder_happy_path_with_task_provided_remote_schema_mode(
         model_metadata={"HF_TASK": task_provided},
         instance_type=instance_type_provided,
         sagemaker_session=sagemaker_session,
+        env_vars={
+            # Force plain HTTPS model download in the container. HF Hub repos are now
+            # backed by Xet storage, and the DLC's hf_transfer path returns 403 against
+            # the Xet CDN, so the container fails the ping health check on startup.
+            "HF_HUB_ENABLE_HF_TRANSFER": "0",
+            "HF_HUB_DISABLE_XET": "1",
+        },
     )
     model = model_builder.build(sagemaker_session=sagemaker_session)
 

tests/integ/sagemaker/serve/test_serve_tei.py

@@ -47,7 +47,12 @@ def model_builder_model_schema_builder(sagemaker_session):
         schema_builder=SchemaBuilder(sample_input, loaded_response),
         env_vars={
             # Add this to bypass JumpStart model mapping
-            "HF_MODEL_ID": "BAAI/bge-m3"
+            "HF_MODEL_ID": "BAAI/bge-m3",
+            # Force plain HTTPS model download in the container. HF Hub repos are now
+            # backed by Xet storage, and the DLC's hf_transfer path returns 403 against
+            # the Xet CDN, so the container fails the ping health check on startup.
+            "HF_HUB_ENABLE_HF_TRANSFER": "0",
+            "HF_HUB_DISABLE_XET": "1",
         },
     )
 

tests/unit/sagemaker/serve/builder/test_js_builder.py

@@ -1272,12 +1272,14 @@ def test_jumpstart_model_provider_calls_jumpstart_speculative_decoding(
 
         mock_js_speculative_decoding.assert_called_once()
 
+    @patch("sagemaker.serve.builder.jumpstart_builder.JumpStartModel")
     @patch("sagemaker.serve.builder.jumpstart_builder._capture_telemetry", side_effect=None)
     @patch.object(ModelBuilder, "_get_serve_setting", autospec=True)
     def test_optimize_quantize_and_compile_for_jumpstart(
         self,
         mock_serve_settings,
         mock_telemetry,
+        mock_js_model,
     ):
         mock_sagemaker_session = Mock()
         mock_metadata_config = Mock()
@@ -1319,6 +1321,10 @@ def test_optimize_quantize_and_compile_for_jumpstart(
 
         model_builder.pysdk_model = mock_pysdk_model
 
+        # Avoid a real JumpStartModel construction (which would hit S3 for the
+        # JumpStart manifest); return a deterministic env for neuron lookup.
+        mock_js_model.return_value.env = {"SAGEMAKER_ENV": "1"}
+
         out_put = model_builder._optimize_for_jumpstart(
             accept_eula=True,
             quantization_config={