First changes (#4723)

Co-authored-by: Bryannah Hernandez <[email protected]>

Author: bryannahm1

src/sagemaker/serve/builder/djl_builder.py

@@ -265,6 +265,7 @@ def _djl_model_builder_deploy_wrapper(self, *args, **kwargs) -> Type[PredictorBa
             # if has not been built for local container we must use cache
             # that hosting has write access to.
             self.pysdk_model.env["TRANSFORMERS_CACHE"] = "/tmp"
+            self.pysdk_model.env["HF_HOME"] = "/tmp"
             self.pysdk_model.env["HUGGINGFACE_HUB_CACHE"] = "/tmp"
 
         if "endpoint_logging" not in kwargs:

src/sagemaker/serve/builder/tei_builder.py

@@ -175,6 +175,7 @@ def _tei_model_builder_deploy_wrapper(self, *args, **kwargs) -> Type[PredictorBa
             # if has not been built for local container we must use cache
             # that hosting has write access to.
             self.pysdk_model.env["TRANSFORMERS_CACHE"] = "/tmp"
+            self.pysdk_model.env["HF_HOME"] = "/tmp"
             self.pysdk_model.env["HUGGINGFACE_HUB_CACHE"] = "/tmp"
 
         if "endpoint_logging" not in kwargs:

src/sagemaker/serve/builder/tgi_builder.py

@@ -214,6 +214,7 @@ def _tgi_model_builder_deploy_wrapper(self, *args, **kwargs) -> Type[PredictorBa
             # if has not been built for local container we must use cache
             # that hosting has write access to.
             self.pysdk_model.env["TRANSFORMERS_CACHE"] = "/tmp"
+            self.pysdk_model.env["HF_HOME"] = "/tmp"
             self.pysdk_model.env["HUGGINGFACE_HUB_CACHE"] = "/tmp"
 
         if "endpoint_logging" not in kwargs:

src/sagemaker/serve/model_server/djl_serving/server.py

@@ -19,6 +19,7 @@
 _DEFAULT_ENV_VARS = {
     "SERVING_OPTS": "-Dai.djl.logging.level=debug",
     "TRANSFORMERS_CACHE": "/opt/ml/model/",
+    "HF_HOME": "/opt/ml/model/",
     "HUGGINGFACE_HUB_CACHE": "/opt/ml/model/",
 }
 

src/sagemaker/serve/model_server/tei/server.py

@@ -18,6 +18,7 @@
 _SHM_SIZE = "2G"
 _DEFAULT_ENV_VARS = {
     "TRANSFORMERS_CACHE": "/opt/ml/model/",
+    "HF_HOME": "/opt/ml/model/",
     "HUGGINGFACE_HUB_CACHE": "/opt/ml/model/",
 }
 

src/sagemaker/serve/model_server/tgi/server.py

@@ -17,6 +17,7 @@
 _SHM_SIZE = "2G"
 _DEFAULT_ENV_VARS = {
     "TRANSFORMERS_CACHE": "/opt/ml/model/",
+    "HF_HOME": "/opt/ml/model/",
     "HUGGINGFACE_HUB_CACHE": "/opt/ml/model/",
 }
 

tests/unit/sagemaker/serve/model_server/tei/test_server.py

@@ -66,6 +66,7 @@ def test_start_invoke_destroy_local_tei_server(self, mock_requests):
             volumes={PosixPath("model_path/code"): {"bind": "/opt/ml/model/", "mode": "rw"}},
             environment={
                 "TRANSFORMERS_CACHE": "/opt/ml/model/",
+                "HF_HOME": "/opt/ml/model/",
                 "HUGGINGFACE_HUB_CACHE": "/opt/ml/model/",
                 "KEY": "VALUE",
                 "SAGEMAKER_SERVE_SECRET_KEY": "secret_key",