Enhance model builder selection logic to include model size

requirements/extras/huggingface_requirements.txt

@@ -0,0 +1,7 @@
+accelerate
+numpy>=1.17
+packaging>=20.0
+psutil
+pyyaml
+torch>=1.10.0
+huggingface_hub

requirements/extras/test_requirements.txt

@@ -39,3 +39,10 @@ tritonclient[http]<2.37.0
 onnx==1.14.1
 # tf2onnx==1.15.1
 nbformat>=5.9,<6
+accelerate
+numpy>=1.17
+packaging>=20.0
+psutil
+pyyaml
+torch>=1.10.0
+huggingface_hub

setup.py

@@ -79,6 +79,7 @@ def read_requirements(filename):
     "feature-processor": read_requirements(
         "requirements/extras/feature-processor_requirements.txt"
     ),
+    "huggingface": read_requirements("requirements/extras/huggingface_requirements.txt"),
 }
 # Meta dependency groups
 extras["all"] = [item for group in extras.values() for item in group]

src/sagemaker/serve/builder/model_builder.py

@@ -20,7 +20,9 @@
 
 from pathlib import Path
 
+from accelerate.commands.estimate import estimate_command_parser, gather_data
 from sagemaker import Session
+from sagemaker.djl_inference import defaults
 from sagemaker.model import Model
 from sagemaker.base_predictor import PredictorBase
 from sagemaker.serializers import NumpySerializer, TorchTensorSerializer
@@ -39,6 +41,7 @@
 from sagemaker.serve.save_retrive.version_1_0_0.metadata.metadata import Metadata
 from sagemaker.serve.spec.inference_spec import InferenceSpec
 from sagemaker.serve.utils.predictors import _get_local_mode_predictor
+from sagemaker.serve.utils.hardware_detector import _get_gpu_info, _get_gpu_info_fallback
 from sagemaker.serve.detector.image_detector import (
     auto_detect_container,
     _detect_framework_and_version,
@@ -65,6 +68,8 @@
     ModelServer.DJL_SERVING,
 }
 
+MIB_CONVERSION_FACTOR = 0.00000095367431640625
+MEMORY_BUFFER_MULTIPLIER = 1.2  # 20% buffer
 
 # pylint: disable=attribute-defined-outside-init
 @dataclass
@@ -567,7 +572,7 @@ def wrapper(*args, **kwargs):
     # It supports two modes of deployment
     # 1/ SageMaker Endpoint
     # 2/ Local launch with container
-    def build(
+    def build(  # pylint: disable=R0911
         self,
         mode: Type[Mode] = None,
         role_arn: str = None,
@@ -616,6 +621,10 @@ def build(
                 )
                 if hf_model_md.get("pipeline_tag") == "text-generation":  # pylint: disable=R1705
                     return self._build_for_tgi()
+                elif self.can_fit_on_single_gpu():
+                    return self._build_for_transformers()
+                elif self.model in defaults.FASTER_TRANSFORMER_SUPPORTED_ARCHITECTURES:
+                    return self._build_for_djl()
                 else:
                     return self._build_for_transformers()
 
@@ -672,3 +681,58 @@ def validate(self, model_dir: str) -> Type[bool]:
         """
 
         return get_metadata(model_dir)
+
+    def total_inference_model_size_mib(self):
+        """Calculates the model size from HF accelerate
+
+        This function gets the model size from accelerate. It also adds a
+        padding and converts to size MiB. When performing inference, expect
+        to add up to an additional 20% to the given model size as found by EleutherAI.
+        """
+        dtypes = "float32"
+        try:
+            if self.env_vars.get("dtypes"):
+                dtypes = self.env_vars.get("dtypes")
+
+            parser = estimate_command_parser()
+            args = parser.parse_args([self.model, "--dtypes", dtypes])
+        except ValueError:
+            logging.error("Args specified incorrect for model %s", self.model)
+
+        output = gather_data(
+            args
+        )  # "dtype", "Largest Layer", "Total Size Bytes", "Training using Adam"
+
+        total_memory_size_mib = MEMORY_BUFFER_MULTIPLIER * output[0][2] * MIB_CONVERSION_FACTOR
+        logger.info("Total memory size MIB: %s", total_memory_size_mib)
+        return total_memory_size_mib
+
+    def can_fit_on_single_gpu(self):
+        """Check if model can fit on a single GPU
+
+        This function gets the GPU info or fallback to set the size of a single GPU.
+        If the size of the model is <= single gpu memory size, returns true.
+        """
+        try:
+            gpu_info = _get_gpu_info(self.instance_type, self.sagemaker_session)
+            logger.info("GPU info %s for instance %s", gpu_info, self.instance_type)
+            single_gpu_size_mib = gpu_info[1] / gpu_info[0]
+        except ValueError:
+            gpu_fallback = _get_gpu_info_fallback(
+                self.instance_type, self.sagemaker_session.boto_region_name
+            )
+            logger.info("GPU fallback picked up %s", gpu_fallback)
+            single_gpu_size_mib = gpu_fallback[1] / gpu_fallback[0]
+
+        if single_gpu_size_mib is None:
+            logger.info("Unable to determine single GPU size for instance %s", self.instance_type)
+            return False
+
+        if self.total_inference_model_size_mib() <= single_gpu_size_mib:
+            logger.info(
+                "Total inference model size MIB %s, single GPU size for instance MIB %s",
+                self.total_inference_model_size_mib(),
+                single_gpu_size_mib,
+            )
+            return True
+        return False

tests/integ/sagemaker/serve/test_model_builder_gpu.py → tests/integ/sagemaker/serve/test_serve_model_builder_gpu.py

@@ -13,9 +13,8 @@
 from __future__ import absolute_import
 
 import pytest
-from sagemaker.serve import Mode
-from sagemaker.serve.builder.model_builder import ModelBuilder
 from sagemaker.serve.builder.schema_builder import SchemaBuilder
+from sagemaker.serve.builder.model_builder import ModelBuilder, Mode
 from tests.integ.sagemaker.serve.constants import (
     HF_DIR,
     PYTHON_VERSION_IS_NOT_310,
@@ -90,10 +89,10 @@ def model_builder(request):
 def test_non_text_generation_model_single_GPU(sagemaker_session, model_builder, model_input):
     iam_client = sagemaker_session.boto_session.client("iam")
     role_arn = iam_client.get_role(RoleName="SageMakerRole")["Role"]["Arn"]
+    model = model_builder.build(role_arn=role_arn, sagemaker_session=sagemaker_session)
     caught_ex = None
     with timeout(minutes=SERVE_SAGEMAKER_ENDPOINT_TIMEOUT):
         try:
-            model = model_builder.build(role_arn=role_arn, sagemaker_session=sagemaker_session)
             logger.info("Running in SAGEMAKER_ENDPOINT mode")
             predictor = model.deploy(
                 mode=Mode.SAGEMAKER_ENDPOINT,
@@ -137,9 +136,9 @@ def test_non_text_generation_model_multi_GPU(sagemaker_session, model_builder, m
     iam_client = sagemaker_session.boto_session.client("iam")
     role_arn = iam_client.get_role(RoleName="SageMakerRole")["Role"]["Arn"]
     caught_ex = None
+    model = model_builder.build(role_arn=role_arn, sagemaker_session=sagemaker_session)
     with timeout(minutes=SERVE_SAGEMAKER_ENDPOINT_TIMEOUT):
         try:
-            model = model_builder.build(role_arn=role_arn, sagemaker_session=sagemaker_session)
             logger.info("Running in SAGEMAKER_ENDPOINT mode")
             predictor = model.deploy(
                 mode=Mode.SAGEMAKER_ENDPOINT,

tests/integ/sagemaker/serve/test_serve_transformers.py

@@ -106,7 +106,7 @@ def test_pytorch_transformers_sagemaker_endpoint(
     with timeout(minutes=SERVE_SAGEMAKER_ENDPOINT_TIMEOUT):
         try:
             logger.info("Deploying and predicting in SAGEMAKER_ENDPOINT mode...")
-            predictor = model.deploy(instance_type=gpu_instance_type, initial_instance_count=1)
+            predictor = model.deploy(instance_type="ml.g4dn.xlarge", initial_instance_count=1)
             logger.info("Endpoint successfully deployed.")
             predictor.predict(input)
         except Exception as e: