Reduce memory consumption in ONNXRT adaptor (#1266)

* reduce memory consumption

Signed-off-by: yuwenz <[email protected]>

Author: yuwenzho

neural_compressor/adaptor/onnxrt.py

@@ -712,13 +712,7 @@ def _detect_domain(self, model):
         # 2. according to input
         # typically, NLP models have multiple inputs,
         # and the dimension of each input is usually 2 (batch_size, max_seq_len)
-        if not model.is_large_model:
-            sess = ort.InferenceSession(model.model.SerializeToString(), providers=["CPUExecutionProvider"])
-        elif model.model_path is not None:  # pragma: no cover
-            sess = ort.InferenceSession(model.model_path, providers=["CPUExecutionProvider"])
-        else:  # pragma: no cover
-            assert False, "Please use model path instead of onnx model object to quantize."
-        input_shape_lens = [len(input.shape) for input in sess.get_inputs()]
+        input_shape_lens = [len(inp.type.tensor_type.shape.dim) for inp in model.model.graph.input]
         if len(input_shape_lens) > 1 and all(shape_len == 2 for shape_len in input_shape_lens):
             is_nlp = True
 
@@ -778,11 +772,15 @@ def _pre_optimize(self, model, level=1):
 
             sess_options.register_custom_ops_library(get_library_path())
         if not model.is_large_model:
-            ort.InferenceSession(model.model.SerializeToString(), sess_options, providers=["CPUExecutionProvider"])
+            sess = ort.InferenceSession(
+                model.model.SerializeToString(), sess_options, providers=["CPUExecutionProvider"]
+            )
         elif model.model_path is not None:  # pragma: no cover
-            ort.InferenceSession(model.model_path, sess_options, providers=["CPUExecutionProvider"])
+            model.model = onnx.ModelProto()  # clean memory for large model
+            sess = ort.InferenceSession(model.model_path, sess_options, providers=["CPUExecutionProvider"])
         else:  # pragma: no cover
             logger.warning("Please use model path instead of onnx model object to quantize")
+        del sess
 
         tmp_model = onnx.load(sess_options.optimized_model_filepath, load_external_data=False)
 

neural_compressor/adaptor/ox_utils/util.py

@@ -85,6 +85,8 @@
     "DmlExecutionProvider": "onnxrt_dml_ep",
 }
 
+MAXIMUM_PROTOBUF = 2147483648
+
 
 def dtype_to_name(dtype_mapping, dtype):
     """Map data type and its string representation."""

neural_compressor/model/onnx_model.py

@@ -18,8 +18,10 @@
 
 import logging
 import os
+import sys
 from pathlib import Path
 
+from neural_compressor.adaptor.ox_utils.util import MAXIMUM_PROTOBUF
 from neural_compressor.model.base_model import BaseModel
 from neural_compressor.utils.utility import LazyImport
 
@@ -41,16 +43,9 @@ def __init__(self, model, **kwargs):
         """
         self._model = model if not isinstance(model, str) else onnx.load(model)
         self._model_path = None if not isinstance(model, str) else model
-        self._is_large_model = False
-        try:
-            ort.InferenceSession(self._model.SerializeToString(), providers=["CPUExecutionProvider"])
-        except Exception as e:  # pragma: no cover
-            if self._model_path is not None:
-                ort.InferenceSession(self._model_path, providers=["CPUExecutionProvider"])
-                self._is_large_model = True
-            else:
-                logger.warning("Please use model path instead of onnx model object to quantize")
-
+        self._is_large_model = self.check_large_model()
+        if self._is_large_model and self._model_path is None:
+            logger.warning("Model size > 2GB. Please use model path instead of onnx model object to quantize")
         self._config = None
         if isinstance(model, str) and os.path.exists(Path(model).parent.joinpath("config.json").as_posix()):
             from transformers import PretrainedConfig
@@ -66,6 +61,26 @@ def __init__(self, model, **kwargs):
         self._get_graph_info()
         self._q_config = None
 
+    def check_large_model(self):
+        """Check model > 2GB."""
+        init_size = 0
+        for init in self._model.graph.initializer:
+            # if initializer has external data location, return True
+            if init.HasField("data_location") and init.data_location == onnx.TensorProto.EXTERNAL:
+                return True
+            # if raise error of initializer size > 2GB, return True
+            try:
+                init_bytes = init.SerializeToString()
+                init_size += sys.getsizeof(init_bytes)
+            except Exception as e:
+                if "exceeds maximum protobuf size of 2GB" in str(e):
+                    return True
+                else:  # pragma: no cover
+                    raise e
+            if init_size > MAXIMUM_PROTOBUF:
+                return True
+        return False
+
     @property
     def is_large_model(self):
         """Check the onnx model is over 2GB."""

test/model/test_onnx_model.py

@@ -203,6 +203,7 @@ def setUp(self):
     def tearDownClass(self):
         shutil.rmtree("./gptj", ignore_errors=True)
         shutil.rmtree("./hf_test", ignore_errors=True)
+        os.remove("model.onnx")
 
     def test_hf_model(self):
         from optimum.onnxruntime import ORTModelForCausalLM
@@ -407,6 +408,53 @@ def test_remove_unused_nodes(self):
         self.model.remove_unused_nodes()
         self.assertEqual(len(self.model.nodes()), 6)
 
+    def test_check_large_model(self):
+        import onnx
+        import torch
+        import torch.nn as nn
+
+        from neural_compressor.model.onnx_model import ONNXModel
+
+        class Net(nn.Module):
+            def __init__(self, in_features, out_features):
+                super(Net, self).__init__()
+                self.fc = nn.Linear(in_features, out_features)
+
+            def forward(self, x):
+                x = self.fc(x)
+                return x
+
+        # model > 2GB
+        model = Net(512, 1024 * 1024)
+        input = torch.randn(512, requires_grad=True)
+        with torch.no_grad():
+            torch.onnx.export(model, (input,), "model.onnx", do_constant_folding=True, opset_version=13)
+        model = onnx.load("model.onnx")
+        model = ONNXModel(model)  # pass ModelProto
+        self.assertTrue(model.check_large_model())
+
+        model = ONNXModel("model.onnx")  # pass string
+        self.assertTrue(model.check_large_model())
+
+        model = onnx.load("model.onnx", load_external_data=False)  # not load init
+        model = ONNXModel(model)
+        self.assertTrue(model.check_large_model())
+
+        # model < 2GB
+        model = Net(10, 10 * 10)
+        input = torch.randn(10, requires_grad=True)
+        with torch.no_grad():
+            torch.onnx.export(model, (input,), "model.onnx", do_constant_folding=True, opset_version=13)
+        model = onnx.load("model.onnx")
+        model = ONNXModel(model)  # pass ModelProto
+        self.assertFalse(model.check_large_model())
+
+        model = ONNXModel("model.onnx")  # pass string
+        self.assertFalse(model.check_large_model())
+
+        model = ONNXModel("model.onnx", load_external_data_for_model=False)  # not load init
+        self.assertFalse(model.check_large_model())
+
 
 if __name__ == "__main__":
     unittest.main()