Support >2G model export - alternative implementation | torchlib(feat)

onnxscript/function_libs/torch_lib/graph_building.py

@@ -1,9 +1,8 @@
 """Graph building functions for torchscript graph backend."""
 from __future__ import annotations
 
+import ctypes
 import logging
-import os
-import tempfile
 import typing
 import warnings
 from typing import Any, Dict, List, Mapping, Optional, Sequence, Tuple, Union
@@ -65,9 +64,6 @@
     None,
 ]
 
-# Be sure to leave ample room for the rest of the proto fields.
-_LARGE_MODEL_SIZE_THRESHOLD = int(2**30 * 1.8)  # 1.8GB
-
 # TODO(justinchuby): Build a context manager to handle source information.
 
 
@@ -378,12 +374,38 @@ def eval_function(  # type: ignore[override]
         return self._graph.add_function_call(function, inputs, attributes)
 
 
+def _add_initializers(
+    model_proto: onnx.ModelProto, initializers: Mapping[str, torch.Tensor]
+) -> None:
+    """Add initializers to the model proto."""
+    tensor_protos = []
+
+    for name, tensor in initializers.items():
+        tensor = tensor.detach().cpu().contiguous()
+        # Take the raw data directly from the tensor to avoid the overhead of
+        # data manipulation in onnx.helper.make_tensor
+        raw_data = bytes(
+            (ctypes.c_ubyte * tensor.element_size() * tensor.numel()).from_address(
+                tensor.data_ptr()
+            )
+        )
+        tensor_proto = onnx.helper.make_tensor(
+            name=name,
+            data_type=_type_utils.JitScalarType.from_dtype(tensor.dtype).onnx_type(),
+            dims=tensor.shape,
+            vals=raw_data,
+            raw=True,
+        )
+        tensor_protos.append(tensor_proto)
+    model_proto.graph.initializer.extend(tensor_protos)
+
+
 @runtime_typing.checked
 def _add_attribute_to_torchscript_node(
     node: torch.Node,
     key: str,
     value: Union[float, int, str, bytes, Sequence[float], Sequence[int], torch.Tensor],
-):
+) -> torch.Node:
     """Initializes the right attribute based on type of value."""
     if isinstance(value, float):
         return node.f_(key, value)
@@ -443,18 +465,6 @@ def _create_op_call_in_torch_graph(
     return node_ouputs
 
 
-def _tensor_rawdata_size(tensor: torch.Tensor) -> int:
-    """Estimate the size of a tensor in bytes.
-
-    Args:
-        tensor: The tensor to estimate the size of.
-
-    Returns:
-        The estimated size of the tensor in bytes.
-    """
-    return tensor.numel() * tensor.element_size()
-
-
 def _shared_functions() -> list[onnx.FunctionProto]:
     """Hack to always include the share ops."""
 
@@ -935,16 +945,13 @@ def to_model_proto(
             # TODO(BowenBao): All local function domain versions are hardcoded as 1.
             unique_custom_domains[function_proto.domain] = 1
 
-        initializers_size = sum(
-            _tensor_rawdata_size(tensor) for tensor in self.initializers.values()
-        )
-
-        large_model = initializers_size > _LARGE_MODEL_SIZE_THRESHOLD
-
-        export_kwargs: dict[str, Any] = dict(
-            initializers=self.initializers
-            if include_initializers and not _flags.EXPERIMENTAL_INITIALIZERS_AS_INPUTS
-            else {},
+        (
+            proto,
+            _,
+            _,
+            _,
+        ) = self._torch_graph._export_onnx(  # type: ignore[attr-defined] # pylint: disable=protected-access
+            initializers={},
             onnx_opset_version=opset_version,
             dynamic_axes={},
             defer_weight_export=False,
@@ -955,46 +962,12 @@ def to_model_proto(
             add_node_names=True,
             node_attr_to_name={},
         )
-
-        # We decided to cache the model to disk when the model is large.
-        # Alternatively, we could build the ONNX `TensorProto`s in memory
-        # and append them to the model proto.
-        # We did not do it because it is harder to get right (vs. PyTorch's battle-tested
-        # implementation) and creating the `TensorProto`s naively (by converting to numpy)
-        # is slow.
-        cache_model_to_disk = large_model and include_initializers
-
-        if cache_model_to_disk:
-            with tempfile.TemporaryDirectory() as temp_dir:
-                onnx_file_path = os.path.join(temp_dir, "exported_model.onnx")
-                export_kwargs["onnx_file_path"] = onnx_file_path
-                (
-                    proto,
-                    _,
-                    _,
-                    _,
-                ) = self._torch_graph._export_onnx(  # type: ignore[attr-defined] # pylint: disable=protected-access
-                    **export_kwargs
-                )
-                onnx_model = onnx.load_from_string(proto)
-                onnx.load_external_data_for_model(onnx_model, temp_dir)
-        else:
-            (
-                proto,
-                _,
-                _,
-                _,
-            ) = self._torch_graph._export_onnx(  # type: ignore[attr-defined] # pylint: disable=protected-access
-                **export_kwargs
-            )
-            onnx_model = onnx.load_from_string(proto)
-
+        onnx_model = onnx.load_from_string(proto)
         onnx_model.functions.extend(function_proto_dict.values())
         onnx_model.functions.extend(_shared_functions())
 
         # Override value_infos with symbolic shapes.
         onnx_model = self._override_with_symbolic_value_info_proto(onnx_model)
-
         # `_export_onnx` only exports opset_imports that is visible to it. It does not
         # export opset_imports for nested functions, since it does not have access to
         # them. We manually add them back and merge with existing opset_imports in the
@@ -1015,21 +988,23 @@ def to_model_proto(
                 common_ops.common_opset.domain, common_ops.common_opset.version
             )
         )
-
         try:
-            if not cache_model_to_disk:
-                # Only check the model if it is in memory.
-                # Otherwise the checker and shape_inference will fail because
-                # we cannot serialize the model.
-                onnx_model = onnx.shape_inference.infer_shapes(
-                    onnx_model, check_type=True, strict_mode=False, data_prop=True
-                )
-                onnx.checker.check_model(onnx_model, full_check=True)
+            # Fill in the shape information before adding initializers,
+            # because the initializers may be too large (>2gb) for the model
+            # to fit in the memory.
+            onnx_model = onnx.shape_inference.infer_shapes(
+                onnx_model, check_type=True, strict_mode=False, data_prop=True
+            )
+            onnx.checker.check_model(onnx_model, full_check=True)
         except (onnx.checker.ValidationError, onnx.shape_inference.InferenceError) as e:
             warnings.warn(f"ONNX model is invalid: {e}", stacklevel=1)
             logging.debug(
                 "ONNX model:\n%s\n\nTorchScript graph:\n%s",
                 onnx.printer.to_text(onnx_model),
                 self.torch_graph,
             )
+
+        if include_initializers:
+            _add_initializers(onnx_model, self.initializers)
+
         return onnx_model