🏪 Support opversion conversion for models with external data (#1322)

## Describe your changes

## Checklist before requesting a review
- [ ] Add unit tests for this change.
- [ ] Make sure all tests can pass.
- [ ] Update documents if necessary.
- [ ] Lint and apply fixes to your code by running `lintrunner -a`
- [ ] Is this a user-facing change? If yes, give a description of this
change to be included in the release notes.
- [ ] Is this PR including examples changes? If yes, please remember to
update [example
documentation](https://github.com/microsoft/Olive/blob/main/docs/source/examples.md)
in a follow-up PR.

## (Optional) Issue link

.pre-commit-config.yaml

@@ -25,7 +25,7 @@ repos:
         exclude: examples/
   - repo: https://github.com/astral-sh/ruff-pre-commit
     # Ruff version.
-    rev: v0.4.5
+    rev: v0.6.0
     hooks:
       - id: ruff
         args: [ --fix ]

docs/source/features/lora.md

@@ -26,4 +26,3 @@ $ olive export-adapters [-h] [--adapter_path ADAPTER_PATH] \
                       [--int4_block_size {16,32,64,128,256}] \
                       [--int4_quantization_mode {symmetric,asymmetric}]
 ```
-

olive/passes/onnx/conversion.py

@@ -10,6 +10,7 @@ from pathlib import Path
 from typing import Any, Dict, Optional, Tuple, Union
 
 import onnx
+import onnx.external_data_helper
 import torch
 from packaging import version
 
@@ -517,13 +518,26 @@ class OnnxOpVersionConversion(Pass):
     def _run_for_config(
         self, model: ONNXModelHandler, config: Dict[str, Any], output_model_path: str
     ) -> ONNXModelHandler:
-        # get current models's opset version
-        model_proto = model.load_model()
+        output_model_path = resolve_onnx_path(output_model_path)
+        # since external data is saved in a separate file, we need to load the model to get the opset version
+        model_proto = onnx.load(model.model_path, load_external_data=False)
+
         model_opset_version = model_proto.opset_import[0].version
         if model_opset_version == config["target_opset"]:
             logger.info("Model is already in target opset version %s.", config["target_opset"])
             return model
 
-        output_model_path = resolve_onnx_path(output_model_path)
-        model_proto = onnx.version_converter.convert_version(model_proto, config["target_opset"])
-        return model_proto_to_olive_model(model_proto, output_model_path, config)
+        converted_model_proto = onnx.version_converter.convert_version(model_proto, config["target_opset"])
+        # copy the external data of original model to the new model
+        dst_init_map = {init.name: init for init in converted_model_proto.graph.initializer}
+        for src_init in model_proto.graph.initializer:
+            if (
+                src_init.name in dst_init_map
+                and src_init.HasField("data_location")
+                and src_init.data_location == onnx.TensorProto.EXTERNAL
+            ):
+                dst_init_map[src_init.name].CopyFrom(src_init)
+        onnx.external_data_helper.load_external_data_for_model(
+            converted_model_proto, str(Path(model.model_path).resolve().parent)
+        )
+        return model_proto_to_olive_model(converted_model_proto, output_model_path, config)

test/unit_test/passes/onnx/test_conversion.py

@@ -64,7 +64,11 @@ def test_onnx_conversion_pass_quant_model(add_quantized_modules, tmp_path):
 def test_onnx_op_version_conversion_pass(target_opset, tmp_path):
     input_model = get_onnx_model()
     # setup
-    p = create_pass_from_dict(OnnxOpVersionConversion, {"target_opset": target_opset}, disable_search=True)
+    p = create_pass_from_dict(
+        OnnxOpVersionConversion,
+        {"target_opset": target_opset},
+        disable_search=True,
+    )
     output_folder = str(tmp_path / "onnx")
 
     onnx_model = p.run(input_model, output_folder)