[5725362] AutoCast Fixes for models with external data (#731)

## What does this PR do?

**Type of change:** Bug fix <!-- Use one of the following: Bug fix, new
feature, new example, new tests, documentation. -->

**Overview:** Fix AutoCast ReferenceRunner to handle large models.
Models above 2GB cannot be serialized to string, which is what
polygraphy is doing under the hood. Use a temporary file instead to save
the modified onnx with all tensors marked as outputs.

## Before your PR is "*Ready for review*"
<!-- If you haven't finished some of the above items you can still open
`Draft` PR. -->

- **Make sure you read and follow [Contributor
guidelines](https://github.com/NVIDIA/Model-Optimizer/blob/main/CONTRIBUTING.md)**
and your commits are signed.
- **Is this change backward compatible?**: Yes <!--- If No, explain why.
-->
- **Did you write any new necessary tests?**: No
- **Did you add or update any necessary documentation?**: No
- **Did you update
[Changelog](https://github.com/NVIDIA/Model-Optimizer/blob/main/CHANGELOG.rst)?**:
No <!--- Only for new features, API changes, critical bug fixes or bw
breaking changes. -->

## Additional Information
<!-- E.g. related issue. -->

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **Improvements**
* Enhanced model processing to better support large ONNX models during
validation and runtime execution
* Added diagnostic logging of model sizes at key processing stages for
improved debugging and performance monitoring

<sub>✏️ Tip: You can customize this high-level summary in your review
settings.</sub>

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Signed-off-by: Gal Hubara Agam <96368689+galagam@users.noreply.github.com>
Signed-off-by: Daniel Korzekwa <dkorzekwa@nvidia.com>

Author: galagam

modelopt/onnx/autocast/precisionconverter.py

@@ -175,7 +175,7 @@ def convert(
             onnx.ModelProto: The converted mixed precision model.
         """
         try:
-            self.model = onnx_utils.check_model(self.model)
+            onnx_utils.check_model(self.model)
         except onnx.checker.ValidationError as e:
             logger.error(f"Internal error: onnx.checker failed on input model {e}")
             raise Exception(

modelopt/onnx/autocast/referencerunner.py

@@ -24,11 +24,13 @@
 import copy
 import io
 import sys
+import tempfile
 from collections import OrderedDict
 
 import numpy as np
 import onnx
 
+from modelopt.onnx import utils as onnx_utils
 from modelopt.onnx.autocast.logging_config import configure_logging, logger
 from modelopt.onnx.quantization.calib_utils import CalibrationDataProvider
 from modelopt.onnx.quantization.ort_utils import _prepare_ep_list
@@ -125,13 +127,42 @@ def _load_inputs(self, inputs):
 
         return data_loader
 
+    def _get_ort_runner(self, model):
+        from polygraphy.backend.onnx import BytesFromOnnx
+        from polygraphy.backend.onnxrt import OnnxrtRunner, SessionFromOnnx
+
+        # Check if model has external data by checking:
+        # 1. If any initializer has data_location set to EXTERNAL (even if data is loaded)
+        # 2. If model size would exceed 2GB (indicating need for external data)
+        needs_external_data = onnx_utils.check_model_uses_external_data(
+            self.model
+        ) or self.model.ByteSize() > 2 * (1024**3)
+        if needs_external_data:
+            logger.debug("Model has external data, using file-based approach")
+            # Get the actual ONNX ModelProto from ModifyOutputs wrapper
+            modified_model = model()
+
+            # Use a persistent temp file, because we need the file to be present in an broader context
+            tmp_file = tempfile.NamedTemporaryFile(suffix=".onnx", delete=False)
+            tmp_file.close()
+            tmp_file_path = tmp_file.name
+            onnx_utils.save_onnx(modified_model, tmp_file_path, save_as_external_data=True)
+            logger.debug(f"Model with all outputs saved to {tmp_file_path}")
+            build_onnxrt_session = SessionFromOnnx(tmp_file_path, providers=self.providers)
+
+        else:
+            # For models without external data, use the original BytesFromOnnx approach (no tmp files)
+            logger.debug("Model has no external data, using BytesFromOnnx approach")
+            serialize_onnx = BytesFromOnnx(model)
+            build_onnxrt_session = SessionFromOnnx(serialize_onnx, providers=self.providers)
+        runners = [OnnxrtRunner(build_onnxrt_session)]
+        return runners
+
     def run(self, inputs=None):
         """Run FP32 inference with provided or random inputs."""
         import onnxruntime as ort
         from polygraphy import constants
-        from polygraphy.backend.onnx import BytesFromOnnx
         from polygraphy.backend.onnx import ModifyOutputs as ModifyOnnxOutputs
-        from polygraphy.backend.onnxrt import OnnxrtRunner, SessionFromOnnx
         from polygraphy.comparator import Comparator
 
         logger.info("Running ONNX Runtime to obtain reference outputs (this may take a while)...")
@@ -140,9 +171,9 @@ def run(self, inputs=None):
 
         model_copy = copy.deepcopy(self.model)
         modify_outputs = ModifyOnnxOutputs(model_copy, outputs=constants.MARK_ALL)
-        serialize_onnx = BytesFromOnnx(modify_outputs)
-        build_onnxrt_session = SessionFromOnnx(serialize_onnx, providers=self.providers)
-        runners = [OnnxrtRunner(build_onnxrt_session)]
+
+        # Load the modified model and create an inference session
+        runners = self._get_ort_runner(modify_outputs)
 
         # Comparator is used despite the fact that we are using ONNXRuntime
         # because it provides the ability to generate random inputs using DataLoader

modelopt/onnx/utils.py

@@ -15,6 +15,7 @@
 
 """Utility functions related to onnx."""
 
+import copy
 import io
 import os
 import tempfile
@@ -552,7 +553,7 @@ def _get_unique_name(old_name):
     return onnx_model, is_modified
 
 
-def check_model(model: onnx.ModelProto) -> onnx.ModelProto:
+def check_model(model: onnx.ModelProto) -> None:
     """Checks if the given model is valid."""
     if model.ByteSize() > (2 * (1024**3)):  # 2GB limit
         with tempfile.TemporaryDirectory() as temp_dir:
@@ -561,10 +562,8 @@ def check_model(model: onnx.ModelProto) -> onnx.ModelProto:
             onnx_tmp_path = os.path.join(temp_dir, f"model_{unique_id}.onnx")
             save_onnx(model, onnx_tmp_path, save_as_external_data=True)
             onnx.checker.check_model(onnx_tmp_path)
-            return onnx.load(onnx_tmp_path)
     else:
         onnx.checker.check_model(model)
-        return model
 
 
 def find_lowest_common_ancestor(node1: Node, node2: Node) -> tuple[str | None, int, int]:
@@ -658,15 +657,16 @@ def save_onnx(model: onnx.ModelProto, onnx_path: str, save_as_external_data: boo
 
     # Set ir_version to 10, remove it once ORT supports ir_version 11
     model.ir_version = 10
-
     if save_as_external_data:
         external_data_path = os.path.basename(onnx_path) + "_data"
         if os.path.exists(external_data_path):
             logger.warning(f"Removing existing external data file: {external_data_path}")
             os.remove(external_data_path)
 
+        # Copy so the onnx.ModelProto object will not be modified
+        model_copy = copy.deepcopy(model)
         onnx.save_model(
-            model,
+            model_copy,
             onnx_path,
             save_as_external_data=True,
             all_tensors_to_one_file=True,
@@ -696,6 +696,21 @@ def get_opset_version(model: onnx.ModelProto) -> int:
     return ai_onnx_domain[0].version
 
 
+def check_model_uses_external_data(model: onnx.ModelProto) -> bool:
+    """Checks if the model uses external data.
+
+    Args:
+        model: Loaded in-memory onnx ModelProto.
+
+    Returns:
+        True if any initializer tensor has data_location set to EXTERNAL.
+    """
+    return any(
+        init.HasField("data_location") and init.data_location == onnx.TensorProto.EXTERNAL
+        for init in model.graph.initializer
+    )
+
+
 def bfloat16_to_float32(bf16_array):
     """Converts a bfloat16 array (as raw data) to a float32 array."""
     uint32_array = bf16_array.astype(np.uint32) << 16

modelopt/torch/_deploy/utils/onnx_utils.py

@@ -45,14 +45,3 @@ def _get_onnx_external_data_tensors(model: onnx.ModelProto) -> list[str]:
         if tensor.HasField("data_location") and tensor.data_location == onnx.TensorProto.EXTERNAL
     ]
     return model_tensors_ext
-
-
-def check_model_uses_external_data(model: onnx.ModelProto) -> bool:
-    """
-    Checks if the model uses external data.
-    """
-    model_tensors = _get_initializer_tensors(model)
-    return any(
-        tensor.HasField("data_location") and tensor.data_location == onnx.TensorProto.EXTERNAL
-        for tensor in model_tensors
-    )

modelopt/torch/_deploy/utils/torch_onnx.py

@@ -42,6 +42,7 @@
 )
 from modelopt.onnx.quantization.qdq_utils import qdq_to_dq, replace_zero_scale_with_smallest_nonzero
 from modelopt.onnx.utils import (
+    check_model_uses_external_data,
     get_input_names,
     get_input_shapes,
     get_node_names,
@@ -55,7 +56,6 @@
 from modelopt.torch.utils._pytree import TreeSpec
 
 from ..utils.onnx_optimizer import Optimizer
-from .onnx_utils import check_model_uses_external_data
 
 ModelMetadata = dict[str, Any]
 ModelType = Any