docs: add external data file handling section to PyTorch export tuto…

docs/tutorials/export-pytorch-model.md

@@ -106,3 +106,54 @@ torch.onnx.export(CustomInverse(), (x,), f, custom_opsets={"com.microsoft": 5})
 ```
 
 Note that you can export a custom op to any version >= the opset version used at registration.
+
+## Handling External Data Files
+
+When exporting PyTorch models using `torch.onnx.export` with opset 17+,
+PyTorch may automatically split the export into two files:
+
+- `model.onnx` — the model graph structure
+- `model.onnx.data` — the weight tensors stored as external data
+
+Both files must be present in the same directory for ONNX Runtime to load
+the model correctly. `InferenceSession` will fail if `model.onnx.data` is missing.
+
+```python
+# Fails if model.onnx.data is not in the same directory
+session = onnxruntime.InferenceSession("model.onnx")
+```
+
+### Option A — Models under 2 GB: merge into a single file
+
+For models whose weights fit within protobuf's 2 GB limit, the split files can be
+merged into a single self-contained `.onnx` file. This simplifies distribution
+to Docker images, cloud buckets, and API endpoints:
+
+```python
+import onnx
+
+model = onnx.load("model.onnx", load_external_data=True)
+onnx.save(model, "model_single.onnx", save_as_external_data=False)
+```
+
+> **Note:** Requires the `onnx` package (`pip install onnx`). Both `model.onnx`
+> and `model.onnx.data` must be in your working directory before running this.
+
+### Option B — Models 2 GB and above: deploy as a file pair
+
+When model weights exceed protobuf's 2 GB limit, merging is not possible.
+The `.onnx.data` file is the ONNX spec's mechanism for handling this case.
+Both files must be co-located at inference time:
+
+```python
+# Both model.onnx and model.onnx.data must be in the same directory
+session = onnxruntime.InferenceSession("/path/to/model/model.onnx")
+# ONNX Runtime resolves model.onnx.data relative to the .onnx file location
+```
+
+For cloud storage (e.g. GCS, S3), upload both files to the same prefix and
+download both to the same local directory before loading:
+
+```bash
+gsutil cp model.onnx model.onnx.data gs://your-bucket/models/
+```