torch.onnx.export produces an ONNX graph with out-of-bounds Gather for torch.as_strided after slicing

@justinchuby is already working on this.

Since Jun 3, 2026.

### Describe the issue

`torch.onnx.export` appears to produce a semantically invalid ONNX graph for a valid PyTorch eager program involving `torch.as_strided` on a sliced tensor.

In PyTorch eager mode, the model runs successfully and returns a tensor with shape `(6, 4)`. However, after ONNX export, ONNX Runtime fails during execution with an out-of-bounds Gather error:

indices element out of data bounds, idx=6 must be within the inclusive range [-6,5]

This looks like an exporter semantic mismatch. The eager operation is valid because `torch.as_strided` constructs a view based on the underlying storage. However, the exported ONNX graph appears to gather from the sliced tensor of length 6, causing generated indices such as 6 to become out of bounds.

### To reproduce

#### Minimal reproduction
```python
#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import json
import os
import platform
import traceback

import onnx
import onnxruntime as ort
import torch
import torch.nn as nn

SEED = 0

class MyModel(nn.Module):
def forward(self, x):
sliced = x[0:6]
return torch.as_strided(sliced, size=(6, 4), stride=(1, 1))

def print_runtime_info() -> None:
info = {
"python": platform.python_version(),
"platform": platform.platform(),
"torch_version": torch.__version__,
"onnx_version": onnx.__version__,
"onnxruntime_version": ort.__version__,
"torch_cuda_available": torch.cuda.is_available(),
"torch_cuda_device_count": torch.cuda.device_count(),
"cuda_visible_devices": os.environ.get("CUDA_VISIBLE_DEVICES", ""),
"ort_available_providers": ort.get_available_providers(),
"seed": SEED,
}
print("[runtime]", json.dumps(info, indent=2, sort_keys=True))

def export_model(model: nn.Module, x: torch.Tensor, path: str) -> None:
os.makedirs(os.path.dirname(path), exist_ok=True)
model.eval()
with torch.no_grad():
torch.onnx.export(
model,
(x,),
path,
input_names=["input_0"],
output_names=["output"],
opset_version=18,
)

def main() -> int:
torch.manual_seed(SEED)
print_runtime_info()

model = MyModel().eval()
x = torch.rand(10, dtype=torch.float32)

with torch.no_grad():
eager = model(x)

print(f"[input] shape={tuple(x.shape)} dtype={x.dtype}")
print(f"[eager] shape={tuple(eager.shape)} dtype={eager.dtype}")
print("[eager] value=")
print(eager)

onnx_path = os.path.abspath("as_strided_after_slice.onnx")
export_model(model, x, onnx_path)
print(f"[export] onnx_path={onnx_path}")

try:
sess = ort.InferenceSession(onnx_path, providers=["CPUExecutionProvider"])
feed = {sess.get_inputs()[0].name: x.numpy()}
out = sess.run(None, feed)
print("[not_reproduced] unexpected success", [getattr(v, "shape", None) for v in out])
return 1
except Exception as exc:
print("[reproduced] exception_type=", type(exc).__name__)
print("[reproduced] exception=", repr(exc))
print(traceback.format_exc())
return 0

if __name__ == "__main__":
raise SystemExit(main())
```

#### Actual behavior
```
PyTorch eager execution succeeds:

[eager] shape=(6, 4)

ONNX export also succeeds:

[torch.onnx] Obtain model graph for `MyModel()` with `torch.export.export(..., strict=False)`... ✅
[torch.onnx] Run decomposition... ✅
[torch.onnx] Translate the graph into ONNX... ✅

However, ONNX Runtime fails when executing the exported model:

[ONNXRuntimeError] : 2 : INVALID_ARGUMENT :
Non-zero status code returned while running Gather node. Name:'n12_2'
Status Message: indices element out of data bounds, idx=6 must be within the inclusive range [-6,5]
```

### Urgency

_No response_

### Platform

Linux

### OS Version

Ubuntu 22.04.4 LTS (x86_64)

### ONNX Runtime Installation

Released Package

### ONNX Runtime Version or Commit ID

1.23.2

### ONNX Runtime API

Python

### Architecture

X64

### Execution Provider

Default CPU

### Execution Provider Library Version

_No response_