Angel Li a463300ccc [CPP FakeTensor] Add fake tensor accessor helpers (#193750) ...

C++ fake tensors are plain torch.Tensor objects, so code that reaches into
Python FakeTensor attributes (fake_device, item_memo, constant, real_tensor,
dispatch_keys) or checks isinstance(x, FakeTensor) silently skips them. This
adds maybe_get_*/maybe_set_* helpers in fake_tensor.py that handle both kinds
(plus the _clear_fake_real_tensor binding they need), and switches
fake_tensor.py itself over to them and to is_fake_tensor. The following
commits in the stack switch the rest of torch over, one file or folder at a
time.

This is intended to be a refactor for today's (Python FakeTensor) behavior:
every new C++ branch is gated on torch._C._is_fake_tensor, which is false for
every tensor unless a C++ fake exists. Equivalences worth checking when
reviewing:

- merge_devices reads the fake device directly instead of t.device. All
  _find_common_device callers run outside in_kernel_invocation, where
  FakeTensor.device already returns fake_device.
- allow_non_fake_inputs_temporarily is added here as a replacement for
  mock.patch.object(mode, "allow_non_fake_inputs", True), which cannot
  restore C++-backed properties. allow_non_fake_inputs is a plain instance
  attribute set in __init__ on the Python mode.
- FakeTensorMode.clear_non_cpu_constants forwards to the converter.

Test Plan:

```
python -m py_compile <touched files>
lintrunner <touched files>
```

Authored with an AI assistant.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/193750
Approved by: https://github.com/ezyang
ghstack dependencies: #200174

2026-10-09 15:05:07 +00:00

..

_internal

[CPP FakeTensor] Add fake tensor accessor helpers (#193750)

2026-10-09 15:05:07 +00:00

__init__.py

Remove NJT (nested jagged tensor) from weights_only safe globals (#194380)

2026-08-24 01:38:42 +00:00