[BUG] `TensorDict.memmap` fails for path-like keys

[egorchakov]

Describe the bug

.memmap fails on a TensorDict with path-like keys (e.g. containing / on *nix)

To Reproduce

# file: tensordict_memmap_issue.py

from tempfile import TemporaryDirectory
import torch
from tensordict import TensorDict


def main() -> None:
    td = TensorDict({"a": torch.zeros(10)})
    with TemporaryDirectory() as tmpdir:
        td.memmap(tmpdir)  # succeeds

    td = TensorDict({"a/b/c": torch.zeros(10)})
    with TemporaryDirectory() as tmpdir:
        td.memmap(tmpdir)  # fails


if __name__ == "__main__":
    main()

uv run --isolated --python 3.13 --with torch==2.8.0,tensordict==0.10.0 tensordict_memmap_issue.py

Traceback (most recent call last):
  File "/Users/evgenii/dev/tensordict_memmap_issue.py", line 17, in <module>
    main()
    ~~~~^^
  File "/Users/evgenii/dev/tensordict_memmap_issue.py", line 13, in main
    td.memmap(tmpdir)  # fails
    ~~~~~~~~~^^^^^^^^
  File "/Users/evgenii/.cache/uv/archive-v0/IFrn6zo5IgF1ZlKrCxPZX/lib/python3.13/site-packages/tensordict/base.py", line 6282, in memmap
    return self._memmap_(
           ~~~~~~~~~~~~~^
        prefix=prefix,
        ^^^^^^^^^^^^^^
    ...<6 lines>...
        existsok=existsok,
        ^^^^^^^^^^^^^^^^^^
    ).lock_()
    ^
  File "/Users/evgenii/.cache/uv/archive-v0/IFrn6zo5IgF1ZlKrCxPZX/lib/python3.13/site-packages/tensordict/_td.py", line 2825, in _memmap_
    _populate_memmap(
    ~~~~~~~~~~~~~~~~^
        dest=dest,
        ^^^^^^^^^^
    ...<5 lines>...
        existsok=existsok,
        ^^^^^^^^^^^^^^^^^^
    )
    ^
  File "/Users/evgenii/.cache/uv/archive-v0/IFrn6zo5IgF1ZlKrCxPZX/lib/python3.13/site-packages/tensordict/_td.py", line 4784, in _populate_memmap
    memmap_tensor = MemoryMappedTensor.from_tensor(
        value.data if value.requires_grad else value,
    ...<4 lines>...
        existsok=existsok,
    )
  File "/Users/evgenii/.cache/uv/archive-v0/IFrn6zo5IgF1ZlKrCxPZX/lib/python3.13/site-packages/tensordict/memmap.py", line 250, in from_tensor
    result = torch.from_file(
        str(filename),
    ...<4 lines>...
        device=torch.device("cpu"),
    )
RuntimeError: unable to open file </var/folders/6c/9tvxzb9x2q74bbfkh1jxtby40000gn/T/tmpsjh7q2xq/a/b/c.memmap> in read-write mode: No such file or directory (2)

Expected behavior

Saving/loading works.

System info

import tensordict, numpy, sys, torch
print(tensordict.__version__, numpy.__version__, sys.version, sys.platform, torch.__version__)

0.10.0 2.3.3 3.13.7 (main, Aug 28 2025, 17:02:04) [Clang 20.1.4 ] darwin 2.8.0

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[vmoens]

Fixed in #1443
LMK what you think!

[egorchakov]

tested with tensordict-nightly==2025.9.24, works great 🙏