Vendored torch_nn and torch_utils

Vendored torch_nn and torch_utils#

CoreRec ships internal copies of PyTorch utility modules under:

  • corerec/torch_nn/

  • corerec/torch_utils/

Why they exist#

These trees were vendored to:

  1. Stabilize legacy engine code that imported internal PyTorch paths before PyTorch 2.x API changes

  2. Support sandbox neural models (corerec/sandbox/collaborative_full/nn_base/) without pinning an old PyTorch fork

  3. Isolate experimental code from the production corerec.engines tier

Production guidance#

  • Production models (corerec/engines/) should prefer public PyTorch APIs (torch.nn, torch.optim) directly

  • Do not import corerec.torch_nn in new production code

  • Sandbox / research code may continue using vendored modules until sandbox models graduate or are removed

Maintenance policy#

Area

Policy

corerec/engines/

No new dependencies on vendored torch trees

corerec/sandbox/

Allowed; migrate on graduation

Type checking

mypy ignores vendored paths (see pyproject.toml)

Trimming

Long-term goal: shrink vendored surface or gate behind optional extra

For framework users#

If you only use production models, you can ignore torch_nn/ entirely — it is not part of the public API.

# ✅ Production import
from corerec.engines import DCN

# ⚠️ Internal / sandbox only
from corerec.torch_nn.modules.linear import Linear  # do not use in app code
Contents