Reliability PyTorch active any

Device placement is explicit and consistent

pytorch-rel-003

Intent

Avoid runtime failures and hidden performance problems from mixed CPU and GPU placement.

Applicability

Applies to PyTorch code that uses accelerators or moves tensors between devices.

What to inspect

Device selection, model placement, batch placement, and forward-pass boundaries.

Pass criteria

The model and all relevant tensors are moved to the same explicit device.

Fail criteria

The diff mixes CPU and GPU tensors implicitly or leaves device handling ambiguous across phases.

Do not flag

Clearly CPU-only code.

Confidence guidance

HIGH when inconsistent placement is directly visible. MEDIUM when helper code may own movement. LOW when accelerator usage is unclear.

Remediation

Select the device explicitly and move both model and tensors consistently.

Pass example

model = model.to(device)
batch = batch.to(device)

Fail example

model = model.cuda()
pred = model(cpu_batch)