Skip to content
My Code, My RulesMy Code, My Rules

PY-PERF0001 · tensor_interoperability_round_trip_count

Find avoidable host, NumPy, or explicit DLPack tensor round trips.

This is a deterministic rule for python. Read its implementation.

Definition

Section titled “Definition”

Resolve explicit Torch, CuPy, and RAPIDS import aliases. Report a recognized destination call only when its argument syntax proves an intermediate cpu, numpy, asnumpy, to_pandas, or explicit to_dlpack conversion from another supported tensor ecosystem. These libraries expose direct CUDA array or DLPack interoperability, so application code should pass the device object directly when the installed versions support that contract.

Evidence

Section titled “Evidence”

Each finding identifies the complete destination call and classifies the unnecessary bridge. The value is the number of destination calls fed through an avoidable host bridge.

Exceptions

Section titled “Exceptions”

Unqualified constructors, unknown aliases, serialization, deliberate host ownership, device changes, and a plain from_dlpack(value) call are excluded. A boundary may keep an explicit bridge when version constraints or lifetime semantics require it.

Examples

Section titled “Examples”

Bad

Section titled “Bad”
array = cp.asarray(tensor.cpu().numpy())
tensor = torch.from_numpy(cp.asnumpy(array))
array = cp.from_dlpack(torch.utils.dlpack.to_dlpack(tensor))

Good

Section titled “Good”
array = cp.asarray(tensor)
tensor = torch.as_tensor(array)

References

Section titled “References”
  • Cites “CuPy documentation”, interoperability with PyTorch and the CUDA Array Interface. Open reference
  • Cites “PyTorch documentation”, torch.as_tensor interoperability. Open reference
  • Cites “cuDF documentation”, CuPy interoperability guide. Open reference