PyTorch PR removes unnecessary tensor copies, boosting autograd speed

PyTorch has merged an optimization pull request (PR #189582) that removes an unnecessary tensor copy operation from the `autograd.Function.apply` path. Previously, every tensor input was unconditionally copied into a vector called `record_function_inputs` during argument unpacking. That vector is only used internally by `RecordFunction` callbacks for profiling and tracing, meaning normal execution — when the profiler is inactive — was paying the cost of converting tensors to IValues (intermediate representations) even though no callback ever consumes them.

By skipping this copy when profiling is off, the PR delivers a measurable performance gain: roughly 0.3–0.4 microseconds per call on a 13-input, 2-output autograd.Function benchmark. While the absolute saving is small, it accumulates across many forward/backward calls in training loops, reducing overall overhead. The change is backward-compatible and has zero impact on profiling functionality — profiled runs still capture the same data. The PR was reviewed and approved by core PyTorch contributors, including Skylion007 and soulitzer, and is part of a broader effort to streamline the autograd engine.