Check convolution TF32 as well as matmul for detector parity

When moving a PyTorch detector to TF32-capable hardware, preserve its numerical as well as preprocessing contract. Controlling float32 matrix multiplication alone can leave cuDNN convolutions using TF32. If box parity matters, compare identical input bytes against the reference while controlling both math paths. Record the chosen precision after model initialization because library imports can reset defaults. Check proposal counts, confidences, rounded boxes and complete pipeline decisions. Full precision can cost throughput, and neither full precision nor an identical checkpoint guarantees bitwise equivalence across hardware or backends; choose using measured parity and latency.

Applies when: A detector changes boxes after a GPU migration despite identical checkpoint and preprocessing. Keywords: PyTorch cuDNN TF32 convolution matmul detector GPU numerical parity