Disabling XLA GPU Autotuning Cuts Memory Use in Large TensorCircuit Workloads

Researchers working with TensorCircuit-NG and the JAX GPU backend found that disabling XLA GPU autotuning via the flag --xla_gpu_autotune_level=0 significantly reduces persistent GPU memory consumption during compilation. In benchmark tests across multiple quantum circuit workloads, post-compile memory dropped from as high as 17 GiB to around 0.5 GiB when autotuning was turned off. The primary gain is memory efficiency rather than speed, though a modest improvement in steady-state runtime was also observed in some cases. XLA autotuning is less useful for large tensor-network contractions because the contraction path is already fixed by tools like OMECO or cotengra, leaving little room for kernel-level optimization. The recommended approach is to run an A/B test with and without the flag, setting both environment variables before the Python process starts and before JAX is imported.