Large-Scale TensorCircuit Contractions on GPUs: Disabling XLA GPU Autotuning

Short summary

Disabling XLA GPU autotuning (--xla_gpu_autotune_level=0) dramatically reduces persistent GPU memory usage during large-scale TensorCircuit contractions on JAX, with minimal runtime impact. Benchmarks show post-compile memory dropping from 8.7 GiB to 0.5 GiB in a 100-qubit workload. The article includes A/B test methodology and diagnostic techniques using the platform allocator to trace memory origins.

• •Disabling XLA GPU autotuning cuts persistent memory from 8.7 GiB to 0.5 GiB in 100-qubit contractions
• •Steady-state runtime is unaffected or slightly improved
• •Platform allocator diagnostics reveal where GPU allocations originate

Generated with AI, which can make mistakes.