============================================================ Lyapunov Speedup Benchmark Job ID: 14632859 | Node: gpub073 Start: Tue Dec 30 06:43:26 CST 2025 ============================================================ NVIDIA A40, 46068 MiB ============================================================ ================================================================================ LYAPUNOV COMPUTATION SPEEDUP BENCHMARK ================================================================================ Batch size: 64 Timesteps: 4 Hidden dims: [64, 128, 256] Device: cuda ================================================================================ [1/6] Benchmarking Baseline... [2/6] Benchmarking Approach A (batched)... [3/6] Benchmarking Approach B (global renorm)... [4/6] Benchmarking Approach A+B (combined)... [5/6] Benchmarking Approach C (compiled baseline)... [6/6] Benchmarking A+B+C (all optimizations)... torch.compile failed: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [2, 64, 256]], which is output 0 of torch::autograd::CopySlices, is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True). ================================================================================ RESULTS ================================================================================ Baseline | Fwd: 10.28ms | Bwd: 7.75ms | Total: 18.03ms | λ: +1.0693 | Mem: 22.6MB A: Batched trajectories | Fwd: 7.20ms | Bwd: 7.43ms | Total: 14.63ms | λ: +1.0849 | Mem: 23.7MB B: Global renorm | Fwd: 9.49ms | Bwd: 6.99ms | Total: 16.48ms | λ: +0.6573 | Mem: 25.0MB A+B: Combined | Fwd: 6.55ms | Bwd: 6.76ms | Total: 13.30ms | λ: +0.6575 | Mem: 26.1MB C: Compiled baseline | Fwd: 8150.22ms | Bwd: 7502.86ms | Total: 15653.07ms | λ: +1.0758 | Mem: 44.5MB A+B+C: All optimized | Fwd: 0.00ms | Bwd: 0.00ms | Total: 0.00ms | λ: +0.0000 | Mem: 0.0MB -------------------------------------------------------------------------------- SPEEDUP vs BASELINE: -------------------------------------------------------------------------------- A: Batched trajectories : 1.23x B: Global renorm : 1.09x A+B: Combined : 1.36x C: Compiled baseline : 0.00x -------------------------------------------------------------------------------- LYAPUNOV VALUE CONSISTENCY CHECK: -------------------------------------------------------------------------------- A: Batched trajectories : λ=+1.0849 (diff=0.0156) ✓ B: Global renorm : λ=+0.6573 (diff=0.4119) ✗ A+B: Combined : λ=+0.6575 (diff=0.4117) ✗ C: Compiled baseline : λ=+1.0758 (diff=0.0065) ✓ ================================================================================ SCALING TESTS ================================================================================ Config | Baseline | A+B | Speedup --------------------------------------------------------------------------------