2026-04-09 04:54:01,412 [INFO] Loading faiss with AVX512 support.
2026-04-09 04:54:01,533 [INFO] Successfully loaded faiss with AVX512 support.
2026-04-09 04:54:03,332 [INFO] Benchmarking full recompute (50 trials)...
2026-04-09 04:54:10,527 [INFO] Benchmarking streaming inference (50 trials)...

============================================================
STREAMING INFERENCE BENCHMARK
============================================================
Context: 256 units, 10 features
Device: cuda:0
------------------------------------------------------------
Full recompute:    128.72 ± 65.50 ms
KV-cached:         133.05 ± 78.01 ms
Speedup:              1.0×
============================================================
2026-04-09 04:54:18,907 [INFO] Results saved to outputs/benchmarks/streaming_results.json