Galahad crash fixed in 1.32.0 — confirmed 3.6x speedup, 19.5% less energy

[i-redspector]

Ran the full A/B benchmark from the GPU AI Benchmarking Report Contest challenge against Galahad after the 1.32.0 release.

Before the fix: crashed every time, on two machines and two vLLM versions, allocating the KV cache block tables.

After the fix: clean start, 36/36 correct answers across 3 runs, and once the disk budget was sized to the actual working set, a 3.6x speedup over plain vLLM (0.91s vs 3.29s median revisit time), with 19.5% less total energy consumed for the same benchmark (measured with nvidia-smi).

Full methodology, raw logs, and results: https://github.com/i-redspector/galahad-benchmark-results