Benchmarking

Two GPUs, Four Crashes, and 384 Benchmark Runs

One question about a 30B model turned into rebuilding my home server's entire local-LLM stack. The hard part was learning to trust my own numbers.