local-llm
- Nine accusations, one lie: re-reading round two's judges
137 reviewed local-model agent runs, 14 "false" verdicts from two judges. On the raw sessions 9 fell, 4 were about docs, and 1 was a lie.
- The third run mattered where it cost the most
739 three-run groups from my agent bench: 19% split on basic tasks, 56% on Zig. So basic drops to two runs, and Zig runs only after the hard tier earns it.
- The race detector saw my requests in order
A mutant removed a lock in my Go proxy and survived -race. The test sent its requests one after another, and that order hid the unlocked read.
- The config said 32k. The engine still ran 64k.
Adding SIGHUP reload to a Go LLM proxy: a whitelist of three settings, a model list that runs ahead of the engines, and a refresh that finished last.
- My isolation guard checked itself
A guard meant to prove my Go tests couldn't reach the live proxy compared a path with the env var it guarded, and passed on any shell that had set it.
- The mutant was killed after it switched a real GPU
A hand-written mutant made my Go CLI skip -proxy. Its test failed on cue, after a real model load that then kept a benchmark off a GPU for almost half an hour.
- The lock map deleted a mutex someone was waiting on
A Go transfer by Flash-Next Q2_0 scored 100 and passed 6 of 6 failure-path tests. One cleanup line let a retried key move the money twice.
- llama.cpp's MoE cache wants gigabytes and long prompts
On a 24 GB RTX 3090, llama.cpp's MoE expert cache read long prompts up to 85% faster, did nothing for a Q8 model, and left a 2-4x gap to Strata.
- The 3090 unlock is real. Where it lands is the catch
gpu_unlock lifted my RTX 3090's BF16 GEMM from 74.9 to 128 TFLOPS. My daily 27B in llama.cpp got 0.7% faster at short context. Here's where the rest went.
- Load gpu_unlock before nvidia on Bazzite for an RTX 3090
Build the gpu_unlock module for the booted Bazzite kernel, load it from the initramfs before nvidia, opt in, verify the unlock, recover, revert.