AirLLM user reports contradicting the compression speed-up claim

Issue titles retrieved 2026-10-03 via GitHub issue search, used as the counterweight to the README's throughput framing: '3x speedup compression claim is false/outdated'; 'Windows + NVIDIA: 28.6 s/token on a 3B model'; 'taking about 40 minutes to generate one sentence'; 'run this with windows amd 9070 XT ROCm 7.12 VERRY SLOW'. Forum-grade evidence (kind=forum), so confidence is deliberately low and the engine record is marked draft. Verified 2026-10-10: Read issue search results on lyogavin/airllm (https://github.com/lyogavin/airllm/issues?q=is%3Aissue+slow); confirmed existence of issues #330 ('3x speedup compression claim is false/outdated'), #364 ('Windows + NVIDIA: 28.6 s/token on a 3B model...'), #186 ('taking about 40 minutes to generate one sentence...'), and #301 ('run this with windows amd 9070 XT ROCm 7.12 VERRY SLOW').