 Frank
|
5df6ef3166
|
autoresearch: harness setup
Benchmark entrypoint: bash autoresearch.sh
Goal: Maximize frank-large decode throughput on the Strix Halo iGPU (gfx1151) via quantization >=q6/fp6, engine flags, spec-decode shape, or runtime swap — context limits unchanged, KV stays bf16.
|
2026-10-01 20:41:51 -07:00 |
|