Tuning local LLMs on a Radeon AI PRO R9700
Speculative decoding gave the biggest single win: n-gram drafting made file rewrites 3.5x faster on Devstral and 2.5x faster on Thinklite, with reasoning speed unchanged.
Speculative decoding gave the biggest single win: n-gram drafting made file rewrites 3.5x faster on Devstral and 2.5x faster on Thinklite, with reasoning speed unchanged.
Arete finished all five fixes; Flint wrote working code for four but stopped within the request limit on only two of those runs.
I saved and reloaded my campaign between three sessions on the macOS port and reached turn 615 without a crash.
Over 29 hours, the AC used 3.84 kWh to keep the living room at 23.4–24.1 °C while the uncooled bedroom reached 29.8 °C.
Two workers hit a merge conflict editing the same build file; in the next batch of eight, I assigned separate insertion points and saw no conflicts.