Files
voculinux/tests
xavierk 64020a132d feat(model): hard-code per-backend ladder thresholds from ticket #19
The whisper.cpp ladder keyed both GPU backends to one provisional pair of
thresholds. Issue #19 (comment 12913) measured each rung's transcribe-time
VRAM delta on the installed build. The thresholds are those deltas plus the
512 MiB desktop headroom:

- vulkan: q8_0 1644 MiB, q5_0 1359 MiB (was 1644 and 1358)
- cuda:   q8_0 1690 MiB, q5_0 1404 MiB

select_ladder_rung and resolve_startup_rung take the installed build's
backend. CPU and unknown backends keep the Vulkan thresholds; the RAM rule
proposed in #19 is not implemented.

The first-run default in config_manager now reads the memory of the
installed build (VRAM for a GPU build, RAM otherwise). It used to read VRAM
whenever an NVIDIA card was present, even on a CPU build.
2026-11-07 01:46:38 +05:30
..