get_recommended_model() now keys to the backend the installed whisper.cpp
build actually has (detect_pywhispercpp_gpu_backend) and to free memory
measured at recommendation time: free VRAM on a discrete GPU for a
Vulkan or CUDA build, free system RAM otherwise. The result is a rung of
the model ladder, so a Vulkan-capable Nvidia card reaches the ladder
instead of the total-memory tiers, and unquantized turbo, the large
family and medium are never a default.
Manual selections are untouched. The informational hardware probe no
longer decides the model. Dead VRAM-total parsing is removed.
free_memory_for_backend() holds the source choice; the floor-rung map is
shared by rung selection and step-down.
q8_0 threshold becomes 1644 MiB (1132+512) and q5_0 becomes 1358 MiB
(846+512) per the recorded baseline recommendation, replacing the
on-disk-size proxies. Tests pin the measured values and the 1644/1358
boundaries.
Free memory picks the load-time rung (large-v3-turbo-q8_0 default,
q5_0 under 1086 MiB free, small/small.en floor) for unpinned default
configs; thresholds derive from catalog sizes plus the protocol's
512 MiB headroom. A manual pin is never overridden, and a failed load
steps one rung down and retries exactly once before surfacing.
CUDA ranks before Vulkan only with a complete architecture-matching
toolkit (nvidia-smi GPU, nvcc, compute capability >= 7.0); otherwise
Vulkan, then CPU. The driver-only CUDA recommendation now demotes to
CPU logic instead of naming an unreachable VRAM ladder.
Refs #20
xavierk
soubarnak