The rknpu2 backend registers as GGML_BACKEND_DEVICE_TYPE_ACCEL, which can
leave the default n_gpu_layers >= 0 by the time common_params_parser_init
builds the -ngl / -ngld options. Upstream's GGML_ASSERT(n_gpu_layers < 0)
then aborts llama-server at startup on every invocation (build/bin was
unusable; --version alone dumped core).
Replace the assert -- as its own comment suggested -- by extending the
help-string default display to handle n_gpu_layers >= 0 (show the number;
-1 = auto, <= -2 = all), at both the main and draft -ngl sites. No behaviour
change for CPU/GPU builds where the default stays -1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EWpfhDgYNA21tETDP9ueBE
mfritsche
9d64138d24