llama.cpp

Test-based VRAM scratch size + context adjustment

#2056

Merged