llama.cpp

cuda: guard the iq4_nl dequantize row kernel against short rows

#29683

Merged