llama.cpp

Assume tied weights if lm_head/output weights is missing.

#5824

Merged