llamafile_sgemm API says it may compute C = A^T x B . When is A matrix transposed in llama.cpp ?

[shalinisalom99]

I would like to know here if A is weights matrix and B is the activations matrix ? Also, is the transposition for A Matrix is done offline during PyTorch -> gguf conversion ? If so, is this the piece of code which does it ?

llama.cpp/convert_hf_to_gguf.py

Line 3095 in 7d51644

if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):

If not, can you please point at what phase in runtime does this transposition happen ?