PR #16225 metal : extend mat-mat multiplication support

metal : support mul_mm with src1->type == GGML_TYPE_F16

ggerganov committed 91 days ago

metal : support mul_mm_id with src1->type == GGML_TYPE_F16

ggerganov committed 91 days ago

metal : mul_mm support ne00 % 32 != 0

ggerganov committed 91 days ago

metal : support mul_mm_id with ne00 % 32 != 0

ggerganov committed 91 days ago

cont : remove unnecessary unrolls

ggerganov committed 91 days ago

cont : simplify data loading

ggerganov committed 91 days ago

metal : optimize mul_mm when output bounds checks are not needed

ggerganov committed 91 days ago

llama.cpp metal : extend mat-mat multiplication support #16225 Merged