ggml: MUL_MAT_ID has a backward pass through frozen experts, so a LoRA trains on a MoE base (GGML_OP_MUL_MAT_ID_BACK; continuum card 14b64aa9) - #82

ggml: MUL_MAT_ID has a backward pass through frozen experts, so a LoRA trains on a MoE base (GGML_OP_MUL_MAT_ID_BACK; continuum card 14b64aa9)#82

joelteply merged 3 commits into

889580fc/seq-state-across-stream-countsCambrianTech/llama.cpp:889580fc/seq-state-across-stream-countsfrom

cormac/mul-mat-id-backwardCambrianTech/llama.cpp:cormac/mul-mat-id-backwardCopy head branch name to clipboard