mirror of
https://github.com/ollama/ollama.git
synced 2026-04-24 09:46:01 +02:00
llama: update to commit 2016f07b (#10352)
This commit is contained in:
4
llama/llama.cpp/src/llama-hparams.h
vendored
4
llama/llama.cpp/src/llama-hparams.h
vendored
@@ -46,6 +46,10 @@ struct llama_hparams {
|
||||
uint32_t n_rel_attn_bkts = 0;
|
||||
uint32_t n_vocab = 0;
|
||||
|
||||
// note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
|
||||
uint32_t n_embd_head_k_mla = 0;
|
||||
uint32_t n_embd_head_v_mla = 0;
|
||||
|
||||
// for WavTokenizer
|
||||
struct llama_hparams_posnet posnet;
|
||||
struct llama_hparams_convnext convnext;
|
||||
|
||||
Reference in New Issue
Block a user