nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 37 Packages Projects Releases Wiki Activity

llama: dynamic head_dim and n_rot for SWA (#20301)

Browse Source 
* llama: dynamic head_dim and n_rot for SWA

* also add gguf_writer wrappers

* fix build

* build_rope_shift arg reorder
This commit is contained in:
Xuan-Son Nguyen
2026-03-09 22:22:39 +01:00
committed by GitHub
parent 23fbfcb1ad
commit 59db9a357d
112 changed files with 419 additions and 346 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-hparams.h
+16 -3
View File
@@ -44,13 +44,20 @@ struct llama_hparams {
uint32_t n_embd;
uint32_t n_layer;
int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
uint32_t n_rot;
uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
uint32_t n_expert = 0;
uint32_t n_expert_used = 0;
uint32_t n_rel_attn_bkts = 0;
// different head size for full_attention and SWA layers
uint32_t n_embd_head_k_full; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
uint32_t n_embd_head_v_full; // dimension of values (d_v) aka n_embd_head
uint32_t n_embd_head_k_swa;
uint32_t n_embd_head_v_swa;
// different RoPE dimensions for full_attention and SWA layers
uint32_t n_rot_full;
uint32_t n_rot_swa;
// note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
uint32_t n_embd_head_k_mla_impl = 0;
uint32_t n_embd_head_v_mla_impl = 0;
@@ -247,12 +254,18 @@ struct llama_hparams {
uint32_t n_gqa(uint32_t il = 0) const;
uint32_t n_rot(uint32_t il = 0) const;
// dimension of main + auxiliary input embeddings
uint32_t n_embd_inp() const;
// dimension of output embeddings
uint32_t n_embd_out() const;
// dimension of key/value embeddings for each head (per layer)
uint32_t n_embd_head_k(uint32_t il = 0) const;
uint32_t n_embd_head_v(uint32_t il = 0) const;
// dimension of key embeddings across all k-v heads
uint32_t n_embd_k_gqa(uint32_t il = 0) const;