ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-05 11:30:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

merge_qkv: llama-4

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-10-29 15:44:24 +02:00
parent 5c138c4d08
commit ed31b6741f
2 changed files with 3 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llama-build-context.cpp
View File

@@ -1373,7 +1373,7 @@ ggml_cgraph * llm_build_context::build_llama() {
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
} else if (inp_attn_scale) {
Qcur = ggml_mul(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_attn_scale);
Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
}
cb(Qcur, "Qcur", il);

5
src/llama-load-tensors.cpp
View File

@@ -416,9 +416,8 @@ bool create_tensors_helper::create_llama4_tensors(const LLM_TN & tn) {
layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
layer.wv = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
use_mmap_buffer &= !merge_qkv(tn, i, 0);
layer.wo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
layer.ffn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);