ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-03 10:30:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

merge_qkv: qwen3 (dense)

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-10-29 15:59:38 +02:00
parent ed31b6741f
commit a2f3b08fbd
2 changed files with 11 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
src/llama-build-context.cpp
View File

@@ -3379,30 +3379,21 @@ ggml_cgraph * llm_build_context::build_qwen3() {
// self-attention
{
auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur, model.layers[il].wq, nullptr,
auto [Qcur, Kcur, Vcur] = llm_build_mul_mat_qkv(gf, cur,
model.layers[il].wqkv, nullptr,
model.layers[il].wq, nullptr,
model.layers[il].wk, nullptr,
model.layers[il].wv, nullptr, 0, il);
model.layers[il].wv, nullptr,
model.layers[il].attn_q_norm, model.layers[il].attn_k_norm, 0, il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, cb, il);
cb(Qcur, "Qcur_normed", il);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, cb, il);
cb(Kcur, "Kcur_normed", il);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, lctx, kv_self, gf,

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

@@ -1015,9 +1015,8 @@ bool create_tensors_helper::create_qwen3_tensors(const LLM_TN & tn) {
layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.wq = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head});
layer.wk = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
layer.wv = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
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});
layer.attn_k_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k});