mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

graph : add optional scale parameter to build_lora_mm [no ci] (#20427)

Browse Source
This commit is contained in:
Richard Davison
2026-03-12 00:22:49 +01:00
committed by GitHub
parent 4a748b8f15
commit 1eea6a2968
6 changed files with 23 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/qwen3.cpp
+3 -12
View File
@@ -30,22 +30,13 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
if (model.layers[il].wq_s) {
Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_s);
}
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
if (model.layers[il].wk_s) {
Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_s);
}
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
if (model.layers[il].wv_s) {
Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_s);
}
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);