mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

graph : reduce topology branching (#18548)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-01-02 19:01:56 +02:00
committed by GitHub
parent d84a6a98be
commit af1e8e1a6c
4 changed files with 14 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/gemma3n-iswa.cpp
+4 -7
View File
@@ -1,7 +1,5 @@
#include "models.h"
llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params),
model(model),
@@ -15,10 +13,9 @@ llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const
inpL = build_inp_embd(model.tok_embd);
// important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
if (ubatch.token) {
inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
cb(inpL, "inp_scaled", -1);
}
inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
cb(inpL, "inp_scaled", -1);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
@@ -248,7 +245,7 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
// equivalent to get_per_layer_inputs() in python code
// output shape: [n_embd_altup, n_layer, n_tokens]
ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
auto inp = std::make_unique<llm_graph_input_embd>();
auto inp = std::make_unique<llm_graph_input_embd>();
ggml_tensor * inp_per_layer;
if (ubatch.token) {
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);