happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into HEAD

This commit is contained in:
Georgi Gerganov 2025-12-14 16:44:30 +02:00
parent 22c7f85b9c 9e6649ecf2
commit 0086c246ee
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
5 changed files with 21 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp
View File

@ -45,7 +45,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
vec4 sum_v = fma(b_val_0, fbits0 + delta_v, vec4(0.0));
sum_v = fma(b_val_1, fbits1 + delta_v, sum_v);
FLOAT_TYPE sum = dot(sum_v, vec4(1.0));
FLOAT_TYPE sum = dot(sum_v, vec4(1.0));
temp[j][n] = fma(dl, sum, temp[j][n]);
ibi += num_blocks_per_row;

4
src/llama-graph.cpp
View File

@ -79,7 +79,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
for (int i = 0; i < n_tokens; ++i) {
const float pos = ubatch->pos[i];
attn_scale_data[i] = std::log(
std::floor((pos + 1.0f) / n_attn_temp_floor_scale) + 1.0
std::floor((pos + f_attn_temp_offset) / n_attn_temp_floor_scale) + 1.0
) * f_attn_temp_scale + 1.0;
}
@ -1278,7 +1278,7 @@ ggml_tensor * llm_graph_context::build_inp_pos() const {
}
ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale, hparams.f_attn_temp_offset);
auto & cur = inp->attn_scale;

5
src/llama-graph.h
View File

@ -133,8 +133,8 @@ public:
// temperature tuning, used by llama4
class llm_graph_input_attn_temp : public llm_graph_input_i {
public:
llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
: n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale, float f_attn_temp_offset)
: n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale), f_attn_temp_offset(f_attn_temp_offset) {}
virtual ~llm_graph_input_attn_temp() = default;
void set_input(const llama_ubatch * ubatch) override;
@ -143,6 +143,7 @@ public:
const uint32_t n_attn_temp_floor_scale;
const float f_attn_temp_scale;
const float f_attn_temp_offset;
};
class llm_graph_input_pos_bucket : public llm_graph_input_i {

1
src/llama-hparams.h
View File

@ -165,6 +165,7 @@ struct llama_hparams {
uint32_t n_no_rope_layer_step = 4;
uint32_t n_attn_temp_floor_scale = 0;
float f_attn_temp_scale = 0.0f;
float f_attn_temp_offset = 0.0f; // offset position index
// gemma3n altup
uint32_t n_altup = 4; // altup_num_inputs

5
src/llama-model.cpp
View File

@ -668,6 +668,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.n_swa = 8192;
hparams.n_attn_temp_floor_scale = 8192;
hparams.f_attn_temp_scale = 0.1f;
hparams.f_attn_temp_offset = 1.0f;
hparams.set_swa_pattern(4); // pattern: 3 chunked - 1 full
}
@ -1646,6 +1647,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
hparams.f_attn_temp_offset = 0.0f;
switch (hparams.n_layer) {
case 27: type = LLM_TYPE_16B; break;
case 60: type = LLM_TYPE_236B; break;
@ -2276,6 +2279,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow, false);
ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f);
hparams.f_attn_temp_offset = 0.0f;
// TODO: maybe add n_attn_temp_floor_scale as a separate KV?
if (hparams.f_attn_temp_scale != 0.0f) {
hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;