extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-08-18 06:10:10 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp (#3084)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-04-28 16:40:23 +03:00
committed by GitHub
parent 28dcdff4c5
commit f3c42399a3
36 changed files with 16940 additions and 12400 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
examples/talk-llama/llama-hparams.h
View File

@@ -36,12 +36,17 @@ struct llama_hparams {
uint32_t n_layer;
uint32_t n_rot;
uint32_t n_swa = 0; // sliding window attention (SWA)
uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention
uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
uint32_t n_expert = 0;
uint32_t n_expert_used = 0;
uint32_t n_rel_attn_bkts = 0;
// note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
uint32_t n_embd_head_k_mla = 0;
uint32_t n_embd_head_v_mla = 0;
// for WavTokenizer
struct llama_hparams_posnet posnet;
struct llama_hparams_convnext convnext;
@@ -75,10 +80,16 @@ struct llama_hparams {
uint32_t time_decay_extra_dim = 0;
uint32_t wkv_head_size = 0;
uint32_t token_shift_count = 2;
uint32_t n_lora_decay = 0;
uint32_t n_lora_iclr = 0;
uint32_t n_lora_value_res_mix = 0;
uint32_t n_lora_gate = 0;
float rope_attn_factor = 1.0f;
float rope_freq_base_train;
float rope_freq_base_train_swa;
float rope_freq_scale_train;
float rope_freq_scale_train_swa;
uint32_t n_ctx_orig_yarn;
float rope_yarn_log_mul;
@@ -105,6 +116,14 @@ struct llama_hparams {
bool use_alibi = false;
bool attn_soft_cap = false;
uint32_t n_moe_layer_step = 0;
bool use_kq_norm = true;
uint32_t n_attn_chunk = 0;
// values below seems to be fixed on llama4
uint32_t n_no_rope_layer_step = 4;
uint32_t n_attn_temp_floor_scale = 8192;
float f_attn_temp_scale = 0.1;
// needed by encoder-decoder models (e.g. T5, FLAN-T5)
// ref: https://github.com/ggerganov/llama.cpp/pull/8141
llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -133,6 +152,8 @@ struct llama_hparams {
// dimension of the recurrent state embeddings
uint32_t n_embd_v_s() const;
bool is_swa(uint32_t il) const;
};
static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");