public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-18 04:35:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2026-01-30 15:56:40 +02:00
parent b529c0610f
commit 953e503fd9
38 changed files with 2299 additions and 730 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/models/deepseek2.cpp
+14 -14
View File
@@ -2,14 +2,11 @@
llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
// lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
const bool is_mla = hparams.is_mla();
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
@@ -43,7 +40,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv();
auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;
auto * inp_attn_k = is_mla ? build_attn_inp_k() : nullptr;
ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -57,6 +55,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// self_attention
{
ggml_tensor * q = NULL;
const bool is_lite = model.layers[il].wq;
if (!is_lite) {
q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
cb(q, "q", il);
@@ -124,14 +125,14 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
// note: rope must go first for in-place context shifting in build_rope_shift()
ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
cb(Qcur, "Qcur", il);
kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
cb(kv_cmpr, "kv_cmpr_reshape", il);
// {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
cb(Kcur, "Kcur", il);
// {kv_lora_rank, 1, n_tokens}
@@ -145,7 +146,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
}
// note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
cur = build_attn(inp_attn,
cur = build_attn(inp_attn_k,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
} else {
@@ -169,11 +170,10 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
Vcur = ggml_cont(ctx0, Vcur);
cb(Vcur, "Vcur_cont", il);
// note: rope must go first for in-place context shifting in build_rope_shift()
ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
cb(Kcur, "Kcur", il);
if (inp_attn_scale) {
@@ -183,7 +183,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
}
// note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
cur = build_attn(inp_attn,
cur = build_attn(inp_attn_kv,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
examples/talk-llama/models/exaone-moe.cpp
+146
View File
@@ -0,0 +1,146 @@
#include "models.h"
llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_k;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
GGML_ASSERT(n_embd_head == hparams.n_rot);
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn_iswa = build_attn_inp_kv_iswa();
ggml_tensor * inp_out_ids = build_inp_out_ids();
const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
for (int il = 0; il < n_transformer_layers; ++il) {
ggml_tensor * inpSA = inpL;
// use RoPE for SWA layers
const bool is_local_layer = hparams.is_swa(il);
// norm
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// self-attention
{
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
cb(Kcur, "Kcur_normed", il);
if (is_local_layer) {
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
}
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn_iswa,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
cb(cur, "attn_out", il);
}
if (il == n_transformer_layers - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// norm
cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
// feed-forward network
if (model.layers[il].ffn_gate_inp == nullptr) {
// dense branch
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL, NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
} else {
// MoE branch
ggml_tensor * moe_out = build_moe_ffn(cur,
model.layers[il].ffn_gate_inp,
model.layers[il].ffn_up_exps,
model.layers[il].ffn_gate_exps,
model.layers[il].ffn_down_exps,
model.layers[il].ffn_exp_probs_b,
n_expert, n_expert_used,
LLM_FFN_SILU, hparams.expert_weights_norm,
true, hparams.expert_weights_scale,
(llama_expert_gating_func_type) hparams.expert_gating_func,
il);
cb(moe_out, "ffn_moe_out", il);
// FFN shared expert
{
ggml_tensor * ffn_shexp =
build_ffn(cur,
model.layers[il].ffn_up_shexp, NULL, NULL,
model.layers[il].ffn_gate_shexp, NULL, NULL,
model.layers[il].ffn_down_shexp, NULL, NULL,
NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(ffn_shexp, "ffn_shexp", il);
cur = ggml_add(ctx0, moe_out, ffn_shexp);
cb(cur, "ffn_out", il);
}
}
cur = ggml_add(ctx0, cur, ffn_inp);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
// final norm
cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur);
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
examples/talk-llama/models/gemma3n-iswa.cpp
+6 -6
View File
@@ -245,12 +245,12 @@ 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>(n_embd);
ggml_tensor * inp_per_layer;
if (ubatch.token) {
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
ggml_set_input(inp->tokens);
res->t_tokens = inp->tokens;
res->t_inp_tokens = inp->tokens;
inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
@@ -258,12 +258,12 @@ ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
res->add_input(std::move(inp));
} else {
// Vision embedding path: use padding token (ID=0) embedding
// TODO: verify if this is the correct behavior in transformers implementation
const int64_t embd_size = model.tok_embd_per_layer->ne[0]; // n_embd_altup * n_layer
// Extract and dequantize padding token embedding (column 0)
ggml_tensor * padding_q = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
ggml_tensor * padding_f32 = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, embd_size);
inp_per_layer = ggml_cpy(ctx0, padding_q, padding_f32);
// Extract and dequantize padding token embedding (row 0)
ggml_tensor * padding = ggml_view_1d(ctx0, model.tok_embd_per_layer, embd_size, 0);
inp_per_layer = ggml_cast(ctx0, padding, GGML_TYPE_F32);
// Reshape to [n_embd_altup, n_layer, 1]
inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, 1);
examples/talk-llama/models/minicpm3.cpp
+1
View File
@@ -9,6 +9,7 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
const uint32_t n_embd_head_qk_rope = hparams.n_rot;
const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
const uint32_t kv_lora_rank = hparams.n_lora_kv;
ggml_tensor * cur;
examples/talk-llama/models/models.h
+4
View File
@@ -167,6 +167,10 @@ struct llm_build_exaone : public llm_graph_context {
llm_build_exaone(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_exaone_moe : public llm_graph_context {
llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params);
};
struct llm_build_falcon : public llm_graph_context {
llm_build_falcon(const llama_model & model, const llm_graph_params & params);
};
examples/talk-llama/models/nemotron-h.cpp
+1 -1
View File
@@ -67,7 +67,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
const llama_model & model,
const int64_t n_embd_head,
const int il) {
// compute Q and K and (optionally) RoPE them
// compute Q and K
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
examples/talk-llama/models/plm.cpp
+1
View File
@@ -5,6 +5,7 @@ llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params &
const uint32_t n_embd_head_qk_rope = hparams.n_rot;
const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
const uint32_t kv_lora_rank = hparams.n_lora_kv;
ggml_tensor * cur;
examples/talk-llama/models/qwen3vl-moe.cpp
+5 -14
View File
@@ -2,7 +2,8 @@
llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16,17 +17,6 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
if (ubatch.embd) {
// Image input: split main embd and deepstack embds
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
for (size_t i = 0; i < n_deepstack_layers; i++) {
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
}
inpL = inpL_main;
}
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
@@ -120,8 +110,9 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
if (ubatch.embd && (size_t)il < n_deepstack_layers) {
cur = ggml_add(ctx0, cur, deepstack_features[il]);
if (il < (int) n_deepstack_layers) {
ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
cur = ggml_add(ctx0, cur, ds);
cb(cur, "deepstack_out", il);
}
examples/talk-llama/models/qwen3vl.cpp
+5 -14
View File
@@ -2,7 +2,8 @@
llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16,17 +17,6 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
if (ubatch.embd) {
// Image input: split main embd and deepstack embds
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
for (size_t i = 0; i < n_deepstack_layers; i++) {
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
}
inpL = inpL_main;
}
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
@@ -113,8 +103,9 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
if (ubatch.embd && (size_t)il < n_deepstack_layers) {
cur = ggml_add(ctx0, cur, deepstack_features[il]);
if (il < (int) n_deepstack_layers) {
ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
cur = ggml_add(ctx0, cur, ds);
cb(cur, "deepstack_out", il);
}