ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-16 05:48:09 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2024-12-17 21:19:28 +02:00
parent eb97b257eb
commit 61edb117a0
6 changed files with 597 additions and 200 deletions
Download Patch File Download Diff File Expand all files Collapse all files

518
examples/talk-llama/llama.cpp
View File

@ -163,6 +163,7 @@ enum llm_arch {
LLM_ARCH_QWEN,
LLM_ARCH_QWEN2,
LLM_ARCH_QWEN2MOE,
LLM_ARCH_QWEN2VL,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PLAMO,
@ -183,6 +184,7 @@ enum llm_arch {
LLM_ARCH_OLMOE,
LLM_ARCH_OPENELM,
LLM_ARCH_ARCTIC,
LLM_ARCH_DEEPSEEK,
LLM_ARCH_DEEPSEEK2,
LLM_ARCH_CHATGLM,
LLM_ARCH_BITNET,
@ -217,6 +219,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_QWEN, "qwen" },
{ LLM_ARCH_QWEN2, "qwen2" },
{ LLM_ARCH_QWEN2MOE, "qwen2moe" },
{ LLM_ARCH_QWEN2VL, "qwen2vl" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PHI3, "phi3" },
{ LLM_ARCH_PLAMO, "plamo" },
@ -237,6 +240,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_OLMOE, "olmoe" },
{ LLM_ARCH_OPENELM, "openelm" },
{ LLM_ARCH_ARCTIC, "arctic" },
{ LLM_ARCH_DEEPSEEK, "deepseek" },
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_BITNET, "bitnet" },
@ -308,6 +312,7 @@ enum llm_kv {
LLM_KV_ATTENTION_SCALE,
LLM_KV_ROPE_DIMENSION_COUNT,
LLM_KV_ROPE_DIMENSION_SECTIONS,
LLM_KV_ROPE_FREQ_BASE,
LLM_KV_ROPE_SCALE_LINEAR,
LLM_KV_ROPE_SCALING_TYPE,
@ -424,6 +429,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" },
{ LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" },
{ LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" },
{ LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
{ LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
@ -898,6 +904,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_QWEN2VL,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_QWEN2MOE,
{
@ -1288,6 +1311,33 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
},
},
{
LLM_ARCH_DEEPSEEK,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" },
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
{ LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
{ LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" },
{ LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" },
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
},
},
{
LLM_ARCH_DEEPSEEK2,
{
@ -1562,6 +1612,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
LLM_CHAT_TEMPLATE_MISTRAL_V7,
LLM_CHAT_TEMPLATE_PHI_3,
LLM_CHAT_TEMPLATE_FALCON_3,
LLM_CHAT_TEMPLATE_ZEPHYR,
LLM_CHAT_TEMPLATE_MONARCH,
LLM_CHAT_TEMPLATE_GEMMA,
@ -1579,6 +1630,7 @@ enum llm_chat_template {
LLM_CHAT_TEMPLATE_EXAONE_3,
LLM_CHAT_TEMPLATE_RWKV_WORLD,
LLM_CHAT_TEMPLATE_GRANITE,
LLM_CHAT_TEMPLATE_GIGACHAT,
LLM_CHAT_TEMPLATE_UNKNOWN,
};
@ -1593,6 +1645,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
{ "mistral-v7", LLM_CHAT_TEMPLATE_MISTRAL_V7 },
{ "phi3", LLM_CHAT_TEMPLATE_PHI_3 },
{ "falcon3", LLM_CHAT_TEMPLATE_FALCON_3 },
{ "zephyr", LLM_CHAT_TEMPLATE_ZEPHYR },
{ "monarch", LLM_CHAT_TEMPLATE_MONARCH },
{ "gemma", LLM_CHAT_TEMPLATE_GEMMA },
@ -1610,6 +1663,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
{ "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 },
{ "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD },
{ "granite", LLM_CHAT_TEMPLATE_GRANITE },
{ "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT },
};
static llm_arch llm_arch_from_string(const std::string & name) {
@ -1794,7 +1848,7 @@ private:
DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
if (!bufLen) {
ret = format("Win32 error code: %s", error_code);
ret = format("Win32 error code: %lx", error_code);
} else {
ret = lpMsgBuf;
LocalFree(lpMsgBuf);
@ -2132,7 +2186,7 @@ struct llama_mmap {
HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
// may fail on pre-Windows 8 systems
pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
if (pPrefetchVirtualMemory) {
// advise the kernel to preload the mapped memory
@ -2474,11 +2528,12 @@ struct llama_hparams {
uint32_t time_decay_extra_dim = 0;
uint32_t wkv_head_size = 0;
float rope_attn_factor = 1.0f;
float rope_freq_base_train;
float rope_freq_scale_train;
uint32_t n_ctx_orig_yarn;
float rope_yarn_log_mul;
float rope_attn_factor = 1.0f;
float rope_freq_base_train;
float rope_freq_scale_train;
uint32_t n_ctx_orig_yarn;
float rope_yarn_log_mul;
int rope_sections[4];
// for State Space Models
uint32_t ssm_d_conv = 0;
@ -2535,6 +2590,9 @@ struct llama_hparams {
if (this->rope_finetuned != other.rope_finetuned) return true;
if (this->n_ctx_orig_yarn != other.n_ctx_orig_yarn) return true;
if (std::equal(std::begin(this->rope_sections),
std::end(this->rope_sections),
std::begin(other.rope_sections))) return true;
if (this->ssm_d_conv != other.ssm_d_conv) return true;
if (this->ssm_d_inner != other.ssm_d_inner) return true;
@ -3378,6 +3436,11 @@ struct llama_context {
// whether we are computing encoder output or decoder output
bool is_encoding = false;
// TODO: find a better way to accommodate mutli-dimension position encoding methods
// number of position id each token get, 1 for each token in most cases.
// when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
int n_pos_per_token = 1;
// output of the encoder part of the encoder-decoder models
std::vector<float> embd_enc;
std::vector<std::set<llama_seq_id>> seq_ids_enc;
@ -4578,9 +4641,6 @@ struct llama_model_loader {
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@ -5344,9 +5404,6 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
default: return "unknown, may not work";
}
@ -5753,6 +5810,13 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_QWEN2VL:
{
std::array<int, 4> section_dims;
ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, section_dims, 4, true);
std::copy(section_dims.begin(), section_dims.begin() + 4, std::begin(hparams.rope_sections));
}
// fall through
case LLM_ARCH_QWEN2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@ -6063,6 +6127,19 @@ static void llm_load_hparams(
model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_DEEPSEEK:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
switch (hparams.n_layer) {
case 28: model.type = e_model::MODEL_20B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_DEEPSEEK2:
{
bool is_lite = (hparams.n_layer == 27);
@ -6398,6 +6475,11 @@ static void llm_load_vocab(
} else if (
tokenizer_pre == "falcon") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_FALCON;
} else if (
tokenizer_pre == "falcon3") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
vocab.tokenizer_ignore_merges = true;
vocab.tokenizer_add_bos = true;
} else if (
tokenizer_pre == "mpt") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MPT;
@ -6409,6 +6491,7 @@ static void llm_load_vocab(
tokenizer_pre == "phi-2" ||
tokenizer_pre == "jina-es" ||
tokenizer_pre == "jina-de" ||
tokenizer_pre == "gigachat" ||
tokenizer_pre == "jina-v1-en" ||
tokenizer_pre == "jina-v2-es" ||
tokenizer_pre == "jina-v2-de" ||
@ -6479,6 +6562,9 @@ static void llm_load_vocab(
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
vocab.tokenizer_add_bos = true;
vocab.tokenizer_clean_spaces = false;
} else if (
tokenizer_pre == "minerva-7b") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MINERVA;
} else {
throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
}
@ -7057,6 +7143,13 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, vocab.max_token_len);
if (model.arch == LLM_ARCH_DEEPSEEK) {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
}
if (model.arch == LLM_ARCH_DEEPSEEK2) {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
@ -8170,6 +8263,7 @@ static bool llm_load_tensors(
}
} break;
case LLM_ARCH_QWEN2:
case LLM_ARCH_QWEN2VL:
{
model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@ -8830,6 +8924,55 @@ static bool llm_load_tensors(
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
}
} break;
case LLM_ARCH_DEEPSEEK:
{
const int64_t n_ff_exp = hparams.n_ff_exp;
const int64_t n_expert_shared = hparams.n_expert_shared;
model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// output
model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
model.output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
for (int i = 0; i < n_layer; ++i) {
auto & layer = model.layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
if (i < (int) hparams.n_layer_dense_lead) {
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
} else {
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
if (n_expert == 0) {
throw std::runtime_error("n_expert must be > 0");
}
if (n_expert_used == 0) {
throw std::runtime_error("n_expert_used must be > 0");
}
// MoE branch
layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0);
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0);
// Shared expert branch
layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, 0);
layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
}
}
} break;
case LLM_ARCH_DEEPSEEK2:
{
const bool is_lite = (hparams.n_layer == 27);
@ -12559,6 +12702,124 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_qwen2vl() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
// inp_pos - contains the positions
lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens * 4);
cb(lctx.inp_pos, "inp_pos", -1);
ggml_set_input(lctx.inp_pos);
struct ggml_tensor * inp_pos = lctx.inp_pos;
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
// norm
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_multi(
ctx0,
ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_multi(
ctx0,
ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// feed-forward network
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, lctx, 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, cb, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cur = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm, NULL,
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
// lm_head
cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_qwen2moe() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
@ -15066,6 +15327,161 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_deepseek() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
// mutable variable, needed during the last layer of the computation to skip unused tokens
int32_t n_tokens = this->n_tokens;
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = build_inp_pos();
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
// norm
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// rope freq factors for llama3; may return nullptr for llama2 and other models
struct ggml_tensor * rope_factors = build_rope_factors(il);
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
if (model.layers[il].bq) {
Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
cb(Qcur, "Qcur", il);
}
struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
if (model.layers[il].bk) {
Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
cb(Kcur, "Kcur", il);
}
struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
if (model.layers[il].bv) {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), 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);
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
n_tokens = n_outputs;
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
if ((uint32_t) il < hparams.n_layer_dense_lead) {
cur = llm_build_ffn(ctx0, lctx, 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, cb, il);
cb(cur, "ffn_out", il);
} else {
// MoE branch
ggml_tensor * moe_out =
llm_build_moe_ffn(ctx0, lctx, 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,
n_expert, n_expert_used,
LLM_FFN_SILU, false,
false, hparams.expert_weights_scale,
cb, il);
cb(moe_out, "ffn_moe_out", il);
// FFN shared expert
{
ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, 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, cb, 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 = lctx.cvec.apply_to(ctx0, cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm, NULL,
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
// lm_head
cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_deepseek2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
@ -16660,6 +17076,11 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_qwen2();
} break;
case LLM_ARCH_QWEN2VL:
{
lctx.n_pos_per_token = 4;
result = llm.build_qwen2vl();
} break;
case LLM_ARCH_QWEN2MOE:
{
result = llm.build_qwen2moe();
@ -16748,6 +17169,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_arctic();
} break;
case LLM_ARCH_DEEPSEEK:
{
result = llm.build_deepseek();
} break;
case LLM_ARCH_DEEPSEEK2:
{
result = llm.build_deepseek2();
@ -16878,8 +17303,8 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch)
if (ubatch.pos && lctx.inp_pos) {
const int64_t n_tokens = ubatch.n_tokens;
ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
auto n_pos = lctx.n_pos_per_token;
ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*n_pos*ggml_element_size(lctx.inp_pos));
}
if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
@ -18364,10 +18789,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
new_type == GGML_TYPE_Q4_0_8_8) {
new_type = GGML_TYPE_Q4_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
new_type = GGML_TYPE_Q4_K;
}
@ -18690,9 +19111,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
@ -19031,14 +19449,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data = (float *) f32_conv_buf.data();
}
int chunk_size_multiplier = 1;
if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);
@ -19051,8 +19461,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const int64_t nrows = tensor->ne[1];
static const int64_t min_chunk_size = 32 * 512;
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
chunk_size_multiplier;
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
@ -19995,6 +20404,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_COMMAND_R:
case LLM_ARCH_OLMO:
case LLM_ARCH_ARCTIC:
case LLM_ARCH_DEEPSEEK:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_CHATGLM:
case LLM_ARCH_GRANITE:
@ -20028,6 +20438,9 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_MINICPM3:
return LLAMA_ROPE_TYPE_NEOX;
case LLM_ARCH_QWEN2VL:
return LLAMA_ROPE_TYPE_MROPE;
// all model arches should be listed explicitly here
case LLM_ARCH_UNKNOWN:
GGML_ABORT("unknown architecture");
@ -21596,7 +22009,7 @@ float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs));
}
} else if ((size_t) i >= ctx->output_ids.size()) {
throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size()));
throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size()));
} else {
j = ctx->output_ids[i];
}
@ -21813,6 +22226,8 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
}
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
return LLM_CHAT_TEMPLATE_PHI_3;
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
return LLM_CHAT_TEMPLATE_FALCON_3;
} else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
return LLM_CHAT_TEMPLATE_ZEPHYR;
} else if (tmpl_contains("bos_token + message['role']")) {
@ -21857,6 +22272,8 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_RWKV_WORLD;
} else if (tmpl_contains("<|start_of_role|>")) {
return LLM_CHAT_TEMPLATE_GRANITE;
} else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {
return LLM_CHAT_TEMPLATE_GIGACHAT;
}
return LLM_CHAT_TEMPLATE_UNKNOWN;
}
@ -21963,6 +22380,15 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) {
ss << "<|assistant|>\n";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_FALCON_3) {
// Falcon 3
for (auto message : chat) {
std::string role(message->role);
ss << "<|" << role << "|>\n" << message->content << "\n";
}
if (add_ass) {
ss << "<|assistant|>\n";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) {
// zephyr template
for (auto message : chat) {
@ -22180,6 +22606,32 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) {
ss << "<|start_of_role|>assistant<|end_of_role|>\n";
}
} else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
// GigaChat template
bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
// Handle system message if present
if (has_system) {
ss << "<s>" << chat[0]->content << "<|message_sep|>";
} else {
ss << "<s>";
}
// Process remaining messages
for (size_t i = has_system ? 1 : 0; i < chat.size(); i++) {
std::string role(chat[i]->role);
if (role == "user") {
ss << "user<|role_sep|>" << chat[i]->content << "<|message_sep|>"
<< "available functions<|role_sep|>[]<|message_sep|>";
} else if (role == "assistant") {
ss << "assistant<|role_sep|>" << chat[i]->content << "<|message_sep|>";
}
}
// Add generation prompt if needed
if (add_ass) {
ss << "assistant<|role_sep|>";
}
} else {
// template not supported
return -1;