mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-11 18:36:41 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

llama: update vendor code to commit ba1cb19c (#8101)

Browse source
This commit is contained in:
Jeffrey Morgan 2024-12-14 14:55:51 -08:00 • committed by GitHub
parent 60f75560a2
commit 7a81daf026
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
273 changed files with 3194 additions and 1900 deletions
Download patch file Download diff file Expand all files Collapse all files

388
llama/llama.cpp vendored
View file

@ -1,5 +1,5 @@
/**
* llama.cpp - commit 40c6d79fb52f995f47507fedfeaae2ac05d9b35c - do not edit this file
* llama.cpp - commit ba1cb19cdd0d92e012e0f6e009e0620f854b6afd - do not edit this file
*
* MIT License
*
@ -190,6 +190,7 @@ enum llm_arch {
LLM_ARCH_QWEN,
LLM_ARCH_QWEN2,
LLM_ARCH_QWEN2MOE,
LLM_ARCH_QWEN2VL,
LLM_ARCH_PHI2,
LLM_ARCH_PHI3,
LLM_ARCH_PLAMO,
@ -246,6 +247,7 @@ 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" },
@ -340,6 +342,7 @@ enum llm_kv {
LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS,
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,
@ -458,6 +461,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_CROSS_ATTENTION_LAYERS, "%s.attention.cross_attention_layers" },
{ 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" },
@ -975,6 +979,23 @@ 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,
{
@ -1113,6 +1134,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" },
{ LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
@ -1778,9 +1801,10 @@ struct LLM_TN {
//
static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_TYPES = {
{ LLAMA_ROPE_SCALING_TYPE_NONE, "none" },
{ LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" },
{ LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" },
{ LLAMA_ROPE_SCALING_TYPE_NONE, "none" },
{ LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" },
{ LLAMA_ROPE_SCALING_TYPE_YARN, "yarn" },
{ LLAMA_ROPE_SCALING_TYPE_LONGROPE, "longrope" },
};
static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
@ -1886,7 +1910,7 @@ 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);
@ -2224,7 +2248,7 @@ 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
@ -2571,11 +2595,12 @@ 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;
@ -2634,6 +2659,9 @@ 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;
@ -3504,6 +3532,11 @@ 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;
@ -4739,9 +4772,6 @@ 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));
@ -5505,9 +5535,6 @@ 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";
}
@ -5756,8 +5783,12 @@ static void llm_load_hparams(
case LLM_ARCH_MINICPM:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale);
ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
switch (hparams.n_layer) {
case 52: model.type = e_model::MODEL_1B; break;
case 40: model.type = e_model::MODEL_2B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
@ -5922,6 +5953,13 @@ 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);
@ -6654,6 +6692,9 @@ 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 {
LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
@ -7248,7 +7289,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp);
}
if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
if (model.arch == LLM_ARCH_MINICPM || model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale);
LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@ -7882,7 +7923,13 @@ static bool llm_load_tensors(
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
}
else {
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
}
if (n_expert == 0) {
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
@ -8396,6 +8443,7 @@ 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);
@ -13064,6 +13112,124 @@ 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);
@ -14015,153 +14181,6 @@ struct llm_build_context {
return gf;
}
// ref: https://arxiv.org/abs/2203.03466
// https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738
// based on the original build_llama() function
struct ggml_cgraph * build_minicpm() {
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);
const int64_t n_embd = hparams.n_embd;
//TODO: if the model varies, these parameters need to be read from the model
const int64_t n_embd_base = 256;
const float scale_embd = 12.0f;
const float scale_depth = 1.4f;
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
// scale the input embeddings
inpL = ggml_scale(ctx0, inpL, scale_embd);
cb(inpL, "inp_scaled", -1);
// 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();
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);
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, nullptr,
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, nullptr,
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, 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);
}
// scale_res - scale the hidden states for residual connection
const float scale_res = scale_depth/sqrtf(float(n_layer));
cur = ggml_scale(ctx0, cur, scale_res);
cb(cur, "hidden_scaled", -1);
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);
}
// scale the hidden states for residual connection
cur = ggml_scale(ctx0, cur, scale_res);
cb(cur, "hidden_scaled_ffn", -1);
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 scaling
const float scale_lmhead = float(n_embd_base)/float(n_embd);
cur = ggml_scale(ctx0, cur, scale_lmhead);
cb(cur, "lmhead_scaling", -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_minicpm3() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
@ -17412,6 +17431,7 @@ static struct ggml_cgraph * llama_build_graph(
switch (model.arch) {
case LLM_ARCH_LLAMA:
case LLM_ARCH_MINICPM:
case LLM_ARCH_GRANITE:
case LLM_ARCH_GRANITE_MOE:
{
@ -17467,6 +17487,11 @@ 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();
@ -17499,10 +17524,6 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_internlm2();
} break;
case LLM_ARCH_MINICPM:
{
result = llm.build_minicpm();
} break;
case LLM_ARCH_MINICPM3:
{
result = llm.build_minicpm3();
@ -17702,8 +17723,8 @@ 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) {
@ -19191,10 +19212,6 @@ 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;
}
@ -19517,9 +19534,6 @@ 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));
}
@ -19860,14 +19874,6 @@ 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);
@ -19880,8 +19886,7 @@ 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;
@ -20859,6 +20864,9 @@ 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");
@ -22434,7 +22442,7 @@ 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];
}