vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/nix/package.nix
#	.github/workflows/build.yml
#	CMakeLists.txt
#	Makefile
#	README.md
#	flake.nix
#	llama.cpp
#	llama.h
#	tests/test-llama-grammar.cpp
This commit is contained in:
Concedo 2024-02-04 19:42:57 +08:00
commit 6dc01297f8
36 changed files with 2917 additions and 10930 deletions
Download patch file Download diff file Expand all files Collapse all files

518
llama.cpp
View file

@ -228,10 +228,11 @@ enum llm_arch {
LLM_ARCH_PLAMO,
LLM_ARCH_CODESHELL,
LLM_ARCH_ORION,
LLM_ARCH_INTERNLM2,
LLM_ARCH_UNKNOWN,
};
static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_LLAMA, "llama" },
{ LLM_ARCH_FALCON, "falcon" },
{ LLM_ARCH_GPT2, "gpt2" },
@ -250,6 +251,7 @@ static std::map<llm_arch, std::string> LLM_ARCH_NAMES = {
{ LLM_ARCH_PLAMO, "plamo" },
{ LLM_ARCH_CODESHELL, "codeshell" },
{ LLM_ARCH_ORION, "orion" },
{ LLM_ARCH_INTERNLM2, "internlm2" },
};
enum llm_kv {
@ -302,11 +304,12 @@ enum llm_kv {
LLM_KV_TOKENIZER_PAD_ID,
LLM_KV_TOKENIZER_ADD_BOS,
LLM_KV_TOKENIZER_ADD_EOS,
LLM_KV_TOKENIZER_ADD_PREFIX,
LLM_KV_TOKENIZER_HF_JSON,
LLM_KV_TOKENIZER_RWKV,
};
static std::map<llm_kv, std::string> LLM_KV_NAMES = {
static std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" },
{ LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" },
{ LLM_KV_GENERAL_ALIGNMENT, "general.alignment" },
@ -356,6 +359,7 @@ static std::map<llm_kv, std::string> LLM_KV_NAMES = {
{ LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" },
{ LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" },
{ LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" },
{ LLM_KV_TOKENIZER_ADD_PREFIX, "tokenizer.ggml.add_space_prefix" },
{ LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" },
{ LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" },
};
@ -366,7 +370,7 @@ struct LLM_KV {
llm_arch arch;
std::string operator()(llm_kv kv) const {
return ::format(LLM_KV_NAMES[kv].c_str(), LLM_ARCH_NAMES[arch].c_str());
return ::format(LLM_KV_NAMES[kv], LLM_ARCH_NAMES[arch]);
}
};
@ -693,7 +697,23 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_INTERNLM2,
{
{ 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_UNKNOWN,
{
@ -751,13 +771,13 @@ struct LLM_TN {
// gguf helpers
//
static std::map<int8_t, std::string> LLAMA_ROPE_SCALING_TYPES = {
static std::map<int32_t, const char *> LLAMA_ROPE_SCALING_TYPES = {
{ LLAMA_ROPE_SCALING_NONE, "none" },
{ LLAMA_ROPE_SCALING_LINEAR, "linear" },
{ LLAMA_ROPE_SCALING_YARN, "yarn" },
};
static int8_t llama_rope_scaling_type_from_string(const std::string & name) {
static int32_t llama_rope_scaling_type_from_string(const std::string & name) {
for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) {
if (kv.second == name) {
return kv.first;
@ -1405,6 +1425,7 @@ enum e_model {
MODEL_13B,
MODEL_14B,
MODEL_15B,
MODEL_20B,
MODEL_30B,
MODEL_34B,
MODEL_40B,
@ -1422,6 +1443,7 @@ static const size_t GiB = 1024*MiB;
struct llama_hparams {
bool vocab_only;
bool rope_finetuned;
uint32_t n_vocab;
uint32_t n_ctx_train; // context size the model was trained on
uint32_t n_embd;
@ -1441,8 +1463,7 @@ struct llama_hparams {
float rope_freq_base_train;
float rope_freq_scale_train;
uint32_t n_yarn_orig_ctx;
int8_t rope_scaling_type_train : 3;
bool rope_finetuned : 1;
int32_t rope_scaling_type_train;
float f_clamp_kqv;
float f_max_alibi_bias;
@ -1646,6 +1667,7 @@ struct llama_vocab {
id special_suffix_id = 32008;
id special_eot_id = 32010;
bool add_space_prefix = true;
int find_bpe_rank(std::string token_left, std::string token_right) const {
// GGML_ASSERT(token_left.find(" ") == std::string::npos);
// GGML_ASSERT(token_left.find("\n") == std::string::npos);
@ -2725,7 +2747,7 @@ struct llama_model_loader {
// load LLaMA models
//
static std::string llama_model_arch_name(llm_arch arch) {
static const char * llama_model_arch_name(llm_arch arch) {
auto it = LLM_ARCH_NAMES.find(arch);
if (it == LLM_ARCH_NAMES.end()) {
return "unknown";
@ -2778,6 +2800,7 @@ static const char * llama_model_type_name(e_model type) {
case MODEL_13B: return "13B";
case MODEL_14B: return "14B";
case MODEL_15B: return "15B";
case MODEL_20B: return "20B";
case MODEL_30B: return "30B";
case MODEL_34B: return "34B";
case MODEL_40B: return "40B";
@ -2790,6 +2813,14 @@ static const char * llama_model_type_name(e_model type) {
default: return "?B";
}
}
static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
switch (type) {
case LLAMA_VOCAB_TYPE_SPM: return "SPM";
case LLAMA_VOCAB_TYPE_BPE: return "BPE";
default: return "unknown";
}
}
static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
model.arch = ml.get_arch();
@ -3053,6 +3084,15 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_INTERNLM2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 32: model.type = e_model::MODEL_7B; break;
case 48: model.type = e_model::MODEL_20B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
default: (void)0;
}
@ -3105,6 +3145,11 @@ static void llm_load_vocab(
vocab.special_unk_id = 0;
vocab.special_sep_id = -1;
vocab.special_pad_id = -1;
const int add_space_prefix_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_ADD_PREFIX).c_str());
if (add_space_prefix_keyidx != -1) {
vocab.add_space_prefix = gguf_get_val_bool(ctx, add_space_prefix_keyidx);
} // The default value of add_space_prefix is true.
} else if (tokenizer_name == "gpt2") {
vocab.type = LLAMA_VOCAB_TYPE_BPE;
@ -3330,12 +3375,12 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
const auto & hparams = model.hparams;
const auto & vocab = model.vocab;
const auto rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
// hparams
LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml.fver));
LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch).c_str());
LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, vocab.type == LLAMA_VOCAB_TYPE_SPM ? "SPM" : "BPE"); // TODO: fix
LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch));
LLAMA_LOG_INFO("%s: vocab type = %s\n", __func__, llama_model_vocab_type_name(vocab.type));
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab);
LLAMA_LOG_INFO("%s: n_merges = %u\n", __func__, (int) vocab.bpe_ranks.size());
LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train);
@ -3356,7 +3401,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff);
LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert);
LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used);
LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str());
LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type);
LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train);
LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train);
LLAMA_LOG_INFO("%s: n_yarn_orig_ctx = %u\n", __func__, hparams.n_yarn_orig_ctx);
@ -4091,8 +4136,35 @@ static bool llm_load_tensors(
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
}
} break;
case LLM_ARCH_INTERNLM2:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
}
for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(i);
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
// layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -4739,126 +4811,6 @@ struct llm_build_context {
ctx0 = nullptr;
}
}
struct ggml_cgraph * build_orion() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, 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, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
cb(inp_pos, "inp_pos", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
cb(KQ_mask, "KQ_mask", -1);
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
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, model.layers[il].attn_norm_b,
LLM_NORM, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(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 = ggml_mul_mat(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 = ggml_mul_mat(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_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
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, model.layers[il].ffn_norm_b,
LLM_NORM, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL,
model.layers[il].ffn_gate, NULL,
model.layers[il].ffn_down, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm, model.output_norm_b,
LLM_NORM, cb, -1);
cb(cur, "result_norm", -1);
// lm_head
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_llama() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
@ -6662,6 +6614,245 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_orion() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, 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, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
cb(inp_pos, "inp_pos", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
cb(KQ_mask, "KQ_mask", -1);
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
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, model.layers[il].attn_norm_b,
LLM_NORM, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
struct ggml_tensor * Qcur = ggml_mul_mat(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 = ggml_mul_mat(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 = ggml_mul_mat(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_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
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, model.layers[il].ffn_norm_b,
LLM_NORM, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL,
model.layers[il].ffn_gate, NULL,
model.layers[il].ffn_down, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm, model.output_norm_b,
LLM_NORM, cb, -1);
cb(cur, "result_norm", -1);
// lm_head
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
struct ggml_cgraph * build_internlm2() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, 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, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0);
cb(inp_pos, "inp_pos", -1);
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0);
cb(KQ_mask, "KQ_mask", -1);
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
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 = ggml_mul_mat(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 = ggml_mul_mat(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 = ggml_mul_mat(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_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
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, cur,
model.layers[il].ffn_up, NULL,
model.layers[il].ffn_gate, NULL,
model.layers[il].ffn_down, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_inp);
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 = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
};
static struct ggml_cgraph * llama_build_graph(
@ -6820,6 +7011,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_orion();
} break;
case LLM_ARCH_INTERNLM2:
{
result = llm.build_internlm2();
} break;
default:
GGML_ASSERT(false);
}
@ -7986,7 +8181,9 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
//
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
if (&fragment == &fragment_buffer.front()) {
raw_text = " " + raw_text; // prefix with space if the first token is not special
if (vocab.add_space_prefix) {
raw_text = " " + raw_text; // prefix with space if the first token is not special
}
}
#ifdef PRETOKENIZERDEBUG
@ -10413,18 +10610,37 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
return result;
}
int32_t llama_max_devices(void) {
return LLAMA_MAX_DEVICES;
size_t llama_max_devices(void) {
return 16;
}
bool llama_mmap_supported(void) {
bool llama_supports_mmap(void) {
return llama_mmap::SUPPORTED;
}
bool llama_mlock_supported(void) {
bool llama_supports_mlock(void) {
return llama_mlock::SUPPORTED;
}
bool llama_supports_gpu_offload(void) {
#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
// Defined when llama.cpp is compiled with support for offloading model layers to GPU.
return true;
#else
return false;
#endif
}
// deprecated:
bool llama_mmap_supported(void) {
return llama_supports_mmap();
}
bool llama_mlock_supported(void) {
return llama_supports_mlock();
}
void llama_backend_init(bool numa) {
ggml_time_init();
@ -10456,8 +10672,8 @@ int64_t llama_time_us(void) {
}
struct llama_model * llama_load_model_from_file(
const char * path_model,
struct llama_model_params params) {
const char * path_model,
struct llama_model_params params) {
ggml_time_init();
llama_model * model = new llama_model;
@ -10833,7 +11049,7 @@ int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int3
int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) {
return snprintf(buf, buf_size, "%s %s %s",
llama_model_arch_name(model->arch).c_str(),
llama_model_arch_name(model->arch),
llama_model_type_name(model->type),
llama_model_ftype_name(model->ftype).c_str());
}
@ -11485,22 +11701,24 @@ struct llama_batch llama_batch_get_one(
};
}
struct llama_batch llama_batch_init(int32_t n_tokens, int32_t embd, int32_t n_seq_max) {
struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
llama_batch batch = { 0, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, };
if (embd) {
batch.embd = (float *) malloc(sizeof(float) * n_tokens * embd);
batch.embd = (float *) malloc(sizeof(float) * n_tokens_alloc * embd);
} else {
batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens);
batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
}
batch.pos = (llama_pos *) malloc(sizeof(llama_pos) * n_tokens);
batch.n_seq_id = (int32_t *) malloc(sizeof(int32_t) * n_tokens);
batch.seq_id = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * n_tokens);
for (int i = 0; i < n_tokens; ++i) {
batch.pos = (llama_pos *) malloc(sizeof(llama_pos) * n_tokens_alloc);
batch.n_seq_id = (int32_t *) malloc(sizeof(int32_t) * n_tokens_alloc);
batch.seq_id = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * (n_tokens_alloc + 1));
for (int i = 0; i < n_tokens_alloc; ++i) {
batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
}
batch.logits = (int8_t *) malloc(sizeof(int8_t) * n_tokens);
batch.seq_id[n_tokens_alloc] = nullptr;
batch.logits = (int8_t *) malloc(sizeof(int8_t) * n_tokens_alloc);
return batch;
}
@ -11511,7 +11729,7 @@ void llama_batch_free(struct llama_batch batch) {
if (batch.pos) free(batch.pos);
if (batch.n_seq_id) free(batch.n_seq_id);
if (batch.seq_id) {
for (int i = 0; i < batch.n_tokens; ++i) {
for (int i = 0; batch.seq_id[i] != nullptr; ++i) {
free(batch.seq_id[i]);
}
free(batch.seq_id);