vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : add support for Nomic Embed (#5468)

Browse source
This commit is contained in:
Jared Van Bortel 2024-02-13 12:03:53 -05:00 • committed by GitHub
parent c4e6dd59e4
commit ea9c8e1143
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 273 additions and 113 deletions
Download patch file Download diff file Expand all files Collapse all files

201
llama.cpp
View file

@ -197,6 +197,7 @@ enum llm_arch {
LLM_ARCH_PERSIMMON,
LLM_ARCH_REFACT,
LLM_ARCH_BERT,
LLM_ARCH_NOMIC_BERT,
LLM_ARCH_BLOOM,
LLM_ARCH_STABLELM,
LLM_ARCH_QWEN,
@ -211,27 +212,28 @@ enum llm_arch {
};
static std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_LLAMA, "llama" },
{ LLM_ARCH_FALCON, "falcon" },
{ LLM_ARCH_GPT2, "gpt2" },
{ LLM_ARCH_GPTJ, "gptj" },
{ LLM_ARCH_GPTNEOX, "gptneox" },
{ LLM_ARCH_MPT, "mpt" },
{ LLM_ARCH_BAICHUAN, "baichuan" },
{ LLM_ARCH_STARCODER, "starcoder" },
{ LLM_ARCH_PERSIMMON, "persimmon" },
{ LLM_ARCH_REFACT, "refact" },
{ LLM_ARCH_BERT, "bert" },
{ LLM_ARCH_BLOOM, "bloom" },
{ LLM_ARCH_STABLELM, "stablelm" },
{ LLM_ARCH_QWEN, "qwen" },
{ LLM_ARCH_QWEN2, "qwen2" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PLAMO, "plamo" },
{ LLM_ARCH_CODESHELL, "codeshell" },
{ LLM_ARCH_ORION, "orion" },
{ LLM_ARCH_INTERNLM2, "internlm2" },
{ LLM_ARCH_MINICPM, "minicpm" },
{ LLM_ARCH_LLAMA, "llama" },
{ LLM_ARCH_FALCON, "falcon" },
{ LLM_ARCH_GPT2, "gpt2" },
{ LLM_ARCH_GPTJ, "gptj" },
{ LLM_ARCH_GPTNEOX, "gptneox" },
{ LLM_ARCH_MPT, "mpt" },
{ LLM_ARCH_BAICHUAN, "baichuan" },
{ LLM_ARCH_STARCODER, "starcoder" },
{ LLM_ARCH_PERSIMMON, "persimmon" },
{ LLM_ARCH_REFACT, "refact" },
{ LLM_ARCH_BERT, "bert" },
{ LLM_ARCH_NOMIC_BERT, "nomic-bert" },
{ LLM_ARCH_BLOOM, "bloom" },
{ LLM_ARCH_STABLELM, "stablelm" },
{ LLM_ARCH_QWEN, "qwen" },
{ LLM_ARCH_QWEN2, "qwen2" },
{ LLM_ARCH_PHI2, "phi2" },
{ LLM_ARCH_PLAMO, "plamo" },
{ LLM_ARCH_CODESHELL, "codeshell" },
{ LLM_ARCH_ORION, "orion" },
{ LLM_ARCH_INTERNLM2, "internlm2" },
{ LLM_ARCH_MINICPM, "minicpm" },
};
enum llm_kv {
@ -375,6 +377,7 @@ enum llm_tensor {
LLM_TENSOR_ATTN_OUT,
LLM_TENSOR_ATTN_NORM,
LLM_TENSOR_ATTN_NORM_2,
LLM_TENSOR_ATTN_OUT_NORM,
LLM_TENSOR_ATTN_ROT_EMBD,
LLM_TENSOR_FFN_GATE_INP,
LLM_TENSOR_FFN_NORM,
@ -387,6 +390,7 @@ enum llm_tensor {
LLM_TENSOR_FFN_UP_EXP,
LLM_TENSOR_ATTN_Q_NORM,
LLM_TENSOR_ATTN_K_NORM,
LLM_TENSOR_LAYER_OUT_NORM,
};
static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES = {
@ -552,12 +556,27 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_TOKEN_TYPES, "token_types" },
{ LLM_TENSOR_POS_EMBD, "position_embd" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_output_norm" },
{ LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_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.layer_output_norm" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
LLM_ARCH_NOMIC_BERT,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
{ LLM_TENSOR_TOKEN_TYPES, "token_types" },
{ LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_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" },
},
@ -1485,6 +1504,7 @@ enum e_model {
MODEL_22M,
MODEL_33M,
MODEL_109M,
MODEL_137M,
MODEL_335M,
MODEL_0_5B,
MODEL_1B,
@ -1620,6 +1640,8 @@ struct llama_layer {
struct ggml_tensor * attn_q_norm_b;
struct ggml_tensor * attn_k_norm;
struct ggml_tensor * attn_k_norm_b;
struct ggml_tensor * attn_out_norm;
struct ggml_tensor * attn_out_norm_b;
// attention
struct ggml_tensor * wq;
@ -1638,6 +1660,8 @@ struct llama_layer {
// normalization
struct ggml_tensor * ffn_norm;
struct ggml_tensor * ffn_norm_b;
struct ggml_tensor * layer_out_norm;
struct ggml_tensor * layer_out_norm_b;
// ff
struct ggml_tensor * ffn_gate; // w1
@ -2855,6 +2879,11 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
static const char * llama_model_type_name(e_model type) {
switch (type) {
case MODEL_22M: return "22M";
case MODEL_33M: return "33M";
case MODEL_109M: return "109M";
case MODEL_137M: return "137M";
case MODEL_0_5B: return "0.5B";
case MODEL_1B: return "1B";
case MODEL_2B: return "2B";
case MODEL_3B: return "3B";
@ -3073,6 +3102,17 @@ static void llm_load_hparams(
model.type = e_model::MODEL_335M; break; // bge-large
}
} break;
case LLM_ARCH_NOMIC_BERT:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn);
ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type);
ml.get_key(LLM_KV_POOLING_LAYER, hparams.pooling_layer);
if (hparams.n_layer == 12 && hparams.n_embd == 768) {
model.type = e_model::MODEL_137M;
}
} break;
case LLM_ARCH_BLOOM:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@ -3875,10 +3915,14 @@ static bool llm_load_tensors(
}
} break;
case LLM_ARCH_BERT:
case LLM_ARCH_NOMIC_BERT:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type});
model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train});
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
model.type_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type});
if (model.arch == LLM_ARCH_BERT) {
model.pos_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train});
}
model.tok_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd});
model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd});
@ -3888,29 +3932,38 @@ static bool llm_load_tensors(
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd});
if (model.arch == LLM_ARCH_BERT) {
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd});
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd});
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa});
layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd});
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa});
} else {
layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
}
layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa});
layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd});
layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "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});
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
if (model.arch == LLM_ARCH_BERT) {
layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd});
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});
layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});
} else {
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
}
layer.layer_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd});
layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i), {n_embd});
}
} break;
case LLM_ARCH_BLOOM:
@ -5773,6 +5826,7 @@ struct llm_build_context {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
struct ggml_tensor * cur;
@ -5789,7 +5843,9 @@ struct llm_build_context {
// token types are hardcoded to zero ("Sentence A")
struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
inpL = ggml_add(ctx0, inpL, type_row0);
inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
if (model.arch == LLM_ARCH_BERT) {
inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
}
cb(inpL, "inp_embd", -1);
// embed layer norm
@ -5805,7 +5861,7 @@ struct llm_build_context {
struct ggml_tensor * cur = inpL;
// self-attention
{
if (model.arch == LLM_ARCH_BERT) {
struct ggml_tensor * Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq);
cb(Qcur, "Qcur", il);
@ -5818,6 +5874,37 @@ struct llm_build_context {
// seems like we just need to do this for Q?
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
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);
} else {
// compute Q and K and RoPE them
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
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, 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);
@ -5828,25 +5915,34 @@ struct llm_build_context {
cur = ggml_add(ctx0, cur, inpL);
// attention layer norm
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, cb, il);
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il);
struct ggml_tensor * ffn_inp = cur;
cb(ffn_inp, "ffn_inp", il);
// feed-forward network
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b,
NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b,
NULL,
LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
if (model.arch == LLM_ARCH_BERT) {
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, model.layers[il].ffn_up_b,
NULL, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b,
NULL,
LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
} else {
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);
// attentions bypass the intermediate layer
cur = ggml_add(ctx0, cur, ffn_inp);
// output layer norm
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, cb, il);
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, cb, il);
// input for next layer
inpL = cur;
@ -7289,6 +7385,7 @@ static struct ggml_cgraph * llama_build_graph(
result = llm.build_refact();
} break;
case LLM_ARCH_BERT:
case LLM_ARCH_NOMIC_BERT:
{
result = llm.build_bert();
} break;