model : Nomic Embed Text V2 with Mixture-of-Experts (MoE) architecture (#12466)
* Nomic Embed Text V2 with Mixture-of-Experts (MoE) architecture - Adds MoE-based embedding model supporting multilingual embeddings. - Selects architecture variant based on hyperparameter detection (MoE layers). - Removes unnecessary subclass initialization checks for clarity. https://www.nomic.ai/blog/posts/nomic-embed-text-v2 Co-authored-by: Jared Van Bortel <jared@nomic.ai> * fix tokenizer * don't rename this tensor --------- Co-authored-by: Jared Van Bortel <jared@nomic.ai>
This commit is contained in:
AT
GitHub
parent
eaea325324
commit
5f5e39e1ba
9 changed files with 247 additions and 110 deletions
|
|
@ -19,6 +19,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
|
|||
{ LLM_ARCH_REFACT, "refact" },
|
||||
{ LLM_ARCH_BERT, "bert" },
|
||||
{ LLM_ARCH_NOMIC_BERT, "nomic-bert" },
|
||||
{ LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" },
|
||||
{ LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
|
||||
{ LLM_ARCH_BLOOM, "bloom" },
|
||||
{ LLM_ARCH_STABLELM, "stablelm" },
|
||||
|
|
@ -106,6 +107,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
|
|||
{ LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" },
|
||||
{ LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" },
|
||||
{ LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" },
|
||||
{ LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" },
|
||||
{ LLM_KV_POOLING_TYPE, "%s.pooling_type" },
|
||||
{ LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
|
||||
{ LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" },
|
||||
|
|
@ -472,6 +474,24 @@ 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_NOMIC_BERT_MOE,
|
||||
{
|
||||
{ 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" },
|
||||
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
|
||||
{ LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" },
|
||||
{ LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" },
|
||||
},
|
||||
},
|
||||
{
|
||||
LLM_ARCH_JINA_BERT_V2,
|
||||
{
|
||||
|
|
|
|||
|
|
@ -23,6 +23,7 @@ enum llm_arch {
|
|||
LLM_ARCH_REFACT,
|
||||
LLM_ARCH_BERT,
|
||||
LLM_ARCH_NOMIC_BERT,
|
||||
LLM_ARCH_NOMIC_BERT_MOE,
|
||||
LLM_ARCH_JINA_BERT_V2,
|
||||
LLM_ARCH_BLOOM,
|
||||
LLM_ARCH_STABLELM,
|
||||
|
|
@ -110,6 +111,7 @@ enum llm_kv {
|
|||
LLM_KV_EXPERT_WEIGHTS_SCALE,
|
||||
LLM_KV_EXPERT_WEIGHTS_NORM,
|
||||
LLM_KV_EXPERT_GATING_FUNC,
|
||||
LLM_KV_MOE_EVERY_N_LAYERS,
|
||||
LLM_KV_POOLING_TYPE,
|
||||
LLM_KV_LOGIT_SCALE,
|
||||
LLM_KV_DECODER_START_TOKEN_ID,
|
||||
|
|
|
|||
|
|
@ -925,28 +925,35 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
|
|||
ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
|
||||
cb(up, "ffn_moe_up", il);
|
||||
|
||||
ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
|
||||
cb(gate, "ffn_moe_gate", il);
|
||||
ggml_tensor * experts = nullptr;
|
||||
if (gate_exps) {
|
||||
cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
|
||||
cb(cur, "ffn_moe_gate", il);
|
||||
} else {
|
||||
cur = up;
|
||||
}
|
||||
|
||||
switch (type_op) {
|
||||
case LLM_FFN_SILU:
|
||||
{
|
||||
gate = ggml_silu(ctx0, gate);
|
||||
cb(gate, "ffn_moe_silu", il);
|
||||
cur = ggml_silu(ctx0, cur);
|
||||
cb(cur, "ffn_moe_silu", il);
|
||||
} break;
|
||||
case LLM_FFN_GELU:
|
||||
{
|
||||
gate = ggml_gelu(ctx0, gate);
|
||||
cb(gate, "ffn_moe_gelu", il);
|
||||
cur = ggml_gelu(ctx0, cur);
|
||||
cb(cur, "ffn_moe_gelu", il);
|
||||
} break;
|
||||
default:
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
|
||||
ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens]
|
||||
cb(par, "ffn_moe_gate_par", il);
|
||||
if (gate_exps) {
|
||||
cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
|
||||
cb(cur, "ffn_moe_gate_par", il);
|
||||
}
|
||||
|
||||
ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
|
||||
experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
|
||||
cb(experts, "ffn_moe_down", il);
|
||||
|
||||
if (!weight_before_ffn) {
|
||||
|
|
|
|||
|
|
@ -66,6 +66,7 @@ struct llama_hparams {
|
|||
float expert_weights_scale = 0.0;
|
||||
bool expert_weights_norm = false;
|
||||
uint32_t expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
|
||||
uint32_t moe_every_n_layers = 0;
|
||||
|
||||
float f_norm_eps;
|
||||
float f_norm_rms_eps;
|
||||
|
|
|
|||
|
|
@ -695,10 +695,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
|
|||
}
|
||||
} break;
|
||||
case LLM_ARCH_NOMIC_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT_MOE:
|
||||
{
|
||||
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_POOLING_TYPE, hparams.pooling_type);
|
||||
ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS, hparams.moe_every_n_layers, 0);
|
||||
|
||||
if (hparams.n_layer == 12 && hparams.n_embd == 768) {
|
||||
type = LLM_TYPE_137M;
|
||||
|
|
@ -2057,6 +2059,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
|
|||
} break;
|
||||
case LLM_ARCH_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT_MOE:
|
||||
{
|
||||
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
|
||||
type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0);
|
||||
|
|
@ -2090,20 +2093,31 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
|
|||
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
|
||||
}
|
||||
|
||||
if (arch == LLM_ARCH_NOMIC_BERT_MOE) {
|
||||
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
|
||||
}
|
||||
|
||||
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
|
||||
|
||||
layer.attn_out_norm = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
|
||||
layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}, 0);
|
||||
|
||||
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
|
||||
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
|
||||
|
||||
if (arch == LLM_ARCH_BERT) {
|
||||
if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
|
||||
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
|
||||
layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, 0);
|
||||
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
|
||||
layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff, n_expert}, 0);
|
||||
layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0);
|
||||
layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
|
||||
} else {
|
||||
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
|
||||
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
|
||||
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
|
||||
|
||||
if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) {
|
||||
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
|
||||
layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, 0);
|
||||
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
|
||||
} else {
|
||||
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
|
||||
}
|
||||
}
|
||||
|
||||
layer.layer_out_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
|
||||
|
|
@ -5730,6 +5744,11 @@ struct llm_build_bert : public llm_graph_context {
|
|||
cur = build_lora_mm(model.layers[il].wqkv, cur);
|
||||
cb(cur, "wqkv", il);
|
||||
|
||||
if (model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
|
||||
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
|
||||
cb(cur, "bqkv", il);
|
||||
}
|
||||
|
||||
Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
|
||||
Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
|
||||
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)));
|
||||
|
|
@ -5782,13 +5801,29 @@ struct llm_build_bert : public llm_graph_context {
|
|||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
// feed-forward network
|
||||
if (model.arch == LLM_ARCH_BERT) {
|
||||
if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
|
||||
// MoE branch
|
||||
cur = build_moe_ffn(cur,
|
||||
model.layers[il].ffn_gate_inp,
|
||||
model.layers[il].ffn_up_exps,
|
||||
nullptr,
|
||||
model.layers[il].ffn_down_exps,
|
||||
nullptr,
|
||||
hparams.n_expert,
|
||||
hparams.n_expert_used,
|
||||
LLM_FFN_GELU,
|
||||
false, false,
|
||||
0.0f,
|
||||
LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
|
||||
cb(cur, "ffn_moe_out", il);
|
||||
} else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) {
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
|
||||
NULL, NULL, NULL,
|
||||
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
|
||||
NULL,
|
||||
LLM_FFN_GELU, LLM_FFN_SEQ, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, NULL, NULL,
|
||||
|
|
@ -5796,6 +5831,7 @@ struct llm_build_bert : public llm_graph_context {
|
|||
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
|
||||
NULL,
|
||||
LLM_FFN_GELU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
} else {
|
||||
cur = build_ffn(cur,
|
||||
model.layers[il].ffn_up, NULL, NULL,
|
||||
|
|
@ -5803,8 +5839,8 @@ struct llm_build_bert : public llm_graph_context {
|
|||
model.layers[il].ffn_down, NULL, NULL,
|
||||
NULL,
|
||||
LLM_FFN_SILU, LLM_FFN_PAR, il);
|
||||
cb(cur, "ffn_out", il);
|
||||
}
|
||||
cb(cur, "ffn_out", il);
|
||||
|
||||
// attentions bypass the intermediate layer
|
||||
cur = ggml_add(ctx0, cur, ffn_inp);
|
||||
|
|
@ -12843,6 +12879,7 @@ llm_graph_result_ptr llama_model::build_graph(
|
|||
case LLM_ARCH_BERT:
|
||||
case LLM_ARCH_JINA_BERT_V2:
|
||||
case LLM_ARCH_NOMIC_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT_MOE:
|
||||
{
|
||||
llm = std::make_unique<llm_build_bert>(*this, params, gf);
|
||||
} break;
|
||||
|
|
@ -13201,6 +13238,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
|
|||
case LLM_ARCH_DBRX:
|
||||
case LLM_ARCH_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT:
|
||||
case LLM_ARCH_NOMIC_BERT_MOE:
|
||||
case LLM_ARCH_STABLELM:
|
||||
case LLM_ARCH_BITNET:
|
||||
case LLM_ARCH_QWEN:
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue