pingu_98/llama_cpp_for_radxa_dragon_wing_q6a
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

model : refactor bias tensor variable names (#22079)

Browse source 
* refactor bias tensor variable names

* use create_tensor_qkv for jina-bert-v2
This commit is contained in:
Sigbjørn Skjæret 2026-04-18 20:12:00 +02:00 committed by GitHub
parent 23b8cc4991
commit 4f02d47339
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
55 changed files with 104 additions and 117 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
src/llama-graph.cpp
View file

@ -1077,9 +1077,9 @@ llm_graph_qkv llm_graph_context::build_qkv(
// fused QKV path
ggml_tensor * qkv = build_lora_mm(layer.wqkv, cur, layer.wqkv_s);
cb(qkv, "wqkv", il);
if (layer.bqkv) {
qkv = ggml_add(ctx0, qkv, layer.bqkv);
cb(qkv, "bqkv", il);
if (layer.wqkv_b) {
qkv = ggml_add(ctx0, qkv, layer.wqkv_b);
cb(qkv, "wqkv_b", il);
}
if (hparams.f_clamp_kqv > 0.0f) {
qkv = ggml_clamp(ctx0, qkv, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
@ -1097,8 +1097,8 @@ llm_graph_qkv llm_graph_context::build_qkv(
// separate Q/K/V path
Qcur = build_lora_mm(layer.wq, cur, layer.wq_s);
cb(Qcur, "Qcur", il);
if (layer.bq) {
Qcur = ggml_add(ctx0, Qcur, layer.bq);
if (layer.wq_b) {
Qcur = ggml_add(ctx0, Qcur, layer.wq_b);
cb(Qcur, "Qcur", il);
}
if (hparams.f_clamp_kqv > 0.0f) {
@ -1107,8 +1107,8 @@ llm_graph_qkv llm_graph_context::build_qkv(
}
Kcur = build_lora_mm(layer.wk, cur, layer.wk_s);
cb(Kcur, "Kcur", il);
if (layer.bk) {
Kcur = ggml_add(ctx0, Kcur, layer.bk);
if (layer.wk_b) {
Kcur = ggml_add(ctx0, Kcur, layer.wk_b);
cb(Kcur, "Kcur", il);
}
if (hparams.f_clamp_kqv > 0.0f) {
@ -1117,8 +1117,8 @@ llm_graph_qkv llm_graph_context::build_qkv(
}
Vcur = build_lora_mm(layer.wv, cur, layer.wv_s);
cb(Vcur, "Vcur", il);
if (layer.bv) {
Vcur = ggml_add(ctx0, Vcur, layer.bv);
if (layer.wv_b) {
Vcur = ggml_add(ctx0, Vcur, layer.wv_b);
cb(Vcur, "Vcur", il);
}
if (hparams.f_clamp_kqv > 0.0f) {

88
src/llama-model.cpp
View file

@ -3098,14 +3098,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const int64_t n_embd_qkv = n_embd_q_ + n_embd_k_ + n_embd_v_;
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
if (layer.wqkv) {
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", bid), {n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", bid), {n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
} else {
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", bid), {n_embd_, n_embd_q_}, flags);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", bid), {n_embd_, n_embd_k_}, flags);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", bid), {n_embd_, n_embd_v_}, flags);
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", bid), {n_embd_q_}, TENSOR_NOT_REQUIRED);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", bid), {n_embd_k_}, TENSOR_NOT_REQUIRED);
layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", bid), {n_embd_v_}, TENSOR_NOT_REQUIRED);
layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", bid), {n_embd_q_}, TENSOR_NOT_REQUIRED);
layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", bid), {n_embd_k_}, TENSOR_NOT_REQUIRED);
layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", bid), {n_embd_v_}, TENSOR_NOT_REQUIRED);
}
};
@ -3138,7 +3138,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
// optional bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@ -3201,7 +3201,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
// No bias for QKV projections as per config: include_bias=false, include_qkv_bias=false
layer.wo =
create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
@ -3336,9 +3336,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
}
// optional bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
if (n_ff > 0) {
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@ -3558,10 +3557,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
layer.wqkv_b = 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.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -3602,8 +3601,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
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);
@ -3719,23 +3718,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i]; // JinaBertLayer
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); //output_dens
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); //output_dens
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); //output_dens
layer.attn_out_norm = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0); //output_norm
layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}, 0);
@ -3783,10 +3775,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
layer.wqkv_b = 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.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -3819,10 +3811,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
@ -3889,7 +3881,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd*3}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd*3}, 0);
layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd*3}, 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);
@ -4068,7 +4060,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0);
@ -4127,7 +4119,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), { n_embd }, 0);
@ -4291,10 +4283,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
layer.wqkv_b = 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.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -4329,7 +4321,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -4646,7 +4638,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
// optional bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -4890,7 +4882,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
}
// feed forward (w/ optional biases)
@ -5152,10 +5144,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
layer.wqkv_b = 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.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -5570,10 +5562,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
layer.wqkv_b = 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.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -5612,10 +5604,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
// attention biases - all have shape n_embd (output dimension of projections)
layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);
layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);
layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -5918,7 +5910,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
// optional bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
@ -5987,7 +5979,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
} else {
if (n_expert != 0) {
const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
@ -6808,7 +6800,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
// optional bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
@ -6890,7 +6882,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
// attention layers (with optional bias)
create_tensor_qkv(layer, i, hidden_size, n_embd_head_k * attn_num_attention_head, attn_num_key_value_head * n_embd_head_k, attn_num_key_value_head * n_embd_head_v, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * attn_num_attention_head, hidden_size}, 0);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED);
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {hidden_size}, 0);
@ -7026,7 +7018,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
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);
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_gate_inp_b = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "bias", i), {n_expert}, 0);
layer.ffn_gate_exps_b = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert}, 0);
@ -7191,7 +7183,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
// optional bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
@ -7422,7 +7414,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
// bias tensors
layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);

9
src/llama-model.h
View file

@ -246,6 +246,8 @@ struct llama_layer {
struct ggml_tensor * wkv_b = nullptr;
struct ggml_tensor * wk_b = nullptr;
struct ggml_tensor * wv_b = nullptr;
struct ggml_tensor * wqkv_b = nullptr;
struct ggml_tensor * wo_b = nullptr;
struct ggml_tensor * wq_cross = nullptr;
struct ggml_tensor * wk_cross = nullptr;
struct ggml_tensor * wv_cross = nullptr;
@ -256,13 +258,6 @@ struct llama_layer {
struct ggml_tensor * wo_enc = nullptr;
struct ggml_tensor * wqkv_gate = nullptr;
// attention bias
struct ggml_tensor * bq = nullptr;
struct ggml_tensor * bk = nullptr;
struct ggml_tensor * bv = nullptr;
struct ggml_tensor * bo = nullptr;
struct ggml_tensor * bqkv = nullptr;
// relative position bias
struct ggml_tensor * attn_rel_b = nullptr;
struct ggml_tensor * attn_rel_b_enc = nullptr;

2
src/models/apertus.cpp
View file

@ -50,7 +50,7 @@ llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur_pos", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/arcee.cpp
View file

@ -55,7 +55,7 @@ llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/bailingmoe.cpp
View file

@ -48,7 +48,7 @@ llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
}

2
src/models/bailingmoe2.cpp
View file

@ -48,7 +48,7 @@ llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const ll
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}

2
src/models/bert.cpp
View file

@ -72,7 +72,7 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
cb(cur, "kqv_out", il);
}

4
src/models/bitnet.cpp
View file

@ -57,8 +57,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
cb(cur, "attn_sub_norm", il);
cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
if (model.layers[il].bo) {
cur = ggml_add(ctx0, cur, model.layers[il].bo);
if (model.layers[il].wo_b) {
cur = ggml_add(ctx0, cur, model.layers[il].wo_b);
}
cb(cur, "attn_out", il);
}

2
src/models/bloom.cpp
View file

@ -33,7 +33,7 @@ llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_para
n_embd_head, n_head, n_head_kv, il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/codeshell.cpp
View file

@ -47,7 +47,7 @@ llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/cohere2-iswa.cpp
View file

@ -58,7 +58,7 @@ llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/command-r.cpp
View file

@ -54,7 +54,7 @@ llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/deci.cpp
View file

@ -59,7 +59,7 @@ llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/deepseek.cpp
View file

@ -49,7 +49,7 @@ llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/dots1.cpp
View file

@ -49,7 +49,7 @@ llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/dream.cpp
View file

@ -43,7 +43,7 @@ llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/exaone.cpp
View file

@ -46,7 +46,7 @@ llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_pa
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/gpt2.cpp
View file

@ -37,7 +37,7 @@ llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params
n_embd_head, n_head, n_head_kv, il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/gptneox.cpp
View file

@ -46,7 +46,7 @@ llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/granite-hybrid.cpp
View file

@ -92,7 +92,7 @@ ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor *
const float kq_scale =
hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
return cur;

2
src/models/granite.cpp
View file

@ -101,7 +101,7 @@ ggml_tensor * llm_build_granite::build_attention_layer(
const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
return cur;

2
src/models/grok.cpp
View file

@ -50,7 +50,7 @@ llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/grovemoe.cpp
View file

@ -50,7 +50,7 @@ llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}

2
src/models/hunyuan-dense.cpp
View file

@ -64,7 +64,7 @@ llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, cons
cb(Qcur, "Qcur_norm", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/hunyuan-moe.cpp
View file

@ -65,7 +65,7 @@ llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const ll
cb(Qcur, "Qcur_norm", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/internlm2.cpp
View file

@ -50,7 +50,7 @@ llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/jais.cpp
View file

@ -27,7 +27,7 @@ llm_build_jais::llm_build_jais(const llama_model & model, const llm_graph_params
n_embd_head, n_head, n_head_kv, il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/jais2.cpp
View file

@ -51,7 +51,7 @@ llm_build_jais2::llm_build_jais2(const llama_model & model, const llm_graph_para
cb(Kcur, "Kcur_rope", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/llama.cpp
View file

@ -70,7 +70,7 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
cb(Kcur, "Kcur_normed", il);
}
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
if (model.layers[il].wo_s) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);

2
src/models/llama4.cpp
View file

@ -84,7 +84,7 @@ llm_build_llama4<iswa>::llm_build_llama4(const llama_model & model, const llm_gr
cb(Kcur, "Kcur_normed", il);
}
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/maincoder.cpp
View file

@ -56,7 +56,7 @@ llm_build_maincoder::llm_build_maincoder(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/mistral3.cpp
View file

@ -67,7 +67,7 @@ llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_grap
}
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/mpt.cpp
View file

@ -56,7 +56,7 @@ llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params &
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}

2
src/models/nemotron-h.cpp
View file

@ -70,7 +70,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
const float kq_scale =
hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
return cur;

2
src/models/nemotron.cpp
View file

@ -51,7 +51,7 @@ llm_build_nemotron::llm_build_nemotron(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/openai-moe-iswa.cpp
View file

@ -48,7 +48,7 @@ llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model,
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il);
cb(cur, "attn_out", il);

2
src/models/paddleocr.cpp
View file

@ -55,7 +55,7 @@ llm_build_paddleocr::llm_build_paddleocr(const llama_model & model, const llm_gr
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1) {

2
src/models/pangu-embedded.cpp
View file

@ -49,7 +49,7 @@ llm_build_pangu_embedded::llm_build_pangu_embedded(const llama_model & model, co
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/phi2.cpp
View file

@ -51,7 +51,7 @@ llm_build_phi2::llm_build_phi2(const llama_model & model, const llm_graph_params
Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/phi3.cpp
View file

@ -60,7 +60,7 @@ llm_build_phi3<iswa>::llm_build_phi3(const llama_model & model, const llm_graph_
cb(Qcur, "Qcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/qwen2.cpp
View file

@ -50,7 +50,7 @@ llm_build_qwen2::llm_build_qwen2(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/qwen2moe.cpp
View file

@ -50,7 +50,7 @@ llm_build_qwen2moe::llm_build_qwen2moe(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/qwen2vl.cpp
View file

@ -53,7 +53,7 @@ llm_build_qwen2vl::llm_build_qwen2vl(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/qwen3.cpp
View file

@ -56,7 +56,7 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
if (model.layers[il].wo_s) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);

2
src/models/qwen3moe.cpp
View file

@ -56,7 +56,7 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
if (model.layers[il].wo_s) {
cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);

2
src/models/qwen3vl-moe.cpp
View file

@ -62,7 +62,7 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/qwen3vl.cpp
View file

@ -62,7 +62,7 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}

2
src/models/rnd1.cpp
View file

@ -58,7 +58,7 @@ llm_build_rnd1::llm_build_rnd1(const llama_model & model, const llm_graph_params
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/seed-oss.cpp
View file

@ -52,7 +52,7 @@ llm_build_seed_oss::llm_build_seed_oss(const llama_model & model, const llm_grap
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/smallthinker.cpp
View file

@ -59,7 +59,7 @@ llm_build_smallthinker<iswa>::llm_build_smallthinker(const llama_model & model,
cb(Kcur, "Kcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/smollm3.cpp
View file

@ -55,7 +55,7 @@ llm_build_smollm3::llm_build_smollm3(const llama_model & model, const llm_graph_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
cb(cur, "attn_out", il);
}

2
src/models/starcoder.cpp
View file

@ -36,7 +36,7 @@ llm_build_starcoder::llm_build_starcoder(const llama_model & model, const llm_gr
n_embd_head, n_head, n_head_kv, il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/starcoder2.cpp
View file

@ -50,7 +50,7 @@ llm_build_starcoder2::llm_build_starcoder2(const llama_model & model, const llm_
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {

2
src/models/t5.cpp
View file

@ -41,7 +41,7 @@ llm_build_t5<false>::llm_build_t5(const llama_model & model, const llm_graph_par
ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);
cur = build_attn(inp_attn_self,
model.layers[il].wo, model.layers[il].bo, model.layers[il].wo_s,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
cb(cur, "kqv_out", il);
}