nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 8 Packages Projects Releases Wiki Activity

model: Add EXAONE 4.5 implementations (#21733)

Browse Source 
* Add EXAONE 4.5 and Add GQA for MMproj

* mtmd: EXAONE 4.5 vision markers and projector path

EXAONE 4.5 uses <vision> and </vision> for image boundaries; Qwen keeps
<|vision_start|> and <|vision_end|>.

Route EXAONE 4.5 through the Qwen2.5-VL-style encode path (window attention
pattern, optional mmproj input norm). Update exaone4_5 projector weights and
convert_hf_to_gguf for mmproj export.

* mtmd: load EXAONE4 nextn tensors correctly

Align EXAONE4 tensor registration with EXAONE_MOE for NextN/MTP slots and avoid skip-flag propagation on duplicated rope_freqs so model loading succeeds for EXAONE 4.5 GGUF.

* Minor fixes

* Address PR feedback

* Address PR feedback

* Fix EXAONE after merge

* Fix EXAONE 4.5 conversion

* Address PR feedback

* Refactor EXAONE 4.5 conversion

* Address PR feedback

* Fix unintended deletion

* Minor fix

---------

Co-authored-by: LG-AI-EXAONE <exaonemodels@lgresearch.ai>
This commit is contained in:
Junwon Hwang
2026-06-01 11:48:53 +02:00
committed by GitHub
co-authored by LG-AI-EXAONE
parent 19620004f5
commit 48b88c3b00
11 changed files with 356 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/exaone4.cpp
+35 -12
View File
@@ -15,6 +15,9 @@ void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
switch (hparams.n_layer) {
case 30: type = LLM_TYPE_1_2B; break;
@@ -38,21 +41,37 @@ void llama_model_exaone4::load_arch_tensors(llama_model_loader &) {
}
for (int i = 0; i < n_layer; ++i) {
const bool is_nextn = hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers;
int flags = 0;
if (is_nextn) {
// NextN/MTP layers are preserved in GGUF but are not executed yet.
flags |= TENSOR_SKIP;
}
auto & layer = layers[i];
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, flags);
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
if (!is_nextn) {
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
}
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags);
layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);
if (is_nextn) {
layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);
layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), {n_embd}, flags);
layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), {n_embd}, flags);
layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
}
}
}
@@ -90,7 +109,11 @@ llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_gra
}
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
// MTP / NextN tail blocks are loaded for compatibility but not executed (same as exaone-moe).
const int n_layer_main = int(n_layer) - int(hparams.nextn_predict_layers);
GGML_ASSERT(n_layer_main > 0);
for (int il = 0; il < n_layer_main; ++il) {
ggml_tensor * inpSA = inpL;
// use RoPE for SWA layers or non-SWA models
@@ -126,7 +149,7 @@ llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_gra
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
cb(cur, "attn_out", il);
}
if (il == n_layer - 1 && inp_out_ids) {
if (il == n_layer_main - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}