Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

model : load hparams.n_layer_nextn before n_layer() calls (#28159)

Browse Source 
* load hparams.n_layer_nextn before n_layer() calls

* remove duplicate loads
This commit is contained in:
Sigbjørn Skjæret
2026-09-01 13:55:45 +02:00
committed by GitHub
parent fe2120bc9d
commit 9d817213a0
21 changed files with 7 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/deepseek4.cpp
+1 -3
View File
@@ -17,15 +17,13 @@ static float dsv4_rope_attn_factor(float freq_scale, float ext_factor) {
}
void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
if (hparams.n_layer_nextn > 0 && hparams.n_layer_nextn < hparams.n_layer_all) {
if (hparams.n_layer_nextn > 0) {
const uint32_t n_layer_main = hparams.n_layer_all - hparams.n_layer_nextn;
const std::string mtp_probe = "blk." + std::to_string(n_layer_main) + ".nextn.eh_proj.weight";
if (ml.get_weight(mtp_probe.c_str()) == nullptr) {
hparams.n_layer_nextn = 0;
}
}
GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < block_count");
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);