LLM-Inference/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-08-12 22:29:39 +04:00
Code Issues Packages Projects Releases Wiki Activity

Fix Antirez DS4 GGUFs (#2256)

Browse Source 
* Fix Antirez DS4 GGUFs

* Let's tell the user what we did
This commit is contained in:
Kawrakow
2026-08-05 07:47:42 +03:00
committed by GitHub
parent 6b55d2c750
commit 6633d19df8
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-hparams.cpp
+8
View File
@@ -1664,6 +1664,14 @@ void llm_load_hparams(
case LLM_ARCH_GLM_DSA:
{
ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
if (model.arch == LLM_ARCH_DEEPSEEK4 && hparams.n_layer == 43 && hparams.nextn_predict_layers > 0) {
LLAMA_LOG_WARN("===============================================================================================\n");
LLAMA_LOG_WARN("Unexpected number of layers (%d) and nextn_predict_layers (%d) for DeepSeek4-Flash\n",
hparams.n_layer, hparams.nextn_predict_layers);
LLAMA_LOG_WARN(" -> setting nextn_predict_layers to zero\n");
LLAMA_LOG_WARN("===============================================================================================\n");
hparams.nextn_predict_layers = 0;
}
// Probe the first appended predictor block, or n_layer - 1 for base GGUFs.
const uint32_t dsv4_probe_offset = std::max<uint32_t>(1, hparams.nextn_predict_layers);
const uint32_t dsv4_probe_layer = hparams.n_layer > dsv4_probe_offset