From 6633d19df8d1f4bcf123fb5ff36f4e67093e4416 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Wed, 5 Aug 2026 07:47:42 +0300 Subject: [PATCH] Fix Antirez DS4 GGUFs (#2256) * Fix Antirez DS4 GGUFs * Let's tell the user what we did --- src/llama-hparams.cpp | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index 2af8704d1..365747ea1 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -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(1, hparams.nextn_predict_layers); const uint32_t dsv4_probe_layer = hparams.n_layer > dsv4_probe_offset