diff --git a/src/llama.cpp b/src/llama.cpp index 077065869ed11..709f7a3a2d1ee 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -8889,7 +8889,7 @@ static struct ggml_tensor * llm_build_kqv( struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); cb(kq, "kq", il); - if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 || model.arch == LLM_ARCH_NEMOTRON) { + if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2 || model.arch == LLM_ARCH_NEMOTRON || model.arch == LLM_ARCH_CHATGLM) { // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847 ggml_mul_mat_set_prec(kq, GGML_PREC_F32);