From: piDack <redacted>
Date: Fri, 23 Aug 2024 07:27:17 +0000 (+0800)
Subject: llama : use F32 precision in GLM4 attention and no FA (#9130)
X-Git-Tag: upstream/0.0.4488~870
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=a07c32ea54850c989f0ef6989da5b955b77b7172;p=pkg%2Fggml%2Fsources%2Fllama.cpp

llama : use F32 precision in GLM4 attention and no FA (#9130)
---

diff --git a/src/llama.cpp b/src/llama.cpp
index bd7f1508..869b584a 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -8885,7 +8885,7 @@ 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);