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

arthw · Nov 15, 2024 · 5ff6c1e · 5ff6c1e
1 parent 62f8b46
commit 5ff6c1e
Showing 1 changed file with 1 addition and 1 deletion.
diff --git 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);