From 220860aa0c15f665a8225424064f6951e3696037 Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Thu, 14 Aug 2025 16:08:31 +0300
Subject: [PATCH] graph : use F32 accumulators for gpt-oss

ggml-ci
---
 src/llama-graph.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 053c72d6d..b0b1c7323 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -1566,6 +1566,11 @@ ggml_tensor * llm_graph_context::build_attn_with_sinks(
 
     if (wo) {
         cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_OPENAI_MOE) {
+            // similar the original build_attn
+            // TODO: this is tmp until we refactor and remove the build_attn_with_sinks() path
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
     }
 
     if (wo_b) {