Commit 5e03a1ff4 for llama.cpp

commit 5e03a1ff450cae37a774684de55d14fb48b1f90b
Author: Mikhail Podvitskii <podvitskiymichael@gmail.com>
Date:   Fri Oct 16 05:32:26 2026 +0200

    graph : add precision handling for MELLUM architecture (#30584)

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 80be4bb57..454c21404 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -2452,7 +2452,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     }

     experts = build_lora_mm_id(down_exps, cur, selected_experts, down_exps_s, slots); // [n_embd, n_expert_used, n_tokens]
-    if (arch == LLM_ARCH_MISTRAL4) {
+    if (arch == LLM_ARCH_MISTRAL4 || arch == LLM_ARCH_MELLUM) {
         // src1 can exceed F16 range
         ggml_prec_set_src(experts, GGML_PREC_F32, 1);
     }