llama-cpp: add gemma 4 graph fix

patches/llamacpp/0004-gemma4-graph-fix.patch

@@ -0,0 +1,24 @@
+From b934a8ca49f9e764fa21d45ff2ce1168a3a7c914 Mon Sep 17 00:00:00 2001
+From: Georgi Gerganov <ggerganov@gmail.com>
+Date: Mon, 6 Apr 2026 11:50:22 +0300
+Subject: [PATCH] models : set gemma 4 FFN MoE prec to F32
+
+---
+ src/llama-graph.cpp | 4 ++--
+ 1 file changed, 2 insertions(+), 2 deletions(-)
+
+diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
+index 0e7d96ca10d..aa8a35721fa 100644
+--- a/src/llama-graph.cpp
++++ b/src/llama-graph.cpp
+@@ -1185,8 +1185,8 @@ ggml_tensor * llm_graph_context::build_ffn(
+ 
+     if (down) {
+         cur = build_lora_mm(down, cur);
+-        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
+-            // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
++        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2 || arch == LLM_ARCH_GEMMA4) {
++            // certain models seem to have numerical issues with half-precision accumulators
+             ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+         }
+     }

services/llama-cpp.nix

@@ -27,6 +27,7 @@ in
       inputs.llamacpp.packages.${pkgs.system}.vulkan.overrideAttrs (old: {
         patches = (old.patches or [ ]) ++ [
           ../patches/llamacpp/0003-gemma4-tokenizer-fix.patch
+          ../patches/llamacpp/0004-gemma4-graph-fix.patch
         ];
       })
     );