llama-cpp: add gemma 4 graph fix

flake.lock

@@ -304,11 +304,11 @@
         "rust-overlay": "rust-overlay"
       },
       "locked": {
-        "lastModified": 1775754862,
+        "lastModified": 1775510693,
-        "narHash": "sha256-8y9cz8+cyeA7KtA7+Q3bXjyFJV5nM38Fc0E4qPw7WDk=",
+        "narHash": "sha256-gZfJ07j/oOciDi8mF/V8QTm7YCeDcusNSMZzBFi8OUM=",
         "owner": "nix-community",
         "repo": "lanzaboote",
-        "rev": "bea51aaee00688794a877f308007590a6cc8e378",
+        "rev": "3fe0ae8cb285e0ad101a9675f4190d455fb05e85",
         "type": "github"
       },
       "original": {
@@ -325,11 +325,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1775754125,
+        "lastModified": 1775614184,
-        "narHash": "sha256-4udYhEvii0xPmRiKXYWLhPakPDd1mJppnEFY6uWdv8s=",
+        "narHash": "sha256-OYwr36LLVIeEqccN1mJ2k6vCsFocboCQJnbtne415Ig=",
         "owner": "TheTom",
         "repo": "llama-cpp-turboquant",
-        "rev": "8590cbff961dbaf1d3a9793fd11d402e248869b9",
+        "rev": "eea498c42716519e58baf2d9600d2e2b41839255",
         "type": "github"
       },
       "original": {
@@ -368,11 +368,11 @@
         "systems": "systems_3"
       },
       "locked": {
-        "lastModified": 1775752089,
+        "lastModified": 1775531897,
-        "narHash": "sha256-+psXqZ1SvQw7L8HgCQINmob9zLnvK433b2k080lBPH0=",
+        "narHash": "sha256-3NIpnV1HxBCwi00iMvj9KcqXkM0VNA72KABj8g0cFFs=",
         "owner": "Infinidoge",
         "repo": "nix-minecraft",
-        "rev": "1beacd3bdadabfac884dedd56176966c141214d8",
+        "rev": "8c7693880cb861e60adeab5480f02dc3e7a390f6",
         "type": "github"
       },
       "original": {
@@ -399,11 +399,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1775595990,
+        "lastModified": 1775305101,
-        "narHash": "sha256-OEf7YqhF9IjJFYZJyuhAypgU+VsRB5lD4DuiMws5Ltc=",
+        "narHash": "sha256-/74n1oQPtKG52Yw41cbToxspxHbYz6O3vi+XEw16Qe8=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "4e92bbcdb030f3b4782be4751dc08e6b6cb6ccf2",
+        "rev": "36a601196c4ebf49e035270e10b2d103fe39076b",
         "type": "github"
       },
       "original": {
@@ -624,11 +624,11 @@
         ]
       },
       "locked": {
-        "lastModified": 1775701952,
+        "lastModified": 1775444042,
-        "narHash": "sha256-xj9u8fz2hTTTELMorqox0hPWrmAvGRnQUEnlj+vCjFo=",
+        "narHash": "sha256-cg19ipIlZaLYgs/5ZPFcDDuOcZlGzfprB5xS4x7bVM4=",
         "owner": "nix-community",
         "repo": "srvos",
-        "rev": "f56f1053ae9f878501d3a8ae1961c73d1d7abce3",
+        "rev": "64c9cc6a274dac7d08c4d53494ffa4acf906e287",
         "type": "github"
       },
       "original": {
@@ -715,11 +715,11 @@
     "trackerlist": {
       "flake": false,
       "locked": {
-        "lastModified": 1775686189,
+        "lastModified": 1775599784,
-        "narHash": "sha256-kzEDJKptaVToSg/wpub0bLjAVRmkYOorjPsNqlpxWdU=",
+        "narHash": "sha256-ZapxbiFEYjJV2nhdowHQ/8+c8Jd5fpBIEKDiPEmyNgI=",
         "owner": "ngosang",
         "repo": "trackerslist",
-        "rev": "ce9c0afc3885d0592caa91f0d4359f315ef7428c",
+        "rev": "6cc71b5b65349081bb713719f5142c200438a327",
         "type": "github"
       },
       "original": {

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
         ];
       })
     );
@@ -50,40 +51,17 @@ in
       "4096"
       "-ub"
       "4096"
-      "--parallel"
-      "2"
     ];
   };
 
   # have to do this in order to get vulkan to work
   systemd.services.llama-cpp.serviceConfig.DynamicUser = lib.mkForce false;
 
-  # ANV driver's turbo3 shader compilation exceeds the default 8 MB thread stack.
-  systemd.services.llama-cpp.serviceConfig.LimitSTACK = lib.mkForce "67108864"; # 64 MB soft+hard
-
   # llama-server tries to create ~/.cache; ProtectSystem=strict + impermanent
   # root make /root read-only. Give it a writable cache dir and point HOME there.
   systemd.services.llama-cpp.serviceConfig.CacheDirectory = "llama-cpp";
   systemd.services.llama-cpp.environment.HOME = "/var/cache/llama-cpp";
 
-  # turbo3 KV cache quantization runs a 14-barrier WHT butterfly per 128-element
-  # workgroup in SET_ROWS. With 4 concurrent slots and batch=4096, the combined
-  # GPU dispatch can exceed the default i915 CCS engine preempt timeout (7.5s),
-  # causing GPU HANG -> ErrorDeviceLost. Increase compute engine timeouts.
-  # Note: batch<4096 is not viable -- GDN chunked mode needs a larger compute
-  # buffer at smaller batch sizes, exceeding the A380's 6 GB VRAM.
-  # '+' prefix runs as root regardless of service User=.
-  systemd.services.llama-cpp.serviceConfig.ExecStartPre = [
-    "+${pkgs.writeShellScript "set-gpu-compute-timeout" ''
-      for f in /sys/class/drm/card*/engine/ccs*/preempt_timeout_ms; do
-        [ -w "$f" ] && echo 30000 > "$f"
-      done
-      for f in /sys/class/drm/card*/engine/ccs*/heartbeat_interval_ms; do
-        [ -w "$f" ] && echo 10000 > "$f"
-      done
-    ''}"
-  ];
-
   # upstream module hardcodes --log-disable; override ExecStart to keep logs
   # so we can see prompt processing progress via journalctl
   systemd.services.llama-cpp.serviceConfig.ExecStart = lib.mkForce (