chore(halo): llama.cpp update

overlays/unstable/default.nix

@@ -20,16 +20,19 @@ final: prev: {
       rocmGpuTargets = [ "gfx1151" ];
     }).overrideAttrs
       (prevAttrs: {
-        src = prev.fetchFromGitHub {
-          owner = "am17an";
+        version = "9264";
+        src = final.fetchFromGitHub {
+          owner = "ggml-org";
           repo = "llama.cpp";
-          rev = "9a532ae4bab1b164052ce60a738f78538b421c66";
-          hash = "sha256-LIGtXaO53Y3Ze0x6uevb2vzcKHbyj2o/8ZHvyZ22uo4=";
-          #postFetch = ''
-          #  echo -n "9a532ae4bab1b164052ce60a738f78538b421c66" > $out/COMMIT
-          #'';
+          tag = "b9264";
+          hash = "sha256-LA4SgE20Dvz1g3degdIx4CYfYhVNEIQM5Q/5rDT/icg=";
+          leaveDotGit = true;
+          postFetch = ''
+            git -C "$out" rev-parse --short HEAD > $out/COMMIT
+            find "$out" -name .git -print0 | xargs -0 rm -rf
+          '';
         };
-        npmDepsHash = "sha256-Po5SWJv3vmcBR7y62G9/CfvI3Lk/MYdjFMTTy2dsgoY=";
+        npmDepsHash = "sha256-Iyg8FpcTKf2UYHuK7mA3cTAqVaLcQPcS0YCa5Qf01Gc=";
         npmRoot = "tools/ui";
       });
 

systems/x86_64-linux/halo/models.ini

@@ -17,7 +17,7 @@ c                   = 131072
 [Qwen3.6-27B]
 hf                  = unsloth/Qwen3.6-27B-MTP-GGUF:Q6_K
 spec-type           = draft-mtp
-spec-draft-n-max    = 6
+spec-draft-n-max    = 4
 threads-batch       = 16
 temp                = 0.6
 top-p               = 0.95