From 976edab3399d5cfe4923a9bbfdde808abd6cc0e8 Mon Sep 17 00:00:00 2001
From: Evan Reichard <evan@reichard.io>
Date: Thu, 30 Apr 2026 07:45:57 -0400
Subject: [PATCH] config(llama-swap): enable preserve_thinking in chat template
 kwargs

---
 modules/nixos/services/llama-swap/config.nix | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/modules/nixos/services/llama-swap/config.nix b/modules/nixos/services/llama-swap/config.nix
index 264b5ea..cb34191 100644
--- a/modules/nixos/services/llama-swap/config.nix
+++ b/modules/nixos/services/llama-swap/config.nix
@@ -140,7 +140,8 @@ in
           --presence-penalty 0.0 \
           --repeat-penalty 1.0 \
           -dev CUDA0 \
-          -fit off
+          -fit off \
+          --chat-template-kwargs "{\"preserve_thinking\": true}"
       '';
       metadata = {
         type = [
@@ -168,7 +169,6 @@ in
           -dev CUDA0 \
           -fit off
       '';
-      # --chat-template-kwargs "{\"enable_thinking\": false}"
       metadata = {
         type = [
           "text-generation"
@@ -194,7 +194,8 @@ in
           -ctk q8_0 \
           -ctv q8_0 \
           -dev CUDA0 \
-          -fit off
+          -fit off \
+          --chat-template-kwargs "{\"preserve_thinking\": true}"
       '';
       # --chat-template-kwargs "{\"enable_thinking\": false}"
       metadata = {