From 0fc0cefb4eecf685c4bbd20912db9c1294d84c1e Mon Sep 17 00:00:00 2001
From: Justin Barlow <7@netlobo.com>
Date: Thu, 8 Oct 2026 09:34:55 -0600
Subject: [PATCH] [Bugfix] Don't int-sort attention-type names in
 --kv-cache-dtype-skip-layers for packed KV dtypes (#60364)

Signed-off-by: lobo235 <7@netlobo.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
---
 vllm/engine/arg_utils.py | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 94dd3e2ef1..5f32ec8deb 100644
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -2285,9 +2285,11 @@ class EngineArgs:
             )
 
             boundary = TurboQuantConfig.get_boundary_skip_layers(model_config)
-            existing = set(cache_config.kv_cache_dtype_skip_layers)
+            # Sorted so the order does not depend on set iteration: the list is
+            # a CacheConfig hash factor. Entries can also be attention type
+            # names (e.g. "sliding_window"), so sort them as strings.
             cache_config.kv_cache_dtype_skip_layers = sorted(
-                existing | set(boundary), key=int
+                set(cache_config.kv_cache_dtype_skip_layers) | set(boundary)
             )
 
         ray_runtime_env = None