From be0dcc29dcfa83659e6857dd73cb527b5986a7f9 Mon Sep 17 00:00:00 2001
From: Yan Ma <yan.ma@intel.com>
Date: Sat, 9 May 2026 09:19:05 +0800
Subject: [PATCH] [XPU] remove q/k/v force contiguous for flash_attn (#40356)

Signed-off-by: Yan Ma <yan.ma@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
---
 vllm/_xpu_ops.py | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/vllm/_xpu_ops.py b/vllm/_xpu_ops.py
index 09f700d0de7..6a2e5e841ce 100644
--- a/vllm/_xpu_ops.py
+++ b/vllm/_xpu_ops.py
@@ -387,14 +387,9 @@ class xpu_ops:
             assert len(window_size) == 2
             real_window_size = (window_size[0], window_size[1])  # noqa: F841
 
-        # In encode attention, k and v maybe not contiguous and current
-        # kernel can't handle it
-        if block_table is None:
-            k = k.contiguous()
-            v = v.contiguous()
         return flash_attn_varlen_func(
             out=out,
-            q=q.contiguous(),
+            q=q,
             k=k,
             v=v,
             cu_seqlens_q=cu_seqlens_q,