diff --git a/tests/mm_router/test_vllm_mm_router_e2e.py b/tests/mm_router/test_vllm_mm_router_e2e.py
index dc71834d073b..0df70dc2d5ef 100644
--- a/tests/mm_router/test_vllm_mm_router_e2e.py
+++ b/tests/mm_router/test_vllm_mm_router_e2e.py
@@ -45,8 +45,8 @@
     pytest.mark.e2e,
     pytest.mark.vllm,
     pytest.mark.multimodal,
+    # Restore xpu_1 after the paged decode kernel is available; see #15417.
     pytest.mark.gpu_1,
-    pytest.mark.xpu_1,
     pytest.mark.model(VLLM_MM_MODEL),
 ]
 
diff --git a/tests/serve/multimodal_profiles/vllm_xpu.py b/tests/serve/multimodal_profiles/vllm_xpu.py
index 65dc7047d4db..4c5955c97f83 100644
--- a/tests/serve/multimodal_profiles/vllm_xpu.py
+++ b/tests/serve/multimodal_profiles/vllm_xpu.py
@@ -61,7 +61,13 @@
         gpu_marker="xpu_1",
         topologies={
             "agg": TopologyConfig(
-                marks=[pytest.mark.post_merge, pytest.mark.xpu_1],
+                marks=[
+                    pytest.mark.skip(
+                        reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417"
+                    ),
+                    pytest.mark.post_merge,
+                    pytest.mark.xpu_1,
+                ],
                 # TODO: re-enable XPU-parallel scheduling with
                 # profiled_vram_gib=9.6 once this has a bounded --kv-bytes profile.
                 timeout_s=300,
@@ -116,7 +122,13 @@
             # (post_merge).
             "agg_router": TopologyConfig(
                 health_check_workers=True,
-                marks=[pytest.mark.pre_merge, pytest.mark.xpu_2],
+                marks=[
+                    pytest.mark.skip(
+                        reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417"
+                    ),
+                    pytest.mark.pre_merge,
+                    pytest.mark.xpu_2,
+                ],
                 gpu_marker="xpu_2",
                 timeout_s=400,
                 profiled_vram_gib=18.7,
@@ -135,7 +147,13 @@
             # single-GPU CI environment.
             "agg_router_chat_processor": TopologyConfig(
                 health_check_workers=True,
-                marks=[pytest.mark.post_merge, pytest.mark.xpu_2],
+                marks=[
+                    pytest.mark.skip(
+                        reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417"
+                    ),
+                    pytest.mark.post_merge,
+                    pytest.mark.xpu_2,
+                ],
                 gpu_marker="xpu_2",
                 timeout_s=400,
                 profiled_vram_gib=18.7,
@@ -153,7 +171,13 @@
             # CI; the content-hash correctness assertion lives in
             # tests/mm_router/test_router_rust_mm_frontend_decode_e2e.py.
             "agg_router_frontend_decode": TopologyConfig(
-                marks=[pytest.mark.post_merge, pytest.mark.xpu_2],
+                marks=[
+                    pytest.mark.skip(
+                        reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417"
+                    ),
+                    pytest.mark.post_merge,
+                    pytest.mark.xpu_2,
+                ],
                 gpu_marker="xpu_2",
                 timeout_s=400,
                 profiled_vram_gib=18.7,
@@ -180,7 +204,12 @@
         topologies={
             "agg_router": TopologyConfig(
                 health_check_workers=True,
-                marks=[pytest.mark.post_merge],
+                marks=[
+                    pytest.mark.skip(
+                        reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417"
+                    ),
+                    pytest.mark.post_merge,
+                ],
                 timeout_s=500,
                 gpu_marker="xpu_2",
                 profiled_vram_gib=19.0,
@@ -196,7 +225,12 @@
         topologies={
             "agg_router": TopologyConfig(
                 health_check_workers=True,
-                marks=[pytest.mark.post_merge],
+                marks=[
+                    pytest.mark.skip(
+                        reason="vLLM XPU v0.29.0 lacks paged decode kernel 8,128,16,false,false,false; see #15417"
+                    ),
+                    pytest.mark.post_merge,
+                ],
                 timeout_s=500,
                 gpu_marker="xpu_2",
                 profiled_vram_gib=16.0,